[LU-11347] Do not use pagecache for SSD I/O when read/write cache are disabled

• Status: Resolved
• Project: Lustre
• Fix Version/s: Lustre 2.12.0
• Type: Improvement
• Priority: Major
• Reporter: Joseph Gmitter (Inactive)
• Assignee: Alex Zhuravlev
• Resolution: Fixed
• Labels: perf_optimization

Description

For SSD optimization, disable pagecache for IO when:

• non-rotational storage is used
• when both read and write caches are disabled

Issue Links

• is related to: LU-5375 Failure on test suite sanity test_151...
• is related to: LU-13680 large allocations in osd_bufs_get() f...
• is related to: LU-12179 allocate continuous pages when disabl...
• is related to: LU-9252 osd: Cannot really disable cache on OSS
• is related to: LU-11164 lvbo_*() methods to reuse env
• is related to: LU-12071 bypass pagecache for large files

Comments

[Joseph Gmitter (Inactive)]

test only patch is at https://review.whamcloud.com/#/c/32875/

 

[Andreas Dilger]

Attached flame graph ( ost_io.svg ) shows find_or_create_page() using 4.25% of CPU under heavy random read workload (using fake IO to avoid storage overhead)..

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/32875/
Subject: LU-11347 osd: do not use pagecache for I/O
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 72372486a5e96dd2078e01a697025da41bbf03c2

[Peter Jones]

Landed for 2.12

[Lukasz Flis]

Alex, Andreas:
According to LU-9252 this patch is needed to disable osd cache.

Would it be possible to have it ported to b2_10?

[Peter Jones]

It would be possible to do this but we have no plans to do this at this time

[Lukasz Flis]

I think it would be beneficial in some cases.
In 2.10.7 and mixed workload cluster - profiling of ost_io threads shows that we spend 35% of the thread time in find_or_create_page and kswapd0 being hyperactive (swap is off)

+   34.99%     0.35%             7  ll_ost_io01_034  [kernel.vmlinux]  [k] find_or_create_page                                                                     

I have already backported the patch - if it's possible to submit it for autotesting it would be great