index.html

<!-- 
	this is the head file that contains the headers section 
	.. upto the start tag of the body section
	please make sure that this file contains other contents before the first file starts
-->
<!DOCTYPE html>
<html>
<head>
	<title>Flash Memory Research: Comprehensive Guide</title>
	<meta name="viewport" content="width=device-width, initial-scale=1">
	<style>
	.collapsible {
	  background-color: #777;
	  color: white;
	  cursor: pointer;
	  padding: 2px;
	  width: 100%;
	  border: none;
	  text-align: left;
	  outline: none;
	  font-size: 15px;
	}

	.active, .collapsible:hover {
	  background-color: #555;
	}

	.content {
	  padding: 0 18px;
	  display: none;
	  overflow: hidden;
	  background-color: #f1f1f1;
	}
	</style>
</head>


<body>
<h1 align="center">Flash Research Comprehensive</h1>

<hr>

<p align="center">Prawar Poudel</p>

<p align="center">prawar(dot)poudel(at)hotmail(dot)com</p>

<hr>

<hr>

<button type="button" class="collapsible"> About this document</button>

<p>This document contains short description about all the research items read by Prawar Poudelduring his/her course of research. The items studied can be categorized into following classes.</p>
<p>Please follow the link below to jump to respective section:</p>

<p>
<ul><li><a href="#nvm_introduction">NVM Introduction</a></li><li><a href="#flash">Flash Memory</a></li><li><a href="#flash_security">Security Primitives in Flash Memory</a></li><li><a href="#flash_puf">Flash PUFs</a></li><li><a href="#flash_sanitization_paper">Sanitization of Flash Memory</a></li><li><a href="#ssd_papers_study">SSD papers</a></li><li><a href="#sram">SRAM</a></li><li><a href="#dram">DRAM</a></li><li><a href="#sram_dram_security">Security Primitives in SRAM and DRAM</a></li><li><a href="#fram">FRAM</a></li><li><a href="#rram">RRAM</a></li><li><a href="#mram">MRAM</a></li><li><a href="#pcm">PCM</a></li><li><a href="#optane">Optane</a></li><li><a href="#file_system_study">Study of File Systems</a></li><li><a href="#magnetic_disk">Magnetic Disks</a></li><li><a href="#other_study">Others</a></li>
</ul>
</p>
<hr>

<hr>

<h2 align="center"><a name="nvm_introduction">NVM Introduction</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: NVM Introduction</i></button>

<div>
<p>
	Memory forms an integral component of every computing systems. These are the parts of the computer that are used for storage and retrival of code and data that are essential for operation of computers. It is these retrival of code and data the makes these computers or computing systems perform all the tasks that we, the user, instruct them to do.
</p>
<p>
	When the term "memory" is used, it generally means the Random Access Memory (RAM) or the memory where the instructions for current (and maybe more) operations are stored. But, this is just a small part of memory system in a computer system. Now a reader might be wondering where the bulk files, multimedia, etc are stored. There is another memory, which is more generally called "storage". It is in this storage system that all bulk files are stored. In the conventional workstation computer settings, Magnetic Disks form the storage system. Nowadays, Solid State Drives (SSDs) are getting traction as well for storage purposes.
</p>
<p>
	In the conventional computer settings, memory or RAM is composed of Dynamic Random Access Memory (DRAM). These memory loose their content when the power is removed, thus they form "Volatile Memory". Another example of volatile memory is SRAM, which is extensively used in caches in conventional PC settings while in embedded systems SRAM is used to serve as RAM.
</p>
<p>
	Magnetic Disks and Optical Disks are some of the non-volatile memory as they do not remove their content when the power is turned off. These devices store the content as physical change in them, and this physical change do not depend on the power being supplied. However for SRAM and DRAM, they need to be supplied with power in order for them to function or retain their content.
</p>
<p>
	Both the kinds of non-volatile memory (NVM) presented above, namely, Magnetic and Optical Disks are not semiconductor based. Thus, when people talk about NVM, these two are kind of left out. We will present a short description of these in the below section, however, they will not be major focus.
</p>
<p>
	Semiconductor NVMs are based on the similar principle as Magnetic and Optical Disks in the sense that the content or the data to be stored in them are etched as a change in their physical property and they do not need a constant supply of power to hold the power. Flash Memory are currently the major semiconductor right now, and it has still not realized its full market potential. As metioned above, flash memories are used for storage purpose in newer computing systems as SSD drives while they have found their majoirty usage in portable storage media. In low-end computing systems, a special kind flash memory is even used for RAM memories. We will see about different kinds of flash memories in later sections.
</p>
<p>
	Other semiconductor NVMs are listed below:
	<ul>
		<li>
			Resistive Random Access Memory (RRAM)
		</li>
		<li>
			Ferroelectric Random Access Memory (FRAM)
		</li>
		<li>
			Magnetic Random Access Memory (MRAM)
		</li>
		<li>
			Phase Change Memory (PCM)
		</li>
		<li>
			3D Cross Point (Optane)
		</li>
	</ul>
</p>
<p>
	Each of these have their separate sections below. Please find the approprite section and study the further descriptions presented. On each of the topics, will present a brief introduction and some relevant papers and their discussion.
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="flash">Flash Memory</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Flash Memory</i></button>

<div>
<p>
	<button type="button" class="collapsible"><i>List of Papers Studied </i></button>
		<div>
			<p>
				Following are the papers studied:
				<ul>
					<li>
						<a href="#flashbasics">Basics of Flash Memory</a>
					</li>
					<li>
						<a href="#flashtypes">Types of Flash Memory</a>
					</li>
					<li>
						<a href="#3dflash">Architectural and Integration Options for 3D NAND Flash Memories</a>
					</li>
					<li>
						<a href="#heatwatch3dnand">HeatWatch: Improving 3D NAND Flash Memory Device Reliability by Exploiting Self-Recovery and Temperature Awareness</a>
					</li>
					<li>
						<a href="#improving3D">Improving 3D NAND Flash Memory Lifetime by Tolerating Early Retention Loss and Process Variation</a>
					</li>
					<li>
						<a href="#mlcnandthresholdvoltage">Threshold Voltage Distribution in MLC NAND Flash Memory: Characterization, Analysis, and Modeling</a>
					</li>
				</ul>
			</p>
		</div>
		<hr>

	<button type="button" class="collapsible"> <a name="3dflash">Architectural and Integration Options for 3D NAND Flash Memories</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The file can be found at <a href="https://www.mdpi.com/2073-431X/6/3/27">this link</a>
					</li>
					<li>
						Different issues are faced upon by the feature size shrinking of flash memory.
					</li>
					<li>
						Intrinsic phenomena like random doping fluctuations, random telegraphic noise and the electron injection statistics significantly reduce the margin between the threshold voltage distributions.
					</li>
					<li>
						Multi bits per cell has made things worse reducing reliability and lifetime.
					</li>
					<li>
						3D NAND flash has provided with a way to overcome all the limitations of planar flash memory.
					</li>
					<li>
						Most of the 3D NAND rely on charge trap flash architecture while some rely on FG architecture.
					</li>
					<li>
						In this paper, they discuss about the charge trap flash memory and 3D NAND architectures based on the CT process.
					</li>
					<li>
						Following are two of the topological classification of different integration solutions:
						<ul>
							<li>
								<b>Control Gate and Channel along the horizontal direction</b>: This includes 3D Stacked option which is the preliminary attempt to achieve 3D integration starting from planar technology.
							</li>
							<li>
								<b>Control Gate along the horizontal direction and channel along the vertical direction</b>: This includes the architectures like BiCS, P-BiCS, and the V-NAND architecture from Samsung including VRAT, VSAT and TCAT.
							</li>
						</ul>
					</li>
				</ul>
				<ol>
					<li>
						3D Stacked Architecture:
						<ul>
							<li>
								first idea and straightforward idea that stacking multiple planar layers of memory arrays would yield 3D layer.
							</li>
							<img src="images/flash_1.PNG" align="center">
							<li>
								As shown in image above.
							</li>
							<li>
								Here, drain and bitline contacts are shared between NAND strings belonging to different layers, while source/wordlines contacts and source/drain selectors are associated to separate layers.
							</li>
							<li>
								Cost and process technologies considerations of this architecture can be derived from those of the planar products
							</li>
							<li>
								the major hurdle is represented by the thermal budget of the manufacturing process to grow and populate additional layers
							</li>
							<li>
								here each layer is manufactured separately and this the architecture is flexible.
							</li>
							<li>
								Since the layers are fabricated independently, there is a significant difference in the threshold voltage distribution with ISPP programming.
							</li>
						</ul>
					</li>
					<hr>
					<li>
						BiCS Architecture:
						<ul>
							<li>
								Control Gates are the different rectangles stacked on top of each other.
							</li>
							<img src="images/flash_2.PNG" align="center">
							<li>
								The bottom rectangle plate is the ensemble of Source Line Selectors terminating flash string.
							</li>
							<li>
								Multiple holes are drilled through the stacks and filled with poly-silicon in order to form a series of vertically arranged NAND flash memory cells.
							</li>
							<li>
								Bitline Selectors (BLS) and Bitlines (BL) contacts are on top of the structure
							</li>
							<li>
								Each cell in the BiCS architecture works in depletion-mode since the poly-silicon constituting the body of the transistor is lightly n-doped with a uniform profile or even left un-doped. This reduces the manufacturing complexity of the p-n junction along the vertical direction of the plugs (also called pillars)
							</li>
							<li>
								The CG plate intersection with a pillar maps a single memory cell. Each NAND Flash string of cells is connected to a BL contact via BLS, whereas the bottom of the string is connected to a common source diffusion formed directly on the process substrate made of silicon.
							</li>
							<img src="images/flash_3.PNG">
							<li>
								
							</li>
						</ul>
					</li>
					<li>
						P-BiCS
						<ul>
							<li>
								BiCS evidenced some critical issues such as poor reliability characteristics of the memory cells interms of endurance and data retention, poor SLS performances (i.e., cut-off) and a high resistance of the SL, which limits the sensing performance
							</li>
							<li>
								To solve these issues, a pipe-shaped BiCS architecture has been developed, namely the P-BiCS. This integration approach adopts a U-shaped vertical AND string
							</li>
							<li>
								<img src="images/flash_5.PNG" align="center">
							</li>
							<li>
								
							</li>
						</ul>
					</li>
					<li>
						VRAT Architecture
						<ul>
							<li>
								Vertical Recess Array Transistor by Samsung
							</li>
							<li>
								Still a charge trap layer is used
							</li>
							<li>
								no staircase structure
							</li>
							<li>
								<img src="images/flash_6.PNG" align="center">
							</li>
							<li>
								
							</li>
						</ul>
					</li>
					<li>
						VSAT Architecture
						<ul>
							<li>
								Vertical Stacked Array Transistor
							</li>
							<li>
								<img src="images/flash_7.PNG" align="center">
							</li>
						</ul>
					</li>
				</ol>
			</p>
		</div>
		<hr>

	<button type="button" class="collapsible"> <a name="heatwatch3dnand">HeatWatch: Improving 3D NAND Flash Memory Device Reliability by Exploiting Self-Recovery and Temperature Awareness</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The original paper can be found at <a href="https://www.cs.cmu.edu/~yixinluo/index_files/heatwatch_hpca18.pdf">this link</a>
					</li>
					<li>
						Damages in 3D NAND can be partially recovered by a process called "self-recovery"
					</li>
					<li>
						Self recovery has been studied for 2D NAND before to improve lifetime. Proposal also included to increase temperature to accelerate self recovery
					</li>
					<li>
						Heatwatch proposed to improve 3D NAND relability. Key idea is to optimize the read reference voltage (adapting it to the dwell time of workload and the current operating temperature).
					</li>
					<li>
						Heatwatch:
						<ul>
							<li>
								efficiently tracks flash memory temperature and dwell time online
							</li>
							<li>
								sends this information to our reliability model to predict the current voltages of flash cells
							</li>
							<li>
								predict the current voltages based on the current cell voltages
							</li>
						</ul>
					</li>
					<li>
						Improves lifetime by 3.85X over the baseline that uses fixed reference voltage
					</li>
					<li>
						Limited lifetime of NAND is because of wearout in flash cells by repeated programming and erase operations. After each PE cycle, the threshold voltage is affected because of the inadvertant trap of electrons in flash cell.
					</li>
					<li>
						Some of these trapped electrons gradually escape during the idle time between the consecutive PE cycle. This time is called "dwell time"
					</li>
					<li>
						This escape or detrapping is called self-recovery. 
					</li>
					<li>
						This can be accelerated by high temperature to the flash cell during the dwell time in 2D FG transistors.
					</li>
					<li>
						But 3D NAND uses CT flash
					</li>
					<li>
						Two effects shown in flash memory is studied in this paper:
						<ul>
							<li>
								Retention Loss: Leakage of charge carriers from the flash cell that contains valid data, that can induce errors
							</li>
							<li>
								Program Variation: Random variation that can cause cell to be set to an incorrect voltage while programming
							</li>
						</ul>
					</li>
					<li>
						Some of the findings made are as follows:
						<ul>
							<li>
								increase in dwell time from 1 to 137 min slows retention loss by 40%
							</li>
							<li>
								lowering temp from 70 to 0 C slows retention loss by 58%
							</li>
							<li>
								increasing temperature from 0 to 70 C during programming increases program variation by 21%
							</li>
							<li>
								effectiveness of self-recovery is correlated with dwell time experienced during the most recent 20 PE cycles
							</li>
						</ul>
					</li>
					<li>
						<img src="images/flash_4.PNG" align="center">
					</li>
					<li>
						Errors in 3D NAND is dominated by retention errors. Thus, reducing retention errors by performing recovery cycles can increase lifetime. Recovery cycle means PE operation where program operatin is followed by an extended dwell time.
					</li>
					<li>
						High temp increases electrom mobility. Thus short retention time in high temperature has same retention loss as long retention time at room temp.
					</li>
					<li>
						And short dwell time at high temp as same self recovery effect as long dwell time at room temp.
					</li>
					<li>
						<b>They refer to other papers [8,12,24] for technique to find the threshld voltage values. </b>
					</li>
					<li>
						Read retry and the above technique is used to detect threshold voltage and fine tune it.
					</li>
				</ul>
			</p>
		</div>
	<hr>


