[LU-1030] new IO engine

• Type: New Feature
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.3.0
• Affects Version/s: Lustre 2.3.0
• Assignee: Jinshan Xiong (Inactive)
• Reporter: Jinshan Xiong (Inactive)

Description

Target of new IO engine:
1. support new grant parameter
2. generate better IO RPC
3. if a page has been set for writeback, and if the covering lock is being canceled, raise prio of writeback page
4. ...

Issue Links

• Trackbacks: Lustre Community Development in Progress Features are being developed for future Lustre releases both at Whamcloud and by other organizations in the Lustre community. These will be eligible for inclusion in future Lustre releases as per our processes

Comments

[Cory Spitz]

Bob, unfortunately, it doesn't look like we have apples to apples here, but the > 15 thread I/O seems a lot worse with b2_3 (around 400 MB/s). Was there an explanation? Granted, we should focus on b2_4 now.

[Bob Glossman (Inactive)]

Testing for b2_3, results file attached. The new curve from 2.3 client/servers looks similar but slightly better than the b2_2+rpc curve in the old results, especially at the mid-range of numbers of thread (4-8). Please note that due to lack of resource I had to slightly reduce the number of OSTs in the tests. With no nodes >16Gb trying to put 2 8G ramdisk OSTs on a node led to lots of OOM. When consulted Jinshan said that numbers from only 15 OSTs were probably good enough for comparison.

[Jinshan Xiong (Inactive)]

all patches have been landed.

[Jinshan Xiong (Inactive)]

Hi Andreas, the patch you're referring to is Chris' testing patch. The original patch is at 2009 and it doesn't change any functionality but move some functions into osc_cache.c; I don't think Chris changed the code instead of adding some test stuff.

[Andreas Dilger]

It appears this may be causing replay-dual.sh test failures that are NOT LU-482 related:

https://maloo.whamcloud.com/test_sets/9a9540c6-a964-11e1-ab65-52540035b04c

Waiting for orphan cleanup...
Waiting 0 secs for  mds-ost sync done.
Waiting 2 secs for  mds-ost sync done.
Waiting for destroy to be done...
before 717896, after 717896

However, it isn't 100% clear that this patch is causing the problem, or only coincidental that Chris was testing for LU-482 repeatedly using this patch. In any case, I don't see any similar problems on master, though ORI-396 at least looks similar.

[Jinshan Xiong (Inactive)]

patch list: http://review.whamcloud.com/

{2009,2460,2270}

and this one is proven to reduce memory usage a lot: http://review.whamcloud.com/2514

[Jinshan Xiong (Inactive)]

write performance benchmark result.

[Jinshan Xiong (Inactive)]

current status of new io engine.