Set msize for 9p mounting #25
Conversation
|
Great, nice to see :-) |
|
Nice to have performance back! and even better \o/ 2% memory degradation for the performance improved is something is really good. looks good +1 |
|
built and ran file. |
MarioCarrilloA
force-pushed
the
0.7.0-clearcontainers
branch
2 times, most recently
from
90bed68 to
b2c3a36
Compare
|
+1 |
src/init.c
Outdated
| @@ -465,7 +465,7 @@ static int hyper_setup_shared(struct hyper_pod *pod) | |||
| } | |||
|
|
|||
| if (mount(pod->share_tag, SHARED_DIR, "9p", | |||
| MS_MGC_VAL| MS_NODEV, "trans=virtio") < 0) { | |||
| MS_MGC_VAL| MS_NODEV, "trans=virtio,msize=524288") < 0) { | |||
There was a problem hiding this comment.
Could you add a comment in the code explaining why this magic number?
By adding the msize value (524288 bytes) provides an enhancement in I/O storage operations. The following results are from I/O operations such as: read, write, random read etc...,using different block sizes, where the x results express how many times is better. Read | bs | random | linear | | ----- | ------ | ------- | | 64K | 6x | 7x | | 256K | 13x | 17x | | 512K | 16x | 13x | | 64MB | 14x | 14x | | 256MB | 15x | 15x | | 512MB | 14x | 15x | Write | bs | random | linear | | ----- | ------ | ------- | | 64K | 3x | 3x | | 256K | 3x | 3x | | 512K | 3x | 3x | | 64MB | 3x | 3x | | 256MB | 3x | 3x | | 512MB | 3x | 3x | Signed-off-by: Mario Alfredo Carrillo Arevalo <[email protected]>
MarioCarrilloA
force-pushed
the
0.7.0-clearcontainers
branch
from
b2c3a36 to
9f8a166
Compare
|
@gorozco1 comments added :) |
|
I'd like to see the review of Chao P (or Anthony Xu). I remember he did something similar in CC 2.0 but with |
|
I picked 128K(msize=131072) because I saw less performance increase when I use a bigger value. I suggest to compare the read/write speed in several configurations(512K vs 128K or even 64K...) to see which is better. |
|
Hi @chao-p , I have done that experiment using several msize configuration(64K, 128K, 256K, 512K etc...), with different block sizes and I/O operations (rand read, rand write, linear read and linear write) and I got the best results with 512K :), however I do not know which tool(s) did you use for that. |
|
@MarioCarrilloA Sounds good, then please use 512K. I used pts/aio-stress in phoronix-test-suite. |
|
We need to see the relationship between this and the backing storage block size. For example in the case of device mapper you see that the xfs logbsize=64k. Not sure how the two will interact.
|
By adding the msize value (524288 bytes)
provides an enhancement in I/O storage operations.
The following results are from I/O operations such as:
read, write, random read etc...,using different block
sizes, where the x results express how many times is better.
Read
Write
Signed-off-by: Mario Alfredo Carrillo Arevalo [email protected]