	<button type="button" class="collapsible"> <a name="improving3D">Improving 3D NAND Flash Memory Lifetime by Tolerating Early Retention Loss and Process Variation</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						Following is derived from the paper at <a href="https://arxiv.org/pdf/1807.05140.pdf">this link.</a>
					</li>
					<li>
						3D NAND uses new flash cell design that vertically stacks dozens of silicon layers in a single chip.
					</li>
					<li>
						Due to the difference in the physical orfanization at circuit level and structural level, there are three new error sources:
						<ul>
							<li>
								Layer to layer process variation: Average error rate of each 3D stacked layer is significantly different. Raw bit error rate at the middle can be 6x that of top layer.
							</li>
							<li>
								Early Retention Loss: A new phenomenon where the number of error due to charge leakage increases quickly within several hours (~3 hours) after programming, and then increases at a much lower rate.
							</li>
							<li>
								Retention Interference: Rate at which charge leaks from a flash cell depends on the data value stored in the neighbouring cell. Charge leaks at a lower rate when the vertically adjacent cell is in a state that holds more charge.
							</li>
						</ul>
					</li>
					<li>
						Different techniques are proposed for mitigation of various process variation  and early retention loss.
						<ul>
							<li>
								LaVAR or Layer Variation Aware Reading: Reduces the effect of layer to layer process variation by fine-tuning the read reference voltage separately for each layer.
							</li>
							<li>
								LI-RAID or Layer Interleaved Redundant Array of Independent Disks: Intelligently groups pages under the RAID error recovery technique using the info about layer to layer process variation
							</li>
							<li>
								ReMAR or Retention Model Aware Reading:  Reduces the retention errors by tracking the retention time of data using the retention model (new model they propose) and adapts read reference voltage to data age
							</li>
							<li>
								ReNAC or Retention Interference Aware Neighbour Cell Assisted Correction: Adapts the read reference voltage to the amount of retention interference a page has experienced in order to re-read the data after a read operation fails.
							</li>
						</ul>
					</li>
					<li>
						These four techniques can be combined together to improve the flash reliability.
					</li>
					<li>
						Most of the studies that have been done are based on planar NAND flash memory that are based on floating gate transistor. 3D NAND are however based on charge trap transistor where the charge is stored within an insulator. Here multiple layers of silicon are stacked together (typically 24 to 96 layers).Thus 3D NAND flash can be manufactured at higher technology node as 30-50 nm as compared to highly compact 10-15 nm for planar NAND.
					</li>
					<li>
						Following are the errors that has been widely documented in planar NAND flash memory:
						<ul>
							<li>
								Process Variation Error: Error due to fabrication process and lithography limitiations.
							</li>
							<li>
								Retention Error: With time, charge leaks off the flash cell after being programmed. Refresh technique is required.
							</li>
							<li>
								Write-Induced Error: Warout error.
							</li>
							<li>
								Cell-to-Cell Interference Error: Increase in threshold voltage of a cell (and eventually RBER) when adjacent flash cell is programmed.
							</li>
							<li>
								Read Induced Error: Application of pass through voltage during read operation causes weak programming effect. The values of flash cell may change in multiple read operations.
							</li>
						</ul>
					</li>
					<li>
						However 3D NAND differs from planar NAND in many ways as following:
						<ul>
							<li>
								Flash Cell Design:
								<ul>
									<li>
										Planar memory uses storage of charge in FG, but 3D uses Charge Trap transistors.
									</li>
									<li>
										In CT flash, a cylindrical substrate is between charge trap insulator layer that is radially wrapped inside control gate. 
									</li>
									<li>
										Charge Trap layer stores the charge, but is insulator (as oppposed to FG).
									</li>
								</ul>
							</li>
							<li>
								Chip Organization:
								<ul>
									<li>
										BL in planar NAND connects (source and drain of) the flash cells from different pages in a block.
									</li>
									<li>
										BL in 3D NAND, which is CT, connects one charge trap cell from each layer of the chip as cells are stacked on top of each other. All the cells in a single stack (z-axis) share the same charge trap insulator.
									</li>
									<li>
										CG in the same layer are connected through a Word Line in Y-axis. The number of WLs in a block is proportional to the length (height) of BL.
									</li>
									<li>
										Multiple blocks are aligned in X-axis to form a flash chip.
									</li>
								</ul>
							</li>
							<li>
								Manufacturing Process Technnology:
								<ul>
									<li>
										3D NAND uses much larger mnufacturing technology (30-50 nm) as compared to planar 3D NAND flash memory (10-15 nm).
									</li>
								</ul>
							</li>
						</ul>
					</li>
					<li>
						They use read-retry command to fine-tune the reference voltage for each read operation. (refs 9 and 14). This changes the read reference voltage by a small step called <i>voltage step</i>.
					</li>
					<li>
						Metrics are
						<ul>
							<li>
								RBER: is the rate at which errors in the data before error correction.
							</li>
							<li>
								Change in Threshold Voltage Distribution: How it changes with different sources of errors. They use read retry method to sweep over all the reference voltages to identify the threshold voltage (using techniques as in refs 14, 64 and 81).
							</li>
						</ul>
					</li>
					<li>
						Layer to Layer Process Variation:
						<ul>
							<li>
								The chips used have 30-40 layers and the observations are normalized to 100 layers for anonymity of the manufacturer. It is also not known from paper how they distinguish between flash cells in differet layers.
							</li>
							<li>
								Variation exists in all three axis. However, most prominent is in the z-axis because of difficulty in stacking multiple layers in top of each other. 
							</li>
							<li>
								Prior work has shown that no etching tech can are able to produce identical 3D NAND cells while punching through multiple stacked layers
							</li>
							<li>
								For characterization, each pages in the blocks are programmed with some random data such that each page has faced 10K PE cycles. Then comparison of collective characteristics of flash cells in a layer is made with another layer. (Repeated in multiple chips)
							</li>
							<li>
								The plot for number of errors in each layer of a 3D NAND flash cell (for MLC) while transitioning from ER<->P1, P1<->P2 and P2<->P3 is plotted while another plot showing error with in MSB and LSB pages are plotted.
							</li>
							<li>
								Error for ER<->P1 is significantly large while the error for P1<->P2 is also large across layers. However error for P2<->P3 remain similar across layers. (This shows that ER threshold voltage varies significantly across layers)
							</li>
							<li>
								MSP and LSB errors vary significantly across layers. <i>This is termed as layer-to-layer process variation.</i>
							</li>
							<li>
								MSB errors are higher than LSB errors in most of the layers. This is because reading MSB requires two reference voltages while reding LSB requires single reference voltages.
							</li>
							<li>
								The top half of the layers have low error rates than bottom half because of flash cell size varies across layers.
							</li>
							<li>
								The observed pattern of error is consistent in another different chip as well.
							</li>
							<li>
								Optimal Read Reference voltage vary across layers as well. The optimal reference voltages (Va and Vb) vary significanlty across layers but Vc does not change much. This is because process variation affects ER and P1 state while threshold voltage distribution of P2 and P3 states are controlled by ISPP that is similar across multiple layers.
							</li>
							<li>
								Va and Vb are low for top half of layers than for bottom half.
							</li>
						</ul>
					</li>
					<li>
						Early Retention Loss:
						<ul>
							<li>
								For a larger retention analysis, they program 11 flash blocks in a chip and write a pseudo random data to each page of the block. The memory is worn out where each block is exposed to different PE cycles (each blocks from 0K to 10K, thus 11 blocks). It is left for 24 days at room temperature.
							</li>
							<li>
								For measuring retention loss, RBER and the threshold voltage distribution is measured at 9 retention times from 7 minutes to 24 days. Only frst 72 flash pages in each bolcks are used in measurement so that other errors would not affect observation.
							</li>
							<li>
								Observation is that retention rate changes very slowly  for planar NAND than 3D NAND. ALthough 3D has very less RBER initially, after 2 hours of retention time, it quickly outgrows planar RBER. RBER increases by the order of 10^4 in 3 hours and 10^6 in 11 days.
							</li>
							<li>
								Retention Error flattens out after time in order of 10^7 seconds.
							</li>
							<li>
								Early retention loss is caused by two possible reasons:
								<ul>
									<li>
										Tunnel Oxide is thinner in 3D NAND for faster programming.(since it should act as insulator but since being thinner is susceptible to leakage).
									</li>
									<li>
										Cells in the same bit line share the same charge trap layer. This causes the charge in one cell to quickly leak to another cell in the same bit line due to <i>Electron Diffusion</i>.
									</li>
								</ul>
							</li>
							<li>
								Optimal Read reference voltage also changes with retention time.
								<ul>
									<li>
										Va remains fairly constant while Vb and Vc decreases with increase in retention times.
									</li>
								</ul>
							</li>
						</ul>
					</li>
					<li>
						Retention Interference
						<ul>
							<li>
								This means that the speed of retention loss of a cell depends on the threshold voltage of a vertically adjacent neighbour cell (shared charge trap layer).
							</li>
							<li>
								If two cells have different threshold voltages, charge can leak from cell with higher threshold voltage to the cell with a lower threshold voltage.
							</li>
							<li>
								Same observation from retention loss is used to observe retention interference. 
							</li>
							<li>
								However, only the neighbouring cells that are programmed before victim cells are used for interference correlation computation. Victim cells in ER state are also ignored because they are significantly affected by program interference even though they are programmed after neighbour cell. (These are done to remove noise due to program interference.)
							</li>
							<li>
								For retention interference, all victim cells are grouped together based on their threshold voltage states and the states of neighboring cells. Then the shift in threshold voltage after 24-day retention time was compared.
							</li>
							<li>
								The observation was that the change in reference voltage in terms of voltage steps shows correlation with the neighboring cell voltage.
							</li>
						</ul>
					</li>
					<li>
						No evidence of program error was found in 3D NAND
					</li>
					<li>
						PE cycling error is present in 3D NAND similar to planar NAND
					</li>
					<li>
						3D NAND face less program interference than planar NAND.
					</li>
					<li>
						Read disturb error is very weak so much so that it does not require any error mitigation.
					</li>
					<li>
						LaVAR: Layer Variation Aware Reading
						<ul>
							<li>
								to account for layer to layer process variation, how much should the reference voltage be offset? (instead of using a single reference voltage for entire block)
							</li>
							<li>
								LaVAR learns the offset for each layer and records them in per chip table in SSD controller
							</li>
							<li>
								uses this variation aware Vopt during read operation by reading appropriate voltage offset for the request from the per chip table that corrsponds to the layer of request.
							</li>
							<li>
								since there are limited number of layers, this can be represented as a table between thet variation agnostic voltage and the optimal voltage for each layer
							</li>
							<li>
								since the change in reference voltage follows similar pattern with PE cycles across different blocks, rigorous analysis of a single block should provide model for all blocks.
							</li>
							<li>
								LaVAR uses read retry functionality in modern NAND flash chips to find variation aware Vopt (optimal read referece voltage).
							</li>
							<li>
								Vc does not need to be changed while Va and Vb needs to be changed and the overhead is 2N bytes for N layers.
							</li>
						</ul>
					</li>
					<li>
						LiRAID: Layer Interleaved RAID
						<ul>
							<li>
								To tolerate chip to chip variation, RAID technique is used in modern SSDs. RAID in SSDs combine pages from multiple chips into a logical unit called a super group and uses one of the pages to store parity information. But they do not consider layer to layer variation
							</li>
							<li>
								Instead of grouping pages in the same layer together in the same group, pages from different chips and different layers are grouped together so that new group has evenly distributed RBERs in LiRAID.
							</li>
							<li>
								In a case with m chips in an SSD, each RAID group contains m pages one from each chip.
							</li>
							<li>
								If there are n wordlines, then LiRAID groups MSB page of wordline 0 with  LSB page of word line n/m, MSB page of 2*n/m and so on, MSB page of wordline (m-2)*n/m, LSB page of (m-1)*n/m. Thus combines LSB and MSB pages and combines different layers in a group.
							</li>
						</ul>
					</li>
					<li>
						ReMARL Retention Model Aware Reading:
						<ul>
							<li>
								Retention loss occurs after programming a page. RBER of early loss is logarithm of the retention time which means majority of threshold voltage  shifts occurs shortly after the programming operation.
							</li>
							<li>
								Idea here is to track the retention time and apply appropriate optimal read reference voltage.
							</li>
							<li>
								Prediction of optimal Va, Vb and Vc are done. Va is not affected by retention time while Vb and Vc are modeled with retention time dependency (result obtained from above).
							</li>
							<li>
								A sweeping read reference voltage learning is performed on one of the block to learn the optimal read reference voltage to construct the model <i>online</i>
							</li>
							<li>
								SSD controller must store the PE cycle count and the program time of each block.
							</li>
						</ul>
					</li>
					<li>
						ReNAC: Retention Interference Aware Neighbour Cell Assisted Correction
						<ul>
							<li>
								Data stored in vertically adjacent cell to predict the amount of retention interference on a victim cell.
							</li>
							<li>
								A model is generated online for retention interference as a function of retention time and neighbour cells state.
							</li>
							<li>
								A read offset is computed for each block and is applied to the read threshold voltage.
							</li>
							<li>
								(This method did not yield significant improvement)
							</li>
						</ul>
					</li>
				</ul>
			</p>
		</div>
		<hr>

