This repository contains the implementation of the parallel intersection-counting algorithm described in the paper:
M. Marzolla, G. Birolo, G. D'Angelo, P. Fariselli, "Parallel Intersection Counting on Shared-Memory Multiprocessors and GPUs", Future Generation Computer Systems volume 159, october 2024, pp. 423-431, ISSN 0167-739X. DOI j.future.2024.05.039
The program has been compiled under Ubuntu Linux 18.04/20.04/22.04 with the following dependencies:
-
The Thrust parallel algorithm library. We used the most recent version from the Git repository
-
NVidia CUDA Toolkit; you should install the version from NVidia web site rather than the version included in Ubuntu's software repository, since the latter might be incompatible with Thrust.
-
The Parallel STL library (part of the GNU C Compiler). GNU STL requires Intel's OneAPI Threading Building Blocks (oneTBB) library, which may or may not be already installed with the GCC compiler. If not, you can install it with
sudo apt install libtbb-dev -
libhts(install withsudo apt install libhts-dev) -
GNU Make
Clone the Thrust library into a local directory:
mkdir -p ~/src
cd ~/src
git clone --recursive https://github.com/NVIDIA/thrust.git
You don't need to compile anything: Thrust is a headers-only library.
The present repository assumes that Thrust is installed in
~/src/thrust, but you can change this by modifying the value of
THRUST_INCLUDE_PATH in the Makefile.
To replicate the tests described in the paper, you need to fetch the
input data into ~/src/intersections-data/ and unpack them using the
following commands:
wget https://si-clusterraspberry.csr.unibo.it/downloads/intersections-data.tar.gz
tar xvfz intersections-data.tar.gz
WARNING the archive is very big (more than 16GB), and becomes even bigger (about 64GB) once decompressed. Please make sure that you have enough disk space.
Clone this repository and build all executables:
git clone https://github.com/mmarzolla/parallel-intersections
cd parallel-intersections
make all
Four programs should be built: intersections_thrust_seq (serial
version), intersections_thrust_omp (OpenMP version for CPU),
intersections_thrust_cuda (CUDA version for the GPU) and
intersections_stl (parallel STL version for the CPU).
Perform a quick check to see if everything works:
make check
Test all implementations with the Chr21 and Exome datasets from
step 2 above:
make test.med
make test.big
There are issues with the NVidia C compiler bundled with Ubuntu 22.04: the serial and OpenMP versions of the parallel intersection-counting program compile and run fine, but the CUDA version does not compile due to the following error:
/usr/include/c++/11/bits/std_function.h:435:145: error: parameter packs not expanded with ‘...’:
435 | function(_Functor&& __f)
| ^
/usr/include/c++/11/bits/std_function.h:435:145: note: ‘_ArgTypes’
/usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not expanded with ‘...’:
530 | operator=(_Functor&& __f)
| ^
/usr/include/c++/11/bits/std_function.h:530:146: note: ‘_ArgTypes’
make: *** [Makefile:112: thrust_count_cuda.o] Error 1
To solve this problem you need to install the NVidia CUDA Toolkit from NVidia web site, rather than the version provided with Ubuntu.
If you encounter this error message:
/home/marzolla/src/thrust/thrust/system/cuda/config.h:116:2: error: #error The version of CUB in your include path is not compatible with this release of Thrust. CUB is now included in the CUDA Toolkit, so you no longer need to use your own checkout of CUB. Define THRUST_IGNORE_CUB_VERSION_CHECK to ignore this.
#error The version of CUB in your include path is not compatible with this release of Thrust. CUB is now included in the CUDA Toolkit, so you no longer need to use your own checkout of CUB. Define THRUST_IGNORE_CUB_VERSION_CHECK to ignore this.
^~~~~
try compiling with the following command at Step 3 above:
NVCFLAGS=-DTHRUST_IGNORE_CUB_VERSION_CHECK make all