[LU-8641] speedup run_metabech () : make cleanup optional

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.8.0
• Fix Version/s: Lustre 2.10.0
• Type: Improvement
• Priority: Minor
• Reporter: Elena Gryaznova
• Assignee: WC Triage
• Resolution: Fixed
• Labels: patch

Description

1
on real HW clusters we run metabench with parameters :

mbench_NFILES=50000
mbench_THREADS=16
+ /test-tools/mpich2/metabench/src/metabench -w /mnt/fs1/d0.metabench -c 50000 -C -S -k -p /test-tools/mpich2/metabench/dictionary
+ chmod 0777 /mnt/fs1
drwxrwxrwx 5 root root 270336 Apr 25 10:02 /mnt/fs1
+ su mpiuser sh -c "/usr/bin/mpirun      -np 112 /test-tools/mpich2/metabench/src/metabench -w /mnt/fs1/d0.metabench -c 50000 -C -S -k -p /test-tools/mpich2/metabench/dictionary "

test itself comleted in 20 minutes with creation rate :

------------- ------------ ---------- ---------- ----------
Total                         5600000   1112.045   5035.769
Elapsed                       5600000   1112.047   5035.758
------------- ------------ ---------- ---------- ----------
Average                         50000   1097.669     45.556
Std Dev                                   11.592      0.489 (   1.06%) (   1.07%)

[2016-04-0025 10:21:30] Leaving par_create_multidir

then test runs rm -rf <dir> with 5 millions file and can not be completed in 1.5 hours.

2. On real cluster it makes sense to leave files not deleted, so total number of files would grow with time.

Issue Links

• is related to: LU-3308 large readdir chunk size slows unlink...

Comments

[Gerrit Updater]

Elena Gryaznova ([e-mail]) uploaded a new patch: http://review.whamcloud.com/22852
Subject: LU-8641 tests: tests: speedup metabench
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 8e33d6aca1411f524e40493f0e39e44b02054e48

[Andreas Dilger]

Instead of using "rm -r $dir" to empty directories, it is likely much faster to use "find $dir -print0 -type f | xargs -0 unlink; rm -r $dir" to delete the files first, then the remaining directories (if any).

Ideally it would be better to actually fix the "rm -r" performance for large directories, rather than just fixing the test script, since this is a performance problem that also affects users. This is discussed in LU-3308, including options to reduce readdir() RPC chunksize (possibly dynamically when contention is present), and alternately keeping the directory contents cached on the client after the DLM lock has been cancelled (as allowed by POSIX) to avoid cache ping-pong.

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/22852/
Subject: LU-8641 tests: tests: speedup metabench
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: a9339b851789348348e77a02bff16f8d3af69091

[Peter Jones]

Landed for 2.10