	<button type="button" class="collapsible"> <a name="mlcnandthresholdvoltage">Threshold Voltage Distribution in MLC NAND Flash Memory: Characterization, Analysis, and Modeling</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						This paper uses <i>read retry</i> command in MLCs of 2Y nm technology for threshold voltage modeling
					</li>
					<li>
						Experimental measurement based characterization of threshold voltage distribution present in some 2Y flash memory
					</li>
					<li>
						For a flash cell with 2-bits, there are 2^2 or 4 states. This needs 2^2-1 or 3 read reference voltages, thus creating 4 regions of threshold voltage distribution.
					</li>
					<li>
						During a read operation, the threshold voltage is a cells is iteratively compared to predefined read reference voltages until the upper and lower bound are identified to determine the n-bit stored value
					</li>
					<li>
						In earlier flash design, read reference voltages are fixed in design time. But newer flash memories below 30 nm have a technique to address the distorted threshold voltage distribution. (Due to different reasons like PE cycling, retention loss, program disturb etc, the otherwise non-overlapping regions of voltage distributions might overlap or corss the fixed boundary. This may lead to value being misread).
					</li>
					<li>
						Read retry allows the voltage to be dynamically adjusted to track changes in distributions.
					</li>
					<li>
						Two new commands are required on the controller side:
						<ul>
							<li>
								Set REF: sets the reference voltaage to a new value 
							</li>
							<li>
								Get REF: checks the set value of the read reference voltage
							</li>
						</ul>
					</li>
					<li>
						As the reference voltage can be finely tuned, we can divide the threshold voltage to multiple bins, and the values read can be changed.
					</li>
					<li>
						Change in the threshold voltage due to PE cycling can also be done using the same technique.
					</li>
					<li>
						For the 3D NAND chip at our disposal, the host controller issues the SET FEATURES (EFh) or SET FEATURES by LUN (D5h) command to feature address 89h with P1 subfeature set to a read retry option, as defined in that feature address. A new NAND array READ operation can now be performed.
					</li>
				</ul>
			</p>
		</div>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="flash_security">Security Primitives in Flash Memory</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Security Primitives in Flash Memory</i></button>

<div>
<p>
	
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="flash_puf">Flash PUFs</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Flash PUFs</i></button>

<div>
<p>
	<p>
		The concept presented here in bullets are presented from the paper <a href="https://spqrlab1.github.io/papers/holcomb_PUFs_date14.pdf">PUFs at a Glance</a>
		<ul>
			<li>
				It is difficult to hold the secret key a secret in practice and have it permanently store as a secret. Some systems may not even have NVMs.
			</li>
			<li>
				Thus derive keys using random physical disorder or manufacturing variations in the devices such that even the manufacturers cannot recreate them.
			</li>
			<li>
				A PUF is deliberately designed making use of such variation that is an disordered physical system. It responds by providing a response to an external stimuli or challenge.
			</li>
			<li>
				PUFs differ from the normal digital circuits or systems in the sense that they work at the stability limits ie they are prone to errors. For the correction of such errors to produce stable output, two approaches are used:
				<ul>
					<li>
						First: Use of helper data. Helper data is derived in earlier generation of response and are stored somewhere in NVM. They can be constructed such that it can be known to the adversary without compromising the secrecy of PUF, ie it need not be secret.
					</li>
					<li>
						Second: Technique is to design the PUF system or protocol with inbuilt error tolerances.
					</li>
				</ul>
			</li>
			<li>
				Weak PUFs:
				<ul>
					<li>
						Characterized by a few challenges, commonly one challenge per instance of PUF
					</li>
					<li>
						The challenge-response mechanism of the PUF should be access restricted.
					</li>
					<li>
						The response of the limited challenges is known to manufacturer and other limited number of parties outside the hardware.
					</li>
				</ul>
			</li>
			<li>
				Strong PUFs:
				<ul>
					<li>
						More complex challenge response behavior. Large number of possible challenges can be applied. And even if the adversary knows a subset of CR pairs, s.he cannot extrapolate the behavior.
					</li>
					<li>
						they can remain unprotected because of availablity of large number of CR pairs.
					</li>
				</ul>
			</li>
		</ul>
	</p>
	<button type="button" class="collapsible"> <i>List of Papers Studied</i></button>
		<p>
			Following is the list of papers studied while creating this document. The description of each papers are maintained and linked here so it would be easy for anyone doing a survey to fetch the appropriate description in timely fashion.
			<ul>
				<li>
					<a href="#prabhu1">Paper: <i>Extracting Device Fingerprints from Flash Memory by Exploiting Physical Variations</i></a>
				</li>
				<li>
					<a href="#wang1">Paper: <i>Flash Memory for Ubiquitous Hardware Security Functions: True Random Number Generation and Device Fingerprints</i></a>
				</li>
				<li>
					<a href="#xu1">Paper: <i>Understanding sources of variations in flash memory for physical unclonable functions</i></a>
				</li>
				<li>
					<a href="#jia1">Paper: <i>Extracting Robust Keys from NAND Flash Physical Unclonable Functions</i></a>
				</li>
				<li>
					<a href="#sanu1">Presentation: <i>Energy Efficient Circuits for Entropy Generation and Secure Encryption</i></a>
				</li>
				<li>
					<a href="#mandadi1">Thesis: <i>Remote Integrity Checking using Multiple PUF based Component Identifiers</i></a>
				</li>
				<li>
					<a href="#clark1">Paper: <i>Reliable techniques for integrated circuit identification and true random number generation using 1.5-transistor flash memory</i></a>
				</li>
				<li>
					<a href="#sakib1">Paper: <i>An Aging-Resistant NAND Flash Memory Physical Unclonable Function</i></a>
				</li>
				<li>
					<a href="#erasesuspend1">Paper: <i>Extraction of Device Fingerprints using Built-In Erase-Suspend Operation of Flash Memory Devices</i></a>
				</li>
				<li>
					<a href="#eFlash">Paper: <i>Lightweight Integrated Design of PUF andTRNG Security Primitives Based on eFlashMemory in 55-nm CMOS</i></a>
				</li>
			</ul>
		</p>
		<hr>
		<button type="button" class="collapsible"> <a name="prabhu1">Paper: <i>Extracting Device fingerprints from Flash Memory by Exploiting Physical Variations</i></a></button>
			<div class="content">
				<p>
					This paper discusses seven techniques for generating fingerprints from flash memory. Published in 2011, the author P Prabhu was affiliated with UCSD while collaborators are from Cornell University. The original paper can be found at <a href="https://dl.acm.org/doi/10.5555/2022245.2022264">this link</a>. Their conclusion is the four of the seven techniques provide usable signatures.
				</p>
				<p>
					The set up consists of Xilinx FPGA connected to the flash chip through a custom built flash controller. The measurements can be made with a resolution of 10ns.
				</p>
				<p>
					<ul>
						<li>
							<p>
								<b>Program Disturb</b>: It is possible for two bits in separate pages to electrically influence each other. Program operation in a page might induce an mild programming force in an unselected word line. For generation of signature, repeated program operation is performed in a page. After each program operation, the state of adjacent page is read out to record the number of program operations required to cause the flip in each bit of adjacent page. This operation is performed for all the bits in the adjacent page. The number of program operations performed is limited to 10000. For the bits that are not affected, the value of 0 is recorded.
							</p>
							<p>
								The signature is page based and thus thousands of signatures can be extracted from a single flash memory. The time it takes is between 1-5 minutes.
							</p>
						</li>
						<li>
							<p>
								<b>Read Disturb</b>: The idea behind read disturb is similar to Program Disturb. Since for read operation, a small current has to be suppled as read current, this current might influence adjacent flash page over a number of times.
							</p>
							<p>
								The number of read operations in a flash page to influence adjacent flash page is limited to 10 million times while the state of adjacent flash page is read after every 1000 read iterations. Despite having good metrics for correlation (high for signature generated from a single page while almost 0 for those from different pages), it is too slow process. It almost took 6 hours to have useful signature for MLC chips.
							</p>
						</li>
						<li>
							<p>
								<b>Program Operation Latency</b>:  This technique is based on per-bit variation in program latency.Program operation is performed one bit at a time and the latency is recorded such that a single bit contributes a single value to the fingerprint. There are two following variations of the technique as discussed below:
							</p>
							<p>
								<ul>
									<li>
										<p>
											<i>Single Page</i>: Following steps are followed for fingerprint generation:
											<ul> 
												<li>Erase the block.</li>
												<li>Program each bit in a page within a block as follows:</li>
												<ul>
													<li>First program operation programs a single '0' followed by ones.</li>
													<li>Second program operation programs two '0's followed by ones.</li>
													<li>.. and so on.</li>
												</ul>
											</ul>
										</p>
									</li>
									<li>
										<p>
											<i>Multi Page:</i>Following steps are followed for fingerprint generation:
											<ul>
												<li>Erase the block</li>
												<li>Program just the nth bit in page n.</li>
												<li>.. and so on.</li>
											</ul>
										</p>
									</li>
								</ul>
							</p>
						</li>
						<li>
							Other proposals that were deemed unuseful are as follows:
							<ul>
								<li>
									<b>Erase Lantency</b>:
									<p>
										Measure the erase latency for each block in a device and use the resulting sequence of values to form a signature. Offered very little variation, and were inconsistent for individual blocks.
									</p>
								</li>
								<li>
									<b>Read Lantency</b>:
									<p>
										Per-page read latency varied very little. In SLC, read latency was constant over entire device.
									</p>
								</li>
								<li>
									<b>Whole Page program Latency</b>:
									<p>
										Whole page program latency offered very little vairation to offer good signature.
									</p>
								</li>					
								<li>
									<b>Program/Erase Wear</b>:
									<p>
										Measuring each bits susceptibility to error because of program/erase induced wear took several hours to extract signature thus making this method ineffective.
									</p>
								</li>		
							</ul>
						</li>
					</ul>
				</p>
			</div>
		<hr>

		<button type="button" class="collapsible"> <a name="wang1">Paper: <i>Flash Memory for Ubiquitous Hardware Security Functions: True Random Number Generation and Device Fingerprints</i></a></button>
			<div class="content">
				<p>
					This paper demonstrates NAND flash memrory as a source of entropy for TRNG and PUF generation, and is work done by Wang et al at Cornell University. The idea is based on repeated partial program operation on the flash memory. The original paper can be found at <a href="https://ieeexplore.ieee.org/document/6234403">this link</a>.
				</p>
				<p>
					For RNG, they make use of the Random Telegraphic Noise (RTN) as the source of randomness. RTN is the alternating capturing and emission of carriers at a defect site in a very small electronic device. This capturing and emission, that is random and exponentially distributed, generates discrete variation in the channel current.
				</p>
				<p>
					To observe this noise, the flash memory needs to be in unreliable state so that noise effects the outputs. Thus partial program operation is used to achieve this state that is in-between erased and programmed state. The initial algorithm is described as follows:
					<ol>
						<li>The flash memory block is partially programmed. The duration of partial program applied is T (unspecified)</li>
						<li>
							The flash memory block is read N times.	
						</li>
						<li>
							Then, for each of the bits, it is checked if there exists RTN.
						</li>
						<li>
							If it does, the number of partial program operation it took is noted, and the bit position is marked as selected. 
						</li>
						<li>
							This partial program operation is repeated until all the bits are marked selected. Go to step 1.
						</li> 
					</ol>  
					The second part of the algorithm, then can be used to generate RNG.
					<ol>
						<li>Partial program the flash cell to appropriate levels (as dictated by the number of partial program operations from above) </li>
						<li>Read each bit M times</li>
						<li>Record the sequence of up-times and down-times. Since it is RTN, the duration of being up-times and down-times are randomly distibuted.</li>
						<li>Produce Random bits from the up-times and down-times. If the time is odd, output 1 else output a 0.</li>
						<li>Perform debiasing</li>
					</ol>
				</p>
				<p>
					For the device fingerprint, repeated partial program operation is utilized. Following describes the steps involved:
					<ol>
						<li>
							initialize bitRank[i] = 0 for i = all bits in a page
						</li>
						<li>
							Perform partial program operation of flash page (T < rated program time)
						</li>
						<li>
							For all the bits in page:
							<ul>
								<li>
									If the bit is programmed and bitRank[this bit] = 0, bitRank[this bit] = partial program number
								</li>
							</ul>
						</li>
						<li>
							1Goto Step2 until 99% bits in the page are programmed
						</li>
					</ol>
					This essentially records the order in which the flash bits attain programmed state from erased state in a flash page.
				</p>
			</div>	
			<hr>

