[LU-12179] allocate continuous pages when disabled page caches

• Status: Open
• Project: Lustre
• Type: Improvement
• Priority: Minor
• Reporter: Qian Yingjin
• Assignee: Qian Yingjin
• Resolution: Unresolved

Description

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

this feature allocates the physical pages per OSS I/O thread in need at the runtime, it may result that  the many allocated pages are  not physical continuous. This may be not good for I/O merge in the block layer due to the limited segements per block I/O request.

For testing purposes, we should pre-allocate continuous pages when OSS thread started as much as possible when disable page chace.

Issue Links

• is related to: LU-11347 Do not use pagecache for SSD I/O when...
• is related to: LU-12071 bypass pagecache for large files
• is related to: LU-13309 performance optimizations for brw

Comments

[Gerrit Updater]

Yingjin Qian ([e-mail]) uploaded a new patch: https://review.whamcloud.com/34644
Subject: LU-12179 osd: allocate continuous pages when disabled page caches
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 8e8d1d32ad55786735823135d6ca126f7fadf40d

[Alex Zhuravlev]

while I understand that SG constraints can prevent large IOs thus limit throughput, I don't see how this is related to pagecache bypassing. i.e. how pagecache can be doing better in this context.

[Qian Yingjin]

IIRC, Ihara did a series of tests using obdfilter-survery,  we found that the performace is bad especially when using 64MB RPC size, and SFA backend can not get large I/Os. After analyzed the debug logs, we found the reason is that the physical pages per I/O thread are not continuous reaching SG constranints.

Current pagechache bypassing allocates pages in need at runtime, when many OSS I/O threads perform IOs concurrently, it may cause that the page allocation is interleaving between each I/O thread, thus many allocated pages per OSS I/O thread is not physical continuous.

If we allocate the pages at the start time of OSS threads, it can  avoid the interleaving page allocation, I think. Thus the pages per I/O thread is much more physical-continnous.

 

Hi Ihara,

Did you do any testing with this patch? any results?

 

Regards,

Qian

 

[Alex Zhuravlev]

> Current pagechache bypassing allocates pages in need at runtime

only once for each thread.

[Qian Yingjin]

From the code it is once for each I/O if the allocated pages are not enough before reach PTLRPC_MAX__BRW_PAGES.

If we allocate pages at the start time of I/O threads and start max_threads during OSS setup, then PTLRPC_MAX_BRW_PAGES pages for each I/O thread are allocated together and it is not interrupt and interleaved by other threads, thus can get much more number of continuous physical pages, I think.

 

If allocating pages during I/O at runtime, then there may be many I/O thread needing to allocate page (via call __page_cache_alloc) for IO concurrently...  and obdfilter-suvery may test with I/O size varying from 1M to 64M, these all will make the allocated pyhiscal pages per I/O thread are not continuous...

[Andreas Dilger]

Now that there are preallocated BRW pages to avoid page cache overhead, since patch https://review.whamcloud.com/32875 "LU-11347 osd: do not use pagecache for I/O", it probably makes sense to update the patch here so those pages can be allocated in larger chunks so that IO submission to the lower storage is more efficient.