Details
-
Bug
-
Resolution: Unresolved
-
Major
-
None
-
None
-
Alma 8 Lustre 2.15.4 client
Centos 7 Lustre 2.12.9 server
-
3
-
9223372036854775807
Description
Environment:
- No DNE
- No DOM
- Opencache disable
Reproducer:
On the first client:
[root@client1 test_bug_2.12-2.15]# mkdir foo [root@client1 test_bug_2.12-2.15]# for i in 1 2 3 4 ; do dd if=/dev/zero bs=1k count=1k of=foo/$i ; done [root@client1 test_bug_2.12-2.15]# cd foo [root@client1 foo]# tail -f 2 ...
On the second client:
[root@client2 test_bug_2.12-2.15]# mv foo/ foo.new ; mkdir foo ; mv foo.new foo/foo [root@client2 test_bug_2.12-2.15]# find . ./foo ./foo/foo ./foo/foo/4 ./foo/foo/1 ./foo/foo/2 ./foo/foo/3
Back on the first one:
^C [root@client1 foo]# ls 1 2 3 4 [root@client1 foo]# cd .. [root@client1 test_bug_2.12-2.15]# ls foo/ 1 2 3 4
Here, we should have:
[root@client1 test_bug_2.12-2.15]# ls foo/ foo
Then, if the directory is touched then the directory entry is sync:
[root@client1 test_bug_2.12-2.15]# touch foo/2 [root@client1 test_bug_2.12-2.15]# ls foo/ foo [root@client1 test_bug_2.12-2.15]# ls foo/foo/ 1 2 3 4
Attachments
Issue Links
- is blocked by
-
LU-11623 Allow caching of open-created dentries
-
- Reopened
-
Activity
"Etienne AUJAMES <eaujames@ddn.com>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/55295
Subject: LU-17583 llite: getattr/open should not revalidate dentry
Project: fs/lustre-release
Branch: b2_15
Current Patch Set: 1
Commit: db81162b82315dc15926a02f4cb9f165c75709e3
"Etienne AUJAMES <eaujames@ddn.com>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/54854
Subject: LU-17583 mdt: don't fetch LOOKUP lock for getattr/open by fid
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: ac6daa3a6fd8ddc87af2111c6258812ca85ab91b
I benched the 2 patches and the master branch:
Evironement
The test is similar than the "LU-10948 llite: Revalidate dentries in ll_intent_file_open" (https://review.whamcloud.com/32157):
- 50 clients (VMs), 20 cores per node
- TCP network
- mpirun --hostfile hfile -np 1000 ./src/mdtest -d /mnt/cache/ -v -F -n 200 -w 1k -C -T -r -i3
Tests results
baseline (master branch: 2.15.61_225_gbb6a2d2) ------- SUMMARY rate (in ops/sec): (of 3 iterations) Operation Max Min Mean Std Dev --------- --- --- ---- ------- File creation 973.425 970.273 971.558 1.655 File stat 1289674.130 1275339.298 1280954.250 7655.213 File read 0.000 0.000 0.000 0.000 File removal 1597.736 1586.937 1593.411 5.711 Tree creation 360.366 25.710 243.243 188.575 Tree removal 54.075 45.861 51.261 4.678 no revalidate patch ------------------- SUMMARY rate (in ops/sec): (of 3 iterations) Operation Max Min Mean Std Dev --------- --- --- ---- ------- File creation 964.990 964.175 964.627 0.415 File stat 1290872.830 1277123.666 1282763.829 7199.393 File read 0.000 0.000 0.000 0.000 File removal 1586.650 1579.566 1583.856 3.772 Tree creation 357.906 127.050 276.751 129.798 Tree removal 54.996 52.189 54.031 1.596 revalidate by name patch ----------------------- SUMMARY rate (in ops/sec): (of 3 iterations) Operation Max Min Mean Std Dev --------- --- --- ---- ------- File creation 965.137 963.721 964.344 0.723 File stat 1288246.143 1275527.402 1280273.795 6945.983 File read 0.000 0.000 0.000 0.000 File removal 1586.897 1579.279 1582.628 3.892 Tree creation 321.945 29.644 206.441 155.491 Tree removal 56.468 41.925 48.071 7.529
So, no major performance regression detected with or without the open/getattr entry revalidation.
If no objection, I will abandon the "lite: getattr/open should revalidate dentry by name" (https://review.whamcloud.com/54607). The patch does not improve the performance significantly.
"Etienne AUJAMES <eaujames@ddn.com>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/54607
Subject: LU-17583 llite: getattr/open should revalidate dentry by name
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 9e69ad31f0116a6eb4089a631a164ee45cd8c0e7
"Etienne AUJAMES <eaujames@ddn.com>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/54354
Subject: LU-17583 llite: ll_getattr should not revalidate dentry
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 37a19ec08985248bb2a8727d7e122b068d7e2ca6
"Etienne AUJAMES <eaujames@ddn.com>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/54352
Subject: LU-17583 tests: bug reproducer
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: e6891440f7989fc4b2f666dd236647c865577f60
I have reproduced the issue with master branch one client (RHEL 8.9) and a 2.15.4 server.
Reverting the 92fadf9 "LU-15200 llite: revalidate dentry if LOOKUP lock fetched" (https://review.whamcloud.com/45599) seems to fix the issue.
int ll_revalidate_it_finish()
...
if (bits & MDS_INODELOCK_LOOKUP) {
if (!ll_d_setup(de, true))
RETURN(-ENOMEM);
d_lustre_revalidate(de); <-------------
ll_getattr checks the entry by fid, it does not resolve it by name, so it cannot revalidate the direntry.
patch https://review.whamcloud.com/32157 needs to be confirmed if stat performance can be boost on the single client.
This is not determined whether if 1.2M stat performance because of 50 clients or not.
Here is what I did test results with/without https://review.whamcloud.com/#/c/fs/lustre-release/+/54354/ on the single client.
master (commit: 424b9ccb00)
master (commit: 424b9ccb00) + patch https://review.whamcloud.com/#/c/fs/lustre-release/+/54354/
I didn't see obvious regression with patch in this configuration.