		<button type="button" class="collapsible"> <a name="xu1">Paper: <i>Understanding sources of variations in flash memory for physical unclonable functions</i></a></button>
			<div class="content">
				<p>
					This paper characterizes the errors and source of errors in flash memory. The authors were affiliated with Cornell University at the time of publication. The original paper can be found at <a href="https://doi.org/10.1109/IMW.2014.6849385"> this link</a>.
				</p>
				<p>
					<i><b>Since this is not focused about PUF but analyzing the nature of variations in flash, will populate later</b></i>
				</p>
			</div>
		<hr>

		<button type="button" class="collapsible"> <a name="jia1">Paper: <i>Extracting Robust Keys from NAND Flash Physical Unclonable Functions</i></a></button>
			<div class="content">
				<p>
					This paper presents three techniques for PUF-based key generator using NAND flash memory. The techniques are partial erase, partial programming and program disturbance. The primary author was a member of <i>Data Assurance and Communication Security Research Center, China</i>. The original paper can be found at <a href="https://doi.org/10.1007/978-3-319-23318-5_24">this link</a>.
				</p>
				<p>
					This paper focuses on robust keys rather than just determinig keys from the flash memories. They find the cells in the flash memory whose output are the most reliable over the life time of the flash memory. The position of such flash cells is the helper data.
				</p>
				<p>
					<b>Partial Erase</b> based operation is more or less similar to the idea proposed in Wang et al. paper presented <a href="#wang1">above</a>. However, since this proposal is based on partial erase opration, the number of (fixed duration) partial erase operation that it takes for each flash cell in a page is recorded and is the identity for that particular cell. In a way, it captures the order in which the cells attain erase state from a programmed state.
				</p>
				<p>
					The total number of partial erase operation is limited however to PENum. For some cells that do not attain the partial erase state even after the max number of partial erase operation performed, they are assigned PENum+1.
				</p>
				<p>
					<b>Partial Program</b> based operation is exactly same as partial erase based operation above. Here, the order in which the cells attain erased state is recorded as the number of (fixed duration) partial program operation required by each flash cells to attain programmed state is recorded. After a fixed number of partial program operations PPNum, the cells that are still not in programmed state are assigned PPNum+1
				</p>
				<p>
					<b>Program Disturb</b> based PUF generation induces disturbance in adjacent flash page by performing repeated program operation in a page. Here, fixed number of program operation is performed and after each of the program operation, the state of adjacent page is recorded to see if any cells might have been programmed enough to change their state to programmed state. Following the same suit, here again the number of program operations in the flash page is recorded for each of the flash cells in adjacent page it is needed to flip their state.
				</p>
				<p>
					<b>For reliable cell selection</b>, they employ two methods as follows:
					<ul>
						<li>
							<i>Bit Map Method:</i>
							<ul>
								<li>
									First the raw PUF values for a certain address location is noted. These address location is genrally adjacent locations and the raw PUF values are obtained from any of the three methods above.
								</li>
								<li>
									Between the adjacent values listed above, a difference is computed and absolute of this value is computed as ADvalues.
								</li>
								<li>
									Based on the need, the top n-values from among all the ADvalues are taken. As helper data, the selected pair is assigned 1 and rest unselected pair are assigned 0.
								</li>
								<li>
									For the final key, the two values used to compute the ADvalues is again checked. For each ADvalues, if the former of the two values is greater, then key at that position is 1 otherwise it is assigned 0. Using 6-flash cells, we get 3-bit key.
								</li>
							</ul>
						</li>
						<li>
							<i>Position Map Method:</i>
							<ul>
								<li>
									Here instead of selecting the adjacent flash cells, the flash values are first sorted from small to large based on raw PUF values.
								</li>
								<li>
									The lowest three values from among the selected raw PUFs are grouped with the highest three values, ie lowest with highest, lowest-second with highest-second and so on.
								</li>
								<li>
									The cell position of each of the raw PUF values are brought in pair. This is the helper data.
								</li>
								<li>
									For each pair, if the former raw PUF is higher than latter, key generated is 1 while 0 otherwise. Using 6 positions, we get 3-bit key.
								</li>
							</ul>
						</li>
					</ul>
				</p>
			</div>
			<hr>

		<button type="button" class="collapsible"> <a name="sanu1">Presentation: <i>Energy Efficient Circuits for Entropy Generation and Secure Encryption</i></a></button>
			<div class="content">
				<p>
					This is not a paper but a tutorial session presented by Sanu Matthew. Matthew was with Circuit Research Labs, Intel @ Hillsboro, Oregon while this presentation was made in ISQED 2020.
				</p>
				<p>
					This is not NAND based, but 14nm CMOS based, and is presented here because of the idea it has. The publication for this can be found at <a href="https://doi.org/10.1109/VLSIC.2018.8502369">this link</a>.
				</p>
				<p>
					The idea presented can be divided into two major categories:
					<ul>
						<li>
							First: There is noise in devices, so suppress the noises. There is process variation in devices, so amplify variation to create a physical identity of a device. Extract <i>static entropy</i>, ie entropy that remains same throughout the lifetime.
						</li>
						<li>
							Second: There is process variation in devices, so suppress the process variation. There is noise in devices, so amplify the noise for randomness or entropy generation. Extract <i>dynamic entropy</i>, ie entropy that is changes.
						</li>
					</ul>
				</p>
				<p>
					The idea for PUF generation or TRNG genration presented is to select the candidate bit for either operation. So, a selection criteria is selected, for example: read the bits multiple times (64 times as presented), and pick the ones with higher bias (either towards 0 or 1 state, determined using entropy computation) for PUF generation. Time varying bits can be used for TRNG purpose. This is based on the idea that the bits are time-variant, thus Temporal Majority Voting (TMV) is used. This idea can be easily ported in case of flash memories combined with the idea of flash cells oscillating. Please refer to <a href="https://ieeexplore.ieee.org/document/8443106">this paper</a>.
				</p>
				<p>
					The idea of self-calibration is presented whereupon entropy is tracked for TRNG. If the output of von-Neumann extractor for TRNG is less than 1bit/cycle, then a different column is chosen for TRNG. (the entropy source bits are arranged in 64x8 organization of rows and columns)
				</p>
			</div>
			<hr>

		<button type="button" class="collapsible"> <a name="mandadi1">Thesis: <i>Remote Integrity Checking using Multiple PUF based Component Identifiers</i></a></button>
			<div class="content">
				<p>
					This document presents PUF generation for NOR flash memory. The author was associated with Virginia Tech and the document is the master's thesis of author. The original document can be found at <a href="https://vtechworks.lib.vt.edu/handle/10919/78200"> this link</a>.
				</p>
				<p>
					The platforms used for demonstration of their idea is Altera DE1-SoC and Altera DE2-115.				
				</p>
				<p>
					The basic idea used in this project is partial programming. The threshold volatge variation is utilized to generate PUF. Here, the address of the flash memory location is the challenge and the response is the bit position of the cell with minimum threshold voltage. Essentially, the idea is to find the flash cell that gets programmed first while applying repeated partial programming operations on a flash memory location.
				</p>
				<p>
					<ol>
						Following are the steps involved:
						<li>A partial programming time T is chosen for a flash location.</li>
						<li>The flash location is erased.</li>
						<li>Program the flash location for duration T</li>
						<li>Read the value of the flash location</li>
						<ul>
							<li>If any one bit are flipped, the location of the flipped bit is encoded into 3-bit form</li>
							<li>If nore than one bit is flipped, T is reduced and go to Step 3</li>
							<li>Else, go to Step 3.</li>
						</ul>
					</ol>
				</p>
			</div>
			<hr>

		<button type="button" class="collapsible"> <a name="clark1">Paper: <i>Reliable techniques for integrated circuit identification and true random number generation using 1.5-transistor flash memory</i></a></button>
			<div class="content">
				<p>
					This document presents PUF generation and TRNG for Superflash NOR  memory. The author was associated with Arizona State University. The original document can be found at <a href="https://doi.org/10.1016/j.vlsi.2017.10.001"> this link</a>.
				</p>
				<p>
					The main idea exploited in this paper is the erase speed variability for the flash cells. The memory used in this research is 1.5T SST Superflash memory that has faster erase because split-gate flash memory has different organzation than stacked gate flash memory.
				</p>
				<p>
					Since such memory are higly efficient, program operation takes only one clock while erase requires multiple cycles. Thus in this research, they employ partial erase operation by interrupting erase operation. Erase operation, however, is yet very fast to be interrupted at nominal operating condition, thus the operating voltage VDD is reduced to lower the internal charge pump voltage. 
				</p>
				<p>
					Interrupting erase operation shows that the number of flash cells that attain erased state increases monotonically with higher erase times. They pick up a time that gives a little more than 50% 1's (flash cells in erased state). The distribution at this point in time shows that the flash cells in each of 1 and 0 state are almost randomly distributed over the block.
				</p>
				<p>
					For authentication, the same operation is repeated but such that it yields less than 50% 1's. The idea is again based on monotonicity: this means that if a binary signature with (for example) 45% 1s is compared with (challenge code) 55% 1s signature, the bits that are 1 in 45% signature should already be at 1 in 55% signature.
				</p>
			</div>
			<hr>

		<button type="button" class="collapsible"> <a name="sakib1">Paper: <i>An Aging-Resistant NAND Flash Memory Physical Unclonable Function</i></a></button>
		<div class="content">
			<p>
				This paper demonstrates a technique for generation of PUF from NAND flash memory such that the PUF is aging-resistant by introducing tunable-parameters. Authors of this papers were associated with The University of Alabama in Huntsville while it was published. The original paper can be found at <a href="https://ieeexplore.ieee.org/document/8993414">this link</a>.
			</p>
			<p>
				The basic idea behind the PUF generation in this article is program disturb. However, this program disturb method differs from previous proposals in the sense that here a single flash memory page is stressed and analyzed (unlike previous implementations where disturbance in adjacent page were observed).
			</p>
			<p>
				Following are the steps involved for PUF generation:
				<ul>
					<li>Erase a flash block.</li>
					<li>In the page of interest, program with all-ones data pattern except the last 32 bits which are programmed to zeros.</li>
					<li>Perform repeated program operation in the same page with the same data pattern.</li>
					<li>This repeated program operation (albeit with 1s) should cause the some flash cells to be weakly programmed.</li>
					<li>Thus, after certain number of program operation, some of the flash cells will start to read as 0 despite being programmed with 1.</li>
					<li>The idea is again based on the sequence in which bits attain the programmed state. Here, after application of certain number of program operation, the state of flash cells is recorded as the signature of the page. The number of program operation is chosen such that sufficient number of flash cells have attained programmed or 0 state.</li>
					<li>Repeating the same operation on the same page of flash memory should yield the same signature except for a few unstable bits.</li>
				</ul>
			</p>
			<p>
				These unstable bits are filtered out to generate more accurate PUF. For this, a plot with the progression of flash page from complete erased to programed state at different stress-levels (number of program operations) is plotted. Based on observation from this plot, two different threshold stress-levels are identified: one at the early stages to identify stable 0 cells (PS0) and second at the late stage to identify stable 1 cells (PS1). This gives the idea about the flash cells that attain the programmed state quickly (stable 0 cells) and cells that resist the change and maintain their state for longest (stable 1s).
			</p>
			<p>
				To generate the PUF of n-bits, at least n/2 stable 0s are generated and rest are stable 1s are obtained, thus removing the unstable flash cells that change their state in between the two thresholds discussed above.
			</p>
			<p>
				To obtain at least n/2 stable 0s, PS0 program operations as discussed above are performed. This will flip at least n/2 bits to 0 and gives information about the early flippers or stable 0s. Next, the page is stressed for PS1 program operations, and ensure that there are only n/2 erased bits left. These are stable 1s. The complete sequence (entire page) is compared at states PS0 and PS1 applied. If the states of the flash cells are same, they contribute to the PUF, else they are unstable bits.
			</p>
			<p>
				For authentication, the number of stresses to be performed is between the PS0 and PS1 (PS0&ltPSU&ltPS1).
			</p>
			<p>
				With usage or aging, the rate of bits that flip from erased to program state with program disturb decreases and might present issue for PUF. To counter this, PSU is adapted with the usage rate that page has endured.
			</p>
			<p>
				PSUnew = PSU+k*npe, where
				<ul>
					<li>k is constant derived from characterization</li>
					<li>npe is the number of program-erase the page has endured</li>
					<li>PSU is the number of program operation required to generate PUF in fresh consition</li>
				</ul>
			</p>
		</div>
		<hr>

