[LU-12748] parallel readahead needs to be optimized at high number of process

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.14.0
• Environment: master
• Severity: 3
• Assignee: Wang Shilong (Inactive)
• Reporter: Shuichi Ihara

Description

parallel readahead is enabled by default in master, it contributes to sequential read performance a lot. 
However, if the number of IO thread is increased (e.g. NP=NCPU), read performance drops and it's lower than without readahead. it needs to be tunning and optimization.
Here is test configuration and resutls.

Client
2 x Platinum 8160 CPU @ 2.10GHz, 192GB memory, 2 x IB-EDR(multi-rail)
CentOS7.6(3.10.0-957.27.2.el7.x86_64)
OFED-4.5

for i in 6 12 24 48; do                                                                                                                                       
        size=$((768/i))                                                                                                                                       
        /work/tools/mpi/gcc/openmpi/2.1.1/bin/mpirun --allow-run-as-root -np $i /work/tools/bin/ior -w -r -t 1m -b ${size}g -e -F -vv -o /scratch0/file  | tee
 ior-1n${i}p-${VER}.log                                                                                                                                       
done

Summaruy of Read Performance(MB/sec)

branch	thr=6	thr=12	thr=24	thr=48
b2_12	9,965	14,551	17,177	18,152
master	15,252	16,026	17,842	16,991
master(pRA=off)	10,253	14,489	17,839	18,658

pRA=off  - disabling parallel readahead (llite.*.read_ahead_async_file_threshold_mb=0)

Issue Links

• is related to: LU-13138 sanity test 101d fails with 'readahead 283s > no-readahead 134s 500M'
• is related to: LU-13258 Bind linux workqueues to specific core

Comments

[Li Xi]

That toally makes sense. Thanks for the explanation, Andreas!

[Andreas Dilger]

Li Xi, note that the ability to change labels on the ticket is one of the reasons that we only mark tickets "Resolved" instead of "Closed". Otherwise, it is necessary to re-open and close the ticket to change it again.

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/37927/
Subject: LU-12748 readahead: limit async ra requests
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 1427a72002e6b57017f1c66eb95f9bebff9ac37f

[James A Simmons]

I was just discussing with Wang about this issue with the LU-13258 work. Thanks for figuring out the cross over. I will update my patch this new limit.

[Gerrit Updater]

Wang Shilong ([e-mail]) uploaded a new patch: https://review.whamcloud.com/37927
Subject: LU-12748 readahead: limit async ra requests
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 447b93d563e555f3255995234b35c4546960768e

[Andreas Dilger]

It looks like the crossover is at about NCPU/2 where the performance of parallel readahead and in-process readahead is the same. If we stop using async readahead at that point it should give the best of both worlds.

[Wang Shilong (Inactive)]

The problem could be that we try to submit too much async ra workers, even we limit number of active workers for workqueue did not help.

I think to fix the problem we could introduce similar idea like what we did for limit RA memory.
introduce another atomic counter to record flighting active async ra.

And we limit flighting active async to number of active cpu cores etc which will give us a balance
from single thread improvements and reduce contention of multiple threads.