		<button type="button" class="collapsible"> <a name="erasesuspend1">Paper: <i>Extraction of Device Fingerprints using Built-In Erase-Suspend Operation of Flash Memory Devices</i></a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The origianl paper can be found at <a href="https://ieeexplore.ieee.org/document/9097209">this link</a>.
					</li>
					<li>
						Utilize erase suspend operation in flash memory. They use Superflash from microchip, but most of their explanation follows from FG NAND flash memory.
					</li>
					<li>
						They generate 16-bit random number in 1ms
					</li>
					<li>
						They perform gradiant analysis of erase operation and since the gradiant is more stable at the later part, erase suspend at that area is more suitable for device fingerprint extraction.
					</li>
					<li>
						For fingerprint:
						<ul>
							<li>
								erase sector
							</li>
							<li>
								program the sector
							</li>
							<li>
								suspend the erase operation
							</li>
							<li>
								note the positions not erased. their fingerprint is the unerased bit position
							</li>
							<li>
								the process is repeated until the required number of bits are reached (lenght of fingerprint)
							</li>
						</ul>
					</li>
				</ul>
			</p>
		</div>
		<hr>
		<button type="button" class="collapsible"> <a name="eFlash">Paper: <i>Lightweight Integrated Design of PUF andTRNG Security Primitives Based on eFlashMemory in 55-nm CMOS</i></a></button>
			<div class="content">
				<p>
					<ul>
						<li>
							Derived from paper at <a href="Lightweight Integrated Design of PUF andTRNG Security Primitives Based on eFlashMemory in 55-nm CMOS">this link.</a>
						</li>
						<li>
							Based on embedded flash memory in 55-nm CMOS
						</li>
						<li>
							Randomness in nonliner IV characteristics and temporal current fluctuations are exploited
						</li>
						<li>
							Elaborate description of protocol from refs 1 and 2. Protocol will be studied later.
						</li>
						<li>
							
						</li>
					</ul>
				</p>
			</div>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="flash_sanitization_paper">Sanitization of Flash Memory</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Sanitization of Flash Memory</i></button>

<div>
<p>
	<button type="button" class="collapsible"> <i>List of sources</i></button>
		<div >
			<p>
				<ul>
					Following are the sources of the contents:
					<li>
						<a href="#proposal">From Proposal</a>
					</li>
					<li>
						<a href="#stateofart">State of the Art</a>
					</li>
					<li>
						<a href="#evanescoM">Evanesco: Motivation</a>
					</li>
					<li>
						<a href="#evanescoP">Evanesco: Proposal</a>
					</li>
					<li>
						<a href="#forensicNAND">Improving the Reliability of Chip-Off Forensic Analysis of NAND Flash Memory Devices</a>
					</li>
					<li>
						<a href="#dataDeletion">Investigation of Data Deletion Vulnerabilities in NAND Flash Memory Based Storage</a>
					</li>
				</ul>
			</p>
		</div>
		<button type="button" class="collapsible"> <a name="proposal"></a><i>From Proposal</i></button>
				<div class="content">
					<p>
						The content here is derived from the proposal that was submitted to NSF by Dr Biswajit Ray and Dr Aleksandar Milenkovic.
					</p>
					<p>
						The major idea is that instant deletion of data from the memory has become extremely important to preserve privacy of the user. According to <a href="http://www.legislation.gov.uk/ukpga/2018/12/contents/enacted"> Data Protection Act (DPA) 2018</a>, the deletion of imformation should be real and should not be recoverable in any way. However, current trend in SSD's do not offer any permanent data deletion strategy.
					</p>
					<p>
						NAND flash or flash memory in general follow the trend of erase before write paradigm, which means before any program operation erase has to be performed. But the granularity of erase (a block) is much larger than program (a page), thus for any modification of data is performed by copying the entire data, modifying it and writing it to a new location. The old data is simply marked as invalid or is unlinked where the original data content still remains. These pages are technically unreachable or unaddressable through the Flash Translation Layer (FTL).
					</p>
					<p>
						For the page level deletion of data, Wei et al introduced the concept of "scrubbing", which means write 0s to all the flash memory location. This corresponds to essentially deletion of data from a page by program operation. Program operation is possible in page-level granularity, thus deletion from page is possible. However, the "scrubbing" technique does not properly delete the data. There has been experimental demonstration where the data is recovered from a "scrubbed" flash memory by analyzing the physical property of the scrubbed flash memory.
					</p>
				</div>
				<hr>

			<button type="button" class="collapsible"><a name="stateofart"></a> <i>State of the Art</i></button>
				<div class="content">
					<p>
						The content here is still derived from the same proposal.
					</p>
					<p>
						<ul>
							<li>Block Erase: Erase the complete block. This means the charges trapped in the Floating Gate is removed. Garbage Collection of FTL uses this mechanism to remove the contents or free space when the drive is almost full. Poor performance because of significant data movement.</li><br>
							<li>Logical Sanitization: To overcome the performance hit by block erase, logical sanitization is performed. This works by simply invalidating the data that are obsolete. The old data still remains in the memory.</li><br>
							<li>Encryption Based Sanitization: Encrypt the user file and store the data and key in separate blocks. Remove the keys, thus securely invalidating the data.</li><br>
							<li>Digital Sanitization (Scrubbing): Write 0s to the memory in page-level deletion strategy.</li><br>
							<li>Analog Sanitization: Perform repeated partial erase operation until completely erased.</li>
						</ul>
					</p>
				</div>
				<hr>

			<button type="button" class="collapsible"> <a name="evanescoM"></a><i>Architectural Support for Sanitization: Introduction and Motivation</i></button>
				<div class="content">
					<p>
						The content here is derived from the paper that can be found at <a href="https://people.inf.ethz.ch/omutlu/pub/evanesco-secure-data-sanitization-for-flash-memory_asplos20.pdf">this link.</a>
					</p>		
					<p>
						It is challenging to erase a file without large performance penalty or reliability issue in modern  NAND. Evanesco is a new technique for high-density 3D NAND flash memory. Evanseco, instead of physically destroying data, blocks access to the data. Two commands, pLock and bLock are designed that block access to the page and block respecively of the data deleted. These locked memories can only be accessed after erase operation, thus the claim is that a strong security is guaranteed. Erase or program based data deletion technique will rapidly reduce the quality of the flash memory.
					</p>	
					<p>
						Performance analysis done on 160 3D TLC NAND memory on FlashBench with their proposal enabled flash model. Benchmarks are workloads collected from enterprise servers and mobile systems. 
					</p>
					<p>
						In the new architecture, a read request to the locked location will always return all 0s.
					</p>
					<p>
						FTL is the special embedded software that is employeed in flash-based storage system. FTL writes in append-only fashion which means new data is stored in a new physical page for performance reasons by avoiding longer block erases. Thus a logical to physical mapping table (L2P) is maintained. For updating a data, the updated data is written to a new free page, the link in LPA table is updated to new physical address, state of new physical page changed to valide from free and the old physical page's state is made invalid.
					</p>
					<p>
						When the system is about to run out of the free pages, a Garbage Collector is invoked. This reclaims the free pages by erasing victim blocks (blocks with invalid pages). If there are valid pages in the victim blocks that are to be erased, these pages have to be copied elsewhere first, and remapped in the L2P. 
					</p>
					<p>
						The trend of having a multple copies of a file while updating (or deleting) a file is called <i>data versioning problem</i> in this paper. Experiment to measure how many invalid versions of a file exist throughout the lifetime of a file is conducted. Tools used are: VerTrace, an extension of IOPro and FlashBench (they made all three of these tools). Essentially, the goal is to keep track of number of valid and invalid pages of a file at any time in the flash memory. Three benchmarks traces, Mobile, MailServer and DBServer are used. Maximum storage emulated is 16GiB. Ultimate goal of the evaluation is to find the number of invlaid versions of file that exists and for how long the invalid versions remain.
					</p>
					<p>
						<ul>
							Types of files:
							<li>Uni-version: File when modified such that content of older version is a subset of new file ie file at time t is subset of file at time t+1.</li>
							<li>Multi-version: Any file that is not Uni-version file.</li>
						</ul>
						<ul>
							Quantities measured for each files are:
							<li>Version Amplification Factor: Maximum of (for all time of observation (ratio of invalid pages to valid pages))</li>
							<li>Total Insecure Interval: Time for which the number of invalid pages is greater than 0.</li>
						</ul>
						The observation show that files have a large number of invalid pages for a long time.
					</p>
					<p>
						Existing technique destroy data by changing Vth of flash cell. Scrubbing technique increases Vth of all flash cell in WL so that Vth distrubutions of different states are mixed together such that original data identification is impossible. 
					</p>
					<p>
						But this technique is not efficient in MLC or TLC NAND as there are multiple pages in  a single Wordline. Thus efficient reprogramming based sanitization technique is proposed for MLC NAND memory. This technique uses one-shot programming technique with lowered voltage such that content can be safely destroyed, and other pages in the same wordline is not impacted. Zero copy overhead is incurred as there is no copy operation needed. Sanitization of LSB and MSB can be independent. But there is always chance of overprogramming, so that the shift of Vth is too much.
					</p>
				</div>
				<hr>

			<button type="button" class="collapsible"> <a name="evanescoP"></a><i>Architectural Support for Sanitization: Proposal</i></button>
				<div class="content">
					<p>
						The content here is derived from the paper that can be found at <a href="https://people.inf.ethz.ch/omutlu/pub/evanesco-secure-data-sanitization-for-flash-memory_asplos20.pdf">this link.</a>
					</p>
					<p>
						The access to the page is controlled by access-permission (AP) flag. Two kinds of AP flags: pAP (for page) and bAP (for block) inside the flash chip that are controlled by two commands pLock and bLock respectively. Command pLock&ltppn&gt locks the physical page number ppn by setting the pAP flag to disable state. Similarly, bLock&ltpbn&gt blocks the access to the physical block number pbn by setting bAP flag to disabled state. No unlock commands are present and unlock is automatically done once the block is erased. Thus data is permanently inaccessible before erase cycle once locked. The logic is implemented into the flash chip itself.
					</p>
					<p>
						<b>Page Level Sanitization:</b> For each page, along with the main area of data storage, some spare storage space for flags storage is available. The flags are stored in this location. For every access made to a page, the data copied to the cache or buffer is only sent to output if the flag value pAP is not disabled. Since in MLC and TLC, multiple flags are needed in each WL. For individual programming of these cells, self-boost program inhibit (SBPI) is used that allows flash cells in a single WL to be selectively programed by choosing different voltage settings for different BLs.
					</p>
					<p>
						<b>Block Level Sanitization:</b> If a large number of pages are to be sanitized, PLock becomes non-trivial and incurs overhead. A single block sanitization can sanitize a large number of pages at once. This only works in 3D as the bAP flag is implemented in SSL (Source Select Line) of 3D NAND as 3D NAND use normal flash cell (that allows for SSL to act as WL) for SSL rather than transistor. 
					</p>
				</div>
				<hr>

			<button type="button" class="collapsible"> <a name="forensicNAND"></a><i>Improving the Reliability of Chip-Off Forensic Analysis of NAND Flash Memory Devices</i></button>
				<div class="content">
					<p>
						The original paper can be found at <a href="https://people.inf.ethz.ch/omutlu/pub/flash-memory-chip-off-forensics-reliability_dfrws17.pdf">this link</a>
					</p>
					<p>
						This paper deals in opposite to what sanitization is about, as it discusses the techniques to extract data from a NAND flash memory that is acquired as part of some investigation.
					</p>
					<p>
						When the chips are obtained intact in the circuit, a method called chip-off is used to remove the chip from the cicruit using some heating technique.
					</p>
					<p>
						Two major observations made in the experiments are:
						<ul>
							<li>From the time device is seized and time digital forensic investigators perform data extraction, a large number of errors are introduced (through charge leakage: data retention errors)</li>
							<li>Using thermal based methods, the errors in the flash memory increases by two or more orders since the application of temprature increases the leakage.</li>
						</ul>
					</p>
					<p>
						This experiment explores a hardware based approach: fine grained read reference voltage control mechanism implemnted in modern NAND, called read retry. 
					</p>
					<p>
						Read retry can compensate change leakage that occurs due to retention loss and thermal based chip removal.
					</p>
					<p>
						Citation number [7], paper found at <a href="http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.444.4809&rep=rep1&type=pdf">this link</a> discusses reverse engineering the content of NAND memory. Thesis at <a href="https://apps.dtic.mil/dtic/tr/fulltext/u2/a509258.pdf">this link</a> also discusses along the same lines.
					</p>
					<p>
						Modern NAND flash have some form of read-retry methods ie by dynamically adjusting read reference voltages, in a fine grained manner.
					</p>
					<p>
						The papers mentions that the flash controllers tries read operation with a number of reference voltages. However, the exact detail of read retry operation is not made public by the manufacturers as in the paper.
					</p>
					<p>
						RBER increases with retention, but with read-retry method of reading the error decreases. But read retry modeA only reduces the error at high PE cycle counts. And most of the times it is not beneficial as it is not controllable and visible.
					</p>

				</div>
				<hr>
		<button type="button" class="collapsible"> <a name="dataDeletion"></a><i>Investigation of Data Deletion Vulnerabilities in NAND Flash Memory Based Storage</i></button>
				<div class>
					<p>
						<ul>
							<li>
								The paper can be found at <a href="https://arxiv.org/pdf/2001.07424.pdf"> this link.</a>
							</li>
							<li>
								NAND suffers from data security vulenrabilities.
							</li>
							<li>
								This paper deals in unreliable deletion in commercial NAND memory.
							</li>
							<li>
								Some observation of this research has led to 100% data recovery.
							</li>
							<li>
								This paper makes four categories of reasons for deletion based vulnerabilities for NAND memory:
								<ul>
									<li>
										Physics Based:
										<ul>
											<li>
												Data deletion vulnerabilties is due to the storage in Flash being charge based.
											</li>
											<li>
												During erase operation all the charge carriers may not be de-trapped. This leads to charge remanence post-erase operation.
											</li>
											<li>
												These remenants charges can be detected using SEM, SPM microscopy techniques and using emperical models the previous state of flash cell can be predicted.
											</li>
										</ul>
									</li>
									Performance Based:
									<ul>
										<li>
											Several performance optimization tricks lead to data deletion vulnerabilities.
										</li>
										<li>
											Even partial programming and erase operation is sensed as complete programming and erase operation due to optimal placement of threshold voltage. This causes remanent charge effect as discussed above.
										</li>
										<li>
											High retention of flash ensures that charge is trapped post-erase operation as well. This may enhance the probability that the information can be recovered even after a long time
										</li>
										<li>
											Due to wear leveling, same physical address may be provided to different physical address. However, the previous written data may still be present.
										</li>
									</ul>
									<li>
										System Complexity Based:
										<ul>
											<li>
												OS instead of deleting still has chance of recovering data. Even after deletion, data may still be in flash memory.
											</li>
											<li>
												Software based recovery tools can be used to recover data using info in the metadata of the file. Even formatting the media does not ensure complete deletion.
											</li>
										</ul>
									</li>
									<li>
										Intentional Reasons
										<ul>
											<li>
												Backdoors and Trojans
											</li>
											<li>
												Side Channel
											</li>
											<li>
												Redundant Data
											</li>
										</ul>
									</li>
								</ul>
							</li>
							<li>
								Data can be recovered from flash memory using different ways as below:
								<ul>
									<li>
										Using software recovery tool
									</li>
									<li>
										Using comprehensive embedded test env
									</li>
									<li>
										by charge-charge constrast measurement of flash devices
									</li>
								</ul>
							</li>
							<li>
								Basic attributes of files are present in the files or folders that vary based on the file system in use. Even record of deleted files are stored.
							</li>
							<li>
								Scanning that a file is deleted and reading the MFT, Master File Table, data can be extracted.
								<ul>
									<li>
										In NTFS, there is a special attribute for deleted files/folders
									</li>
									<li>
										In FAT, ASCII 0xE5 is used to mark deleted files/folders.
									</li>
									<li>
										The first 42 bytes in MFT contains file record header in which Flag field (2 Bytes) provides info about whether file is deleted or still in use. (if 1, means still in use)
									</li>
									<li>
										MFT contains other informations as well.
									</li>
									<li>
										Read http://www.ntfs.com/file-recovery-concepts.htm
									</li>
								</ul>
							</li>
							<li>
								Experiments performed in commercial NAND Flash USB. OS and device vendors do not delete conclusively and do not reliably sanitize data was the conclusion.
							</li>
							<li>
								Expt performed in Windows 10 and Linux Ubuntu 14.04
							</li>
							<li>
								USB media formatted to both FAT32 and NTFS. 
							</li>
							<li>
								<i>Recuva</i> and <i> EaseUS </i> recovery software used.
							</li>
							<li>
								Commands explored were: <i>Shift+Delete</i> and <i>Format Option</i> in Windows while <i>rm -rf</i> in Linux and <i>shred</i> command.
							</li>
							<li>
								Data were recovered successfull in all cases except <i>shred</i> in Linux.
							</li>
							<li>
								Experiments performed in two settings:
								<ul>
									<li>
										Dump Recovered Data to Another storage Media:
										<ul>
											<li>
												Recovered Successfully
											</li>
											<li>
												Deleted again using quick format option
											</li>
											<li>
												data again recovered using recovery software using <i>deep scan</i>
											</li>
											<li>
												Full format tried on recovered data
											</li>
											<li>
												Data unable to recover, but because of the limitation of the recovery software used.
											</li>
										</ul>
									</li>
									<li>
										Dump Recovered Data to Same Storage Media:
										<ul>
											<li>
												Similar delete-recover-delete option tried
											</li>
											<li>
												With each iteration, recovered data quality degraded.
											</li>
										</ul>
									</li>
								</ul>
							</li>
						</ul>
					</p>
				</div>
				<hr>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="ssd_papers_study">SSD papers</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: SSD papers</i></button>

<div>
<p>
	<button type="button" class="collapsible"> <i>List of Papers Studied</i></button>
			<p>
				Following is the list of paper studied about the SSDs.
				<ul>
					<li>
						<a href="#maneas1">A Study of SSD Reliability in Large Scale Enterprise Storage Deployments</a>
					</li>
					<li>
						<a href="#openssd">A Study of OpenSSD</a>
					</li>
				</ul>
			</p>
			<hr>

		<button type="button" class="collapsible"> <a name="maneas1">Paper: <i>A Study of SSD Reliability in Large Scale Enterprise Storage Deployments</i></a></button>
			<div class="content">
				<p>
					<li>
						The original paper can be found at <a href="https://www.usenix.org/system/files/fast20-maneas.pdf">this link</a>. This paper was presented in USENIX FAST2020 and was awarded best paper.
					</li>
					<li>
						Large scale study focused on enterprise storage systems, study conducted on 1.4 million SSDs of NetApp which is a major storage vendor.
					</li>
					<li>
						Netapp storage system employ WAFL file systems and DATA ONTAP operating system which uses software RAID to provide resiliency against drive failures.
					</li>
					<li>
						Data over the nw is serviced using file-based protocols as NFS and CIFS/SMB or block-based protocols such as iSCSI
					</li>
					<li>
						Netapp systems in field send weekly NetApp Active IQ bundles that track a  very large set of system and device parameters. This study is based in mining this collection of NetApp Active IQ messages.
					</li>
					<li>
						Different types of failures are categorized. The most severe of them that prompts replacement of drives was SCSI error. These error are due to ECC errors. Majority of other errors were recovered by RAID reconstruction.
					</li>
					<li>
						Replacement rate: number of device failures divided by number of device years.
					</li>
					<li>
						For the causes of errors or the factors impacting replacement rates:
						<ul>
							<li>
								Usage and Age: A increasing failure rates of long period 12-15 months in the beginning, and 6-12 months of decreasing failure rates before finally stabilizing for 3D-TLC and eMLC (enterprise MLC) drives.
							</li>
							<li>
								3D-TLC failure rates is higher than other types, thus the replacement reateis higher. Also 3d-TLC uses 10-15X times more for spare blocks
							</li>
							<li>
								Higher capacity drives have higher replacement rates and more severe failures too (unresponsive drives).
							</li>
							<li>
								High density drives have higher replacment rates.
							</li>
							<li>
								EMphasize importance of firmware updates
							</li>
							<li>
								Most of the systems make use of only 15% of rated life of device, thus there should be no concern to updatd to QLC.
							</li>
						</ul>
					</li>
				</p>
			</div>

			<hr>

		<button type="button" class="collapsible"> <a name="openssd">A Study of OpenSSD</a></button>
			<div class="content">
				<p>
					<ul>
						<li>
							The content is derived from <a href="http://www.openssd.io/">this link.</a> The tutorial on use of this platform is at <a href="https://github.com/Cosmos-OpenSSD/Cosmos-plus-OpenSSD/blob/master/doc/Cosmos%2B%20OpenSSD%202017%20Tutorial.pdf">this link</a>
						</li>
						<li>
							It is a open source solid state drive project designed for research.
						</li>
						<li>
							In addition to the software, they have a read hardware board called Cosmos OpenSSD platform (consisting Zynq platform).
						</li>
						<li>
							The full design includes Controller Firmware, Controller Hardware along with other modules like Controller Processor and Performance Monitor.
						</li>
						<li>
							Earlier version supported PCIe (AHCI) while the newer platform Cosmos+ supports NVMe interface.
						</li>
						<li>
							BCH ECC implementation.
						</li>
					</ul>
				</p>
			</div>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="sram">SRAM</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: SRAM</i></button>

<div>
<p>
	
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="dram">DRAM</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: DRAM</i></button>

<div>
<p>
	
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="sram_dram_security">Security Primitives in SRAM and DRAM</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Security Primitives in SRAM and DRAM</i></button>

<div>
<p>
	
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="fram">FRAM</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: FRAM</i></button>

<div>
<p>
	<p>
		FRAM or FeRAM is a class of memory that achieves non-volatility with a structure similar to DRAM. The capacitor in the DRAM memory has a dielectric layer that needs to periodic recharge to hold the charge, thus making DRAM a volatile memory while it is replaced by Ferroelectric material in FRAM. Presence of this Ferroelectric material (typically Lead Zirconate Titanate, PZT) makes FRAM non-volatile <a href="https://en.wikipedia.org/wiki/Ferroelectric_RAM">(ref)</a>.
	</p>
	<hr>
	<p>
		Writing to the FRAM is peformed in a process similar to storing charge in capacitors. A field is applied across the dielectric by charging the plates on either sides of it. This causes the atoms inside the dielectric to orient in "up" or "down" orientation, thus storing '1' or '0'. This change in polarity because of application of electric field produces a power-efficient binary switch <a href="https://www.cypress.com/products/f-ram-nonvolatile-ferroelectric-ram">(ref)</a>. The ferroelectric crystal and its orientation is not affected by magnetic field.
	</p>
	<hr>
	<p>
		<b>Destructive Read</b>: Reading in FRAM is done through the transistor in 1T-1C structure. The transistor forces the cell into a particular state. For example let us assume the transistor forces the cell into a '0' state. If the cell was already in '0' state, there is no change in the output lines. However, if the cells was in a '1' state, this reorientation of atoms in the dielectric causes a brief pulse of current in the output lines are the electrons are pushed out. Detection of this pulse indicates there was a '1' in FRAM cell. This process causes the original value to be overwritten, thus the reading process is desctructive and needs to be re-written.
	</p>
	<hr>
	<p>
		<b>Ferroelectric Property</b>: The dielectric or the ferroelectric material has a crystal structure. The PbZT crystal is organized in a <i>perovskite</i> structure where the Lead atoms are at the outermost layer and O atom inside them. The Z/Ti atom in the center form the cat-ion and has two equal and low-energy states. By the application of the electric field, the cat-ion will move in the direction of the field applied. This causes the low-energy state to be aligned in the direction of applied field while and conversely high-energy state will be aligned in the opposite direction. THe mnovement of cation causes either "up-polarization" or "down polarization" thus creating the two states <a href="https://www.cypress.com/file/46186/download">(ref)</a>.
	</p>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="rram">RRAM</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: RRAM</i></button>

<div>
<p>
	<p>
			Resistive RAM or ReRAM works in the fashion similar to PCM <a href="https://en.wikipedia.org/wiki/Resistive_random-access_memory">(ref)</a>, i.e. by changing the resistance of the dielectric solid-state meterial. Here a metal oxide is sandwiched between two metal electrodes <a href="https://dl.acm.org/doi/pdf/10.1145/2463209.2488867">(ref)</a>. RRAM works by creating defects in the oxide layer to create a filmaent or a condictive path by the application of high voltage.
		</p>
		<hr>
		<p>
			The filament can be broken resulting in a high resistance state and can be formed to result in a low resistance state.
		</p>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="mram">MRAM</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: MRAM</i></button>

<div>
<p>
	<button type="button" class="collapsible"><i>Introduction </i></button>
		<div>
			<p>
				MRAM stands for Magnetoresistive Random Access Memory. Unlike other conventional memory technology where information is stored in the form of presence or absence of charge particles or current flows, information in MRAM is stored by magnetic storage elements.
			</p>
			<p>
				Structurally, MRAM consists of MRAM cells just like any other memory. Each MRAM cell contains two magnetic plates that are separated by an insulating layer. These plates are ferromagnetic in nature where one of the plates in permanent while the magnetization on other can be changed by influence of external applied field.
			</p>
			<p>
				Based on the magnetization of the two plates being in parallel or antiparallel direction, the resistance of the MRAM cell is different. Measure of this electric resistance is the reading process. If the two plates have the same magnetic alignment, the logic state is said to be '1' while if the two plates have different magnetic alignment, the logic state is said to be '0'. (In parallel orientation, the likelihood of electron tunneling through the insulating layer is more than if they are in antiparallel orientation. Thus in parallel orientation, resistance offered is low and is read as 1 and vice versa.)
			</p>
			<p>
				The two layers or plates can be referred to as (i) free layer, and (ii) fixed layer.
			</p>
		</div>
		<hr>
		<h3>Magnetic Tunnel Junction</h3>
		<p>
			A structure that consists of two layers of magnetic metals that is separated by an insulating layer in between them is the Magnetic Tunnel Junction. Here, the insulating layer between the magnetic layers is very thin that allows electrons to tunnel through one of the magnetic layers to another if a bias voltage is applied between the electrods.
		</p>
		<hr>
		<h3>Tunnel Magnetoresistance</h3>
		<p>
			The current in the tunnel, tunneling current, can be changed based on the orientation of the magnetic plates relative to each other. This change in the current based on the magnetic alignment of the plates is called Tunnel Magnetoresistance.
		</p>
		<hr>
		<h3>Toggle MRAM</h3>
		<p>
			The structure consists of two magnetic layers and an insulating layer in between. When a bias is applied, the electrons that are spin polarized by the magnetic layers tunnel across the dielectric. 
		</p>
		<hr>
		<h3>STT MRAM</h3>
		<p>
			A newer technique called Spin Torque Transfer MRAM that uses Spin Torque Transfer technology. Under the Spin Torque Transfer technology, an electric current that is unpolarized (unpolarized current contains 50% electrons in each orientation) is passed through the fixed layer. This causes the electric current to be spin polarized and thus creates spin-polarized current.
		</p>
		<p>
			When this spin-polarized current is directed into the free layer, the angular momentum can be transferred to this layer. This changes the orientation of the free layer. Here the torque developed while changing the orientation of the current is transferred to the free layer. This creates parallel orientation between the fixed and free layer. If it is required to change the orientation from parallel to antiparallel, the current direction is reversed. Here, the electrons are sent from the free layer to the fixed layer; majority electrons pass to the fixed layer while a minority electrons are reflected; the reflected electrons transfers the angular momentum to the free layer thus changing the orientation to antiparallel. 
		</p>
		<hr>
		<h3>Spin Orbit MRAM</h3>
		<p>
			SOT MRAM is a more recent kind of MRAM technology where an additional layer of heavy metal is attached to the free layer. When a current is passed through the heavy metal layer, a spin polarized current is created in direction perpendicular to the unpolarized current. This transfers its angular momentum to the free layer thus performing switching operation on free layer <a href="https://doi.org/10.1016/j.sysarc.2018.11.005">(ref)</a>. The basic principal is based on the idea that a charge current in a heavy metal creates a spin current in a transverse direction or Hall Effect <a href="10.1109/JXCDC.2017.2762699">(ref)</a>. While this is still a under-development technology, researcher at ETH Zurich have reported writing operations in the range of 100 picoseconds <a href="https://scitechdaily.com/next-generation-magnetic-memory-breakthrough-writing-data-in-under-a-nanosecond/">(ref)</a> and <a href="https://www.nature.com/articles/s41565-019-0607-7">(ref)</a>.
		</p>
		<hr>
		<h3>Write Disturb in MRAM</h3>
		<p>
			Writing to the memory is performed by changing the magnetic orientation of the plate whose magnetization can be changed. Conventional method of writing requires substantial amount of current on the word and bit lines (the fields being perpendicular to each other). Another negative point of MRAM is the induced field might effect multiple cells, thus limiting the scaling below 100nm <a href="https://doi.org/10.1016/j.mattod.2017.07.007">(ref)</a>. This problem where the field overlaps multiple cells causes <b>write disturb</b> in MRAM.
		</p>
		<hr>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="pcm">PCM</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: PCM</i></button>

<div>
<p>
	<p>
			Phase Change Memories(PCM) utilize the unique propoerty of some material to remain in multiple states as the basics of information storage. Since the change induced in the state remain until the state is again changed to some other state, or reverted back, the storage is non-volatile.
		</p>
		<hr>
		<p>
			The material that is commonly used is <i>Chalcogenide glass</i>, which can stay in amorphous as well as crystalline solid state. In older generation of PRAM, the state of the Chalcogenide glass was changed by production of heat by passing current through a heating element. This heating element is made of TiN, and was used to change the state of the glass to amorphous solid state by quick application (heated over 600 celcuis). Holding the heat for some time in the crystallization temperature range (cooled) would switch the state back to the crystalline state. Alternative technologies for changing states using laser have been proposed in research <a href="https://doi.org/10.1038/nnano.2011.96">(ref)</a>.
		</p>
		<hr>
		<p>
			Amorphous state offers more resistance and thus represents a binary 0 state while crystalline state represents a binary 1 state.
		</p>
		<hr>
		<p>
			The material used in the PCM can exhibit multiple intermediate states between the amorphous and crystalline solid state. This allows for multiple bits storage in a single cell of the PCM. Intel and ST Microelectronis are currently working on a design with four states, two of which are partially crystalline state in addition to the previous two states. <a href="https://en.wikipedia.org/wiki/Phase-change_memory">(ref)</a>.
		</p>
		<hr>
		<p>
			PRAMs are highly temperature sensitive as data can be lost with application of high temperature. However, they offer fast write times and have an endurance of about 100 million write cycles.
		</p>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="optane">Optane</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Optane</i></button>

<div>
<p>
	<button type="button" class="collapsible"><i>List of Papers Studied </i></button>
		<div>
			<p>
				Following are the papers studied:
				<ul>
					<li>
						<a href="#3dxpwiki">3D X Point From Wikipedia</a>
					</li>
					<li>
						<a href="#optaneperformance">Platform Storage Performance With 3D XPoint Technology</a>
					</li>
					<li>
						<a href="#optaneevolution">Early Evaluation of Intel Optane Non-VolatileMemory with HPC I/O Workloads</a>
					</li>
					<li>
						<a href="#optanecontract">Towards an Unwritten Contract of Intel Optane SSD</a>
					</li>
				</ul>
			</p>	
		</div>
		<hr>

		<button type="button" class="collapsible"> <a name="3dxpwiki">3D X Point From Wikipedia</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The file can be found at <a href="https://en.wikipedia.org/wiki/3D_XPoint">this link</a>
					</li>
					<li>
						NVM technology developed by Intel and Micron. Available under name Optane by Intel and Quantx by Micron.
					</li>
					<li>
						Bit storage is based on change of bulk resistance. But differs from other PCMs in that it uses chalcogenide materials for both selector and storage parts. (3DXPoint is thought to be a subset of ReRAM also)
					</li>
					<li>
						Full technical detail is not revealed by Intel or Micron. But it has been stated to use electrical resistance and is not based on electrons (like flash etc)
					</li>
				</ul>
			</p>
		</div>
		<hr>

		<button type="button" class="collapsible"> <a name="optaneperformance">Platform Storage Performance With 3D XPoint Technology</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The file can be found at <a href="https://ieeexplore.ieee.org/document/8003284">this link</a>
					</li>
					<li>
						This paper reviews the potentialities on computing introduced by the 3-D XPoint technology in changing the memory-storage hierarchy.
					</li>
					<li>
						Work on 3DXPoint started in 2008.
					</li>
					<li>
						Three use cases of 3D X Point:
						<ul>
							<li>
								Storage: disk operation
							</li>
							<li>
								DRAM extension:
							</li>
							<li>
								Persistant Memory:
							</li>
						</ul>
					</li>
					<li>
						read later
					</li>
				</ul>
			</p>
		</div>
		<hr>

		<button type="button" class="collapsible"> <a name="optaneevolution">Early Evaluation of Intel Optane Non-VolatileMemory with HPC I/O Workloads</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The file can be found at <a href="https://arxiv.org/pdf/1708.02199.pdf">this link</a>
					</li>
				</ul>
			</p>
		</div>
		<hr>

		<button type="button" class="collapsible"> <a name="optanecontract">Towards an Unwritten Contract of Intel Optane SSD</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The file can be found at <a href="https://research.cs.wisc.edu/adsl/Publications/hotstorage-contract19.pdf"> this link</a>
					</li>
					<li>
						This paper describes Optane as well as formalized an unwritten contract of the Optane SSD violating which will result in 11x worse read latency and limited throughput.
					</li>
					<li>
						Optane is Intel's 3D X Point memory with Micron.
					</li>
					<li>
						It is available in various form factors including as a caching layer between the DRAM and block-device (as Optane memory), as a block device itself (as Optane SSD) and Optane DC Persistent Memory.
					</li>
					<li>
						Of all the above, Optane SSD is the most cost effective and widely available option.
					</li>
					<li>
						Following are the six (seven) rules for immediate performance of Optane SSD:
						<ol>
							<li>
								Uses should issue small requests for low latency (>=4KB) and keep a small number of outstanding IOs (Access with Low Request Scale rune)
							</li>
							<li>
								Optane SSD should not consider sequential workloads unlike HDD/Flash SSDs. (Random Access is OK)
							</li>
							<li>
								Avoid parallel access to a single chunk to avoid contention among requests (Avoid Crowded Access rule)
							</li>
							<li>
								Control the overall load of both reads and writes to control for optimal latency.
							</li>
							<li>
								Never issue request for less than 4KB to exploit bandwidth of Optane SSD
							</li>
							<li>
								Requests issued to Optane SSD should align to eight sectors for best latency.
							</li>
							<li>
								When serving sustained workloads, there is no cost of garbage collection in Optane SSD. (Forget garbage collection)
							</li>
						</ol>
					</li>
				</ul>
				<ol>
					<li>
						Access with Low Request Scale:
						<ul>
							<li>
								Optane claimed to be upto 1000x faster than SSDs. But is it?
							</li>
							<li>
								Optane SSD users should issue small requests and maintain small number of outstanding IOs.
							</li>
							<li>
								Needed to extract low latency but also to exploit full bandwidth of Optane SSD
							</li>
							<li>
								Variables of analysis are (i) request size and (ii) queue depth. Analysis of read and writes are performed.
							</li>
							<li>
								For read operation, large request size and large queue depth  does not work better with optane SSD.
							</li>
							<li>
								For write operations, they are almost comparable in all the cases while optane still being poor for large request size and large queue depth.
							</li>
							<li>
								Optane internally uses RAID like organization of memory dies.
							</li>
							<li>
								The interleaving degree (number of channels) of the Optane and SSD are examined through experiments (ref 18 and 19) and found out to be 7 and 128 respectively. The info of 7 channels is also found in hardware description ( ref 3).
							</li>
							<li>
								This shows Optane has limited internal parallelism.
							</li>
							<li>
								The limited parallelism is one reason that Optane performs better with small number of queue depth.
							</li>
						</ul>
					</li>
					<li>
						Random Access is OK
						<ul>
							<li>
								In SSD and HDD, better performance is seen with sequential access than random access. But, Optane is random access block device.
							</li>
							<li>
								Experiments were performed for this for both SSD and Optane based system to prove.
							</li>
							<li>
								Flash SSD performs better on sequential while Optane has comparable performance for read operations.
							</li>
							<li>
								For smaller sized requests, Optane actually favors random writes over sequential writes. Similarly, flash SSD favors sequential writes only for small request size while in other cases, they are comparable.
							</li>
							<li>
								Optane prefers random access because of the ability to perform in-place updates in 3D X Point memory. In Optane, there is no difference in address translation cost for random versus sequential workloads.
							</li>
						</ul>
					</li>
					<li>
						Avoided Crowded Accesses
						<ul>
							<li>
								The client should not issue parallel access to a single chunk in Optane based system because Optane SSD contains shared resources.
							</li>
							<li>
								Experiments performed to study the performance by issuing parallel requests for different sector within same chunk.
							</li>
							<li>
								Latency increases with increase in queue depth.
							</li>
						</ul>
					</li>
					<li>
						Control Overall Load
						<ul>
							<li>
								For optimal latency in Optane SSD, client must control overall load of both reads and writes
							</li>
							<li>
								Observation derived from performance of Optane serving mixed reads and writes. In the experiment, we issue random 4KB requests, varying the percentageof writes from 0% to 100%, with QD= 64 (large enough toachieve full throughput for both Optane SSD and Flash SSD).
							</li>
							<li>
								In Optane, reads and writes are treated equally. Latency plot overlaps for the plots with different ratio of writes/total access.
							</li>
							<li>
								Latency was not a function of writes vs reads, but is dependent upon other factors of overall load.
							</li>
							<li>
								For flash SSD, write operation increases the latency.
							</li>
						</ul>
					</li>
					<li>
						Avoid Tiny Accesses
						<ul>
							<li>
								Client must not issue less than 4KB request.
							</li>
							<li>
								Latency might be same for small requests but for maximizing throughput it is better to issue 4KB request.
							</li>
						</ul>
					</li>
					<li>
						Issue 4KB Aligned Requests
						<ul>
							<li>
								For best latency, request should be aligned to 8 sectors that is 4KB.
							</li>
							<li>
								In the experiment,we measure the latency of individual read requests (QD= 1);each read is issued to a position A+offset, where A is a random position aligned to 32KB and offset is a 512-byte sector within that 32KB.
							</li>
							<li>
								A periodic latency observation is made which shows Optane favor aligned requests.
							</li>
							<li>
								Best is when request are aligned to 8 sectors
							</li>
						</ul>
					</li>
					<li>
						Forget Garbage Collection
						<ul>
							<li>
								No need to worry about garbage collection of Optane
							</li>
							<li>
								In Flash SSDs, after the device is full, any further writes is slow because it triggers garbage collection. But the write latency foe Optane is sustained which shows garbage collection has no cost.
							</li>
							<li>
								Optane SSD has LBA_based mapping. I have to study this myself later on.
							</li>
							<li>
								Flash performs log-structured based file system thus the throughput pattern occurs when we read according to written order.
							</li>
						</ul>
					</li>
				</ol>
			</p>
		</div>
		<hr>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="file_system_study">Study of File Systems</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Study of File Systems</i></button>

<div>
<p>
	<button type="button" class="collapsible"> <i>List of Papers Studied</i></button>
			<p>
				Following is the list of paper studied about the SSDs.
				<ul>
					<li>
						<a href="#ext4">The new ext4 filesystem: current status and future plans</a>
					</li>
					<li>
						<a href="#matias1">From Open-Channel SSDs to Zoned Namespaces</a>
					</li>
					<li>
						<a href="#zoned1">NVMe Zoned Namespaces</a>
					</li>
					<li>
						<a href="#parallelfirmware1">Scalable Parallel Flash Firmware for Many-Core Architectures</a>
					</li>
					<li>
						<a href="#flashsearch1">Read as Needed: Building WiSER, a Flash Optimized Search Engine</a>
					</li>
				</ul>
			</p>
			<hr>

		<button type="button" class="collapsible"> <a name="ext41">Paper: <i>The new ext4 filesystem: current status and future plans</i></a></button>
			<div class="content">
				<p>
					<ul>
						<li>
							The original paper can be found at <a href="https://pdfs.semanticscholar.org/b8dd/ec47f9fab1eddb5c9cacf703781dd5337b87.pdf"> this link</a>
						</li>
					</ul>
				</p>
			</div>
		
		<hr>

		<button type="button" class="collapsible"> <a name="matias1">Presentation: <i>From Open-Channel SSDs to Zoned Namespace</i></a></button>
			<div class="content">
				<p>
					<ul>
						<li>
							This is a presentation that can be found at <a href="https://www.usenix.org/sites/default/files/conference/protected-files/nsdi19_slides_bjorling.pdf">this link</a>. It was presented by Matais Bjorling, Director Solid-State System Software of Western Digital at USENIX 2019.
						</li>
						<li>
							Presents forward-looking statements regarding Solid State techs, product development efforts, software development etc
						</li>
						<li>
							Open Channel SSD architectures:
							<ul>
								<li>
									No FTL present but leaves the management of device to OS. Linux 4.4 kernel supports Open Channel using NVM Express specification.
								</li>
								<li>
									Chunks: Zoned namespace
								</li>
								<li>
									Parallel Groups: Endurance management groups.
								</li>
							</ul>
						</li>
					</ul>
				</p>
			</div>
			<hr>

		<button type="button" class="collapsible"> <a name="zoned1">Website: <i>NVMe Zoned Namespaces</i></a></button>
			<div class="content">
				<p>
					<ul>
						<ul>
							Let us First talk about zoned storage itself <a href="https://zonedstorage.io/introduction/zoned-storage/">(this link)</a>:
							<li>
								It is a class of storage devices that is divided into zones. Each zones have a write pointer that keeps track of position of next data to be written.
							</li>
							<li>
								Data in a zone cannot be overwritten and must be erased first.
							</li>
							<li>
								SSDs can also implement zoned interface to reduce write amplification, reduce the device DRAM needs and improve quaity of service at scale.
							</li>
							<li>
								Support for zoned storage was added to Linux in version 4.10.0, and is based on Zoned Block Device abstraction. The interface associated with ZBD is an extension to traditional Linux block device interface.
							</li>
							<li>
								Zoned storage is used in Shingled Magnetic Recording (SMR) which is different from Conventional Magnetic Recording (CMR) used in hard disk drives. SMR provides more areal density than CMR because SMR removes the gaps that are placed in CMR.
							</li>
							<li>
								CMR places gaps in tracks of HDDs to account for Track MisRegistration (TMR) budget. This impacts the density negatively as the area is not fully utilized. SMR removes needs for gaps by writing tracks in overlapping manner, forming pattern similar to shingles on a roof.
							</li>
							<li>
								While performing a write, the write head (which area bit wider than conventional ones) overlap part of another track thus leaving a narrower track band for reading.
							</li>
							<li>
								The overlapping tracks are gouped into bands called zones of fized capacity for more effective data organization and partial update capacity.
							</li>
							<li>
								Gaps are however present between zones to prevent overwrite between zones.
							</li>
							<li>
								Because of the shingled organization, data has to be written sequentially. To overwrite, entire band or zone has to be re-written.
							</li>
						</ul>
						<li>
							The original content can be found at <a href="https://zonedstorage.io/introduction/zns/">this link</a>.
						</li>
						<li>
							Zoned namespace divides the logical address space of a namespace into zones.
						</li>
						<li>
							Each zones provides an LBA range that must be written sequentially and if written again must be explicitly reset.
						</li>
						<li>
							It introduces a new type of NVMe drive that has several benefits over traditional NVMe SSDs.
						</li>
						<li>
							ZNS is similar to HDDs Shingled Magnetic Recording zones that must be written sequentially and explicilty reset before rewritting.
						</li>
						<li>
							The problem of log-on-log is solved naturally. Please read <a href="https://www.usenix.org/system/files/conference/inflow14/inflow14-yang.pdf">this paper</a> for log on log concept.
						</li>
						<li>
							
						</li>
					</ul>
				</p>
			</div>
			<hr>

		<button type="button" class="collapsible"> <a name="parallelfirmware1">Paper: <i>Scalable Parallel Flash Firmware for Many-core Arch (read later)</i></a></button>
			<div class="content">
				<p>
					<ul>
						<li>
							Published in FAST'20, the original link can be found at <a href="https://www.usenix.org/system/files/fast20-zhang_jie.pdf">this link</a>.
						</li>
						<li>
							NVMe allows users to exploit all levels of parallelism, but current firmware design are not scalable and have difficult time in handling high IO requests in parallel
						</li>
						<li>
							They propose DeepFlash that is manycore-based storage platform that can process more than a million IO requests in a seconds while hiding long latencies imposed by internal media.
						</li>
						<li>
							Many to many threading model that can be scaled horizontally.
						</li>
						<li>
							Implement prototype in many core processor that empolys dozens of light weight cores and perform parallel IO. Deepflash can serve around 4.5GB/s
						</li>
					</ul>
					<ul>
						Intro:
						<li>
							NVMe is designed to make use of parallelism at all levels of SSDs. NVMe provides 64K deep queues.
						</li>
						<li>
							
						</li>
					</ul>
				</p>
			</div>
			<hr>

		<button type="button" class="collapsible"> <a name="flashsearch1">Paper: <i>Read as Needed: Building WiSER, a flash optimized search engine(read later)</i></a></button>
			<div class="content">
				<p>
					<ul>
						<li>
							Published in FAST'20, the original link can be found at <a href="https://www.usenix.org/system/files/fast20-he.pdf">this link</a>.
						</li>
						<li>
							
						</li>
					</ul>
				</p>
			</div>


			<hr>
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="magnetic_disk">Magnetic Disks</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Magnetic Disks</i></button>

<div>
<p>
	
</p>
</div>

<hr>

<hr>

<h2 align="center"><a name="other_study">Others</a></h2>

<button type="button" class="collapsible"> <i>Please Click Here to Expand: Others</i></button>

<div>
<p>
	<button type="button" class="collapsible"><a href="#otherstudy"></a> <i>List of Papers Studied </i></button>
		<div>
			<p>
				Following are the papers studied:
				<ul>
					<li>
						<a href="#iopro">Improving File System Performance of Mobile Storage Systems Using a Decoupled Defragmenter</a>
					</li>
					<li>
						<a href="#flashbench">FlashBench: A workbench for a rapid development of flash-based storage devices</a>
					</li>
				</ul>
			</p>	
		</div>
		<hr>
		<button type="button" class="collapsible"> <a name="iopro">Improving File System Performance of Mobile Storage Systems Using a Decoupled Defragmenter</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The file can be found at <a href="https://www.usenix.org/system/files/conference/atc17/atc17-hahn.pdf"> this link</a>
					</li>
					<li>
						Investigate file fragment problem in mobile flash storage
					</li>
					<li>
						Observation based on real android smartphones:
						<ul>
							<li>
								Defragmentation is essential because fragmentation is recurring problem
							</li>
							<li>
								Fragmentation in flash affects differently than in HDDs. In flash, the logical and physical fragmentation is decoupled and the degradation in perf comes mostly from logical fragmentation.
							</li>
							<li>
								They propose two defragger tools: janusdL for logical and janusdP for physical
							</li>
							<li>
								the combined tool is janusd, that has similar performance to e4defrag without effecting flash lifetime.
							</li>
						</ul>
					</li>
					<li>
						Introduction:
						<ul>
							<li>
								When a file system is fragmented, it has to allocate multiple storage areas. In HDDs this causes longer seek times thus periodic defragmentation is recommended. But in flash, it is not recommended. This is because seek operation is not required in flash memories and thus fragmentation negligibly affect performance. Frequent deframentation might lead to effect in lifetime. This paper performs a systematic study of these things.
							</li>
							<li>
								Two questions in this paper: 1. How does defragmentation affect the performance? 2. How long does defragmentation last?
							</li>
							<li>
								They show that defragmentation improves the performance like app launching time by upto 52%.
							</li>
							<li>
								File fragmentation recurs after a week or so even after full file defragmentation. Since the apps update quickly, defragmentation disappears qiuckly.
							</li>
							<li>
								Regular defragger like e4defrag is not good for frequent defragmentation because it affects the lifetime of flash memory. Invoking each week may lead to reduction of life by 10% or more.
							</li>
							<li>
								Flash memories in mobiles have dcoupled fragmentation because the performance of physical IO is dependent upon the avaiilable of IO rathar than upon logical fragmentation. Regardless of how much the file is fragmented in logical layer, the device IO performance at the physical layer is unaffected. HOwever, the logically fragmented files significantly increase processing times in block IO layer and device driver because of large number of block IO requests.
							</li>
							<li>
								Thus defragmentation of logical space effictevly is needed.
							</li>
							<li>
								And, an independent physical defragger is needed as well because logical defragger would not be able to identify physical fragmented file.
							</li>
							<li>
								validated on emulated mobile flash storage: simeMMC and simUFS.
							</li>
						</ul>
					</li>
					<li>
						File Fragmentation Study
						<ul>
							<li>
								Based on 14 used android smartphones from 5 different manufacturers with 6 months use
							</li>
							<li>
								Fragmentation Metric is degree of fragmentation which they define as ratio of number of extent allocated to the file x by the number of extent required by the file. (in ext4 file format, each extent is 128 MB which is a chunk of memory unit allocated). The higher the number of extents, the higher is fragmentation.
							</li>
							<li>
								FIle fragmentation is a recurring problem especially when file system utilization is high.
							</li>
							<li>
								App launching time was longer with time and it is also when the file utlization waas high. App launching time is time when the icon is pressed to the time when the app is loaded fully.
							</li>
							<li>
								But if defragmentation is performed, large number of file are copied here and there, thus impacting the life of the flash.
							</li>
							<li>
								To understand how logical fragmentation affects overhead in system software stack, they built IOPro that can profile complete Android stack from app level to device driver level. Time spent at different layers in the stack can be measured using IOPro.
							</li>
							<li>
								For this research they measured time spent in Ext4 file system, page cache, block layer, the device driver and mobile flash storage.
							</li>
							<li>
								Observation: Time spent in   block layer, device driver layer, the flash storage layer increases with DOFlogical. Whle time spent in file system layer and page cahce layer are barely affected.
							</li>
						</ul>
					</li>
					<li>
						Design and Implementation of Defragmentation Tool:
						<ul>
							<li>
								<i>study later</i>
							</li>
						</ul>
					</li>
				</ul>
			</p>
		</div>
		<hr>
		<button type="button" class="collapsible"> <a name="flashbench">FlashBench: A workbench for a rapid development of flash-based storage devices</a></button>
		<div class="content">
			<p>
				<ul>
					<li>
						The file can be found at <a href="https://ieeexplore.ieee.org/document/6380706"> this link</a>
					</li>
					<li>
						Performance of NAND decreases with scaling, thus to counter the demerits, a cross-layer system level approach that employs system level approach is emerging as a promising solution. eg Data Compression and high speed accelerator etc considering the endurance is dependent upon the amount of data written to flash.
					</li>
					<li>
						To make use of changes in hardware, the software modules shold also be redesigned eg. address mapping, garbage collection etc.
					</li>
					<li>
						Data deduplication can be used to reduce the amount of data written to flash.
					</li>
					<li>
						Cross system layer solution is good in handling the challenges of newer NAND memory but it increases the complexity and cost of design.
					</li>
					<li>
						In this paper, they present a flash storage development environment called FlashBench
						<ul>							
							<li>
								Flashbench provides multiple design environments at different design levels	
							</li>
							<li>
								Flashbench enables rapid development and validation of tightly coupled sw/hw modules by providing high level sw/hw dev platform
							</li>
							<li>
								FLashbench helps to implement flexible and portable software modules.
							</li>
						</ul>
					</li>
					<li>
						On the hardware side, they have three kinds of models:
						<ul>
							<li>
								Abstract: using abstract device models based functional simulators using parameters values
							</li>
							<li>
								Emulated: emulate real storage device using DRAM chips and external DRAM disks
							</li>
							<li>
								Flash Hardware prototype: FPGA based model that provides more accurate development framework for developing hardware models using real NAND flash chips.
							</li>
						</ul>
					</li>
					<li>
						On the software side:
						<ul>
							<li>
								support various hardware device models offered by hardware platform.
							</li>
							<li>
								capable of handling IO requests from host system and deliver them to software through host-side interface.
							</li>
							<li>
								Implemented in Linux OS using ANSI-C and hardware prototyping in Xilinx FPGA uisng Bluespec Systemverilog
							</li>
						</ul>
					</li>
				</ul>
			</p>
		</div>
		<hr>
</p>
</div>

<hr>

<hr>
<!-- 
	This file contains the tail part of the html file to be created
	.. make sure the script is present here
	.. make sure the body section is closed properly
 -->

	<script>
		var coll = document.getElementsByClassName("collapsible");
		var i;

		for (i = 0; i < coll.length; i++) {
		  coll[i].addEventListener("click", function() {
		    this.classList.toggle("active");
		    var content = this.nextElementSibling;
		    if (content.style.display === "block") {
		      content.style.display = "none";
		    } else {
		      content.style.display = "block";
		    }
		  });
		}
	</script>
</body>
</html>