[LU-5473] Test failure sanity test_51b: test_51b failed: fnum: No space left on device Created: 11/Aug/14 Updated: 21/Sep/16 Resolved: 21/Sep/16 |
|
| Status: | Resolved |
| Project: | Lustre |
| Component/s: | None |
| Affects Version/s: | Lustre 2.7.0, Lustre 2.5.3 |
| Fix Version/s: | Lustre 2.9.0 |
| Type: | Bug | Priority: | Major |
| Reporter: | Maloo | Assignee: | Nathaniel Clark |
| Resolution: | Fixed | Votes: | 0 |
| Labels: | HB, zfs | ||
| Issue Links: |
|
||||
| Severity: | 3 | ||||
| Rank (Obsolete): | 15250 | ||||
| Description |
|
This issue was created by maloo for Nathaniel Clark <nathaniel.l.clark@intel.com> This issue relates to the following test suite run: The sub-test test_51b failed with the following error:
Info required for matching: sanity 51b |
| Comments |
| Comment by Oleg Drokin [ 15/Aug/14 ] |
|
I think this is some sort of test env issue? mkdir(/mnt/lustre/d51.sanity/d61732) error: No space left on device total: 61732 creates in 390.32 seconds: 158.16 creates/second /usr/lib64/lustre/tests/sanity.sh: line 4004: /mnt/lustre/d51.sanity/fnum: No space left on device |
| Comment by Jian Yu [ 18/Aug/14 ] |
|
The same failure occurred while verifying patch http://review.whamcloud.com/11435 on Lustre b2_5 branch with FSTYPE=zfs over IB network: The configuration was: MDSCOUNT=1 MDSSIZE=2097152 OSTCOUNT=2 OSTSIZE=8388608 |
| Comment by Bruno Faccini (Inactive) [ 18/Aug/14 ] |
|
+1 at https://testing.hpdd.intel.com/test_sets/890a458c-2675-11e4-a34b-5254006e85c2. |
| Comment by Jian Yu [ 18/Aug/14 ] |
|
By searching on Maloo, I found that all of the failures occurred under the following configuration: MDSCOUNT=1 MDSSIZE=2097152 OSTCOUNT=2 OSTSIZE=8388608 NETTYPE=o2ib FSTYPE=zfs The same test passed on Lustre 2.5.2 with FSTYPE=zfs and NETTYPE=tcp. Here is a for-test-only patch to perform the test against Lustre 2.5.2 with FSTYPE=zfs and NETTYPE=o2ib: http://review.whamcloud.com/11510 |
| Comment by Jian Yu [ 25/Aug/14 ] |
|
Test results showed that the same failure occurred on Lustre 2.5.2 with FSTYPE=zfs and NETTYPE=o2ib. Dmesg on MDS: LustreError: 15047:0:(osd_handler.c:211:osd_trans_start()) lustre-MDT0000: failed to start transaction due to ENOSPC. Metadata overhead is underestimated or grant_ratio is too low. Just increased the MDSSIZE to run the test again. |
| Comment by Jian Yu [ 06/Sep/14 ] |
|
After increasing MDSSIZE from 2GB to 4GB, the tests passed with FSTYPE=zfs and NETTYPE=o2ib: I just changed the size to 3GB to see if the tests pass. And will create a TEI ticket. |
| Comment by Jian Yu [ 07/Sep/14 ] |
|
After increasing the MDSSIZE to 3GB, the tests also passed: I created TEI-2623. |
| Comment by Jian Yu [ 08/Sep/14 ] |
|
One more instance on Lustre b2_5 branch: |
| Comment by Dmitry Eremin (Inactive) [ 16/Sep/14 ] |
|
+1 https://testing.hpdd.intel.com/test_sets/9269b584-3d6a-11e4-af25-5254006e85c2 |
| Comment by nasf (Inactive) [ 22/Sep/14 ] |
|
Another failure instance: |
| Comment by Andreas Dilger [ 03/Oct/14 ] |
|
The test_51b() code itself checks to see if the filesystem is reporting at least $NUMTEST free inodes on the MDT where the test directory is located, and at least 4kB of free space for each file. For creating NUMTEST=70000 empty directories this should consume about 272MB of space on a ZFS MDT, so it is surprising that we can't fit this on a 2GB MDT. I pushed http://review.whamcloud.com/12185 to add some debugging to this test to see why it is failing, and to collect more information about ZFS space usage per inode before and after the test on the MDT even if it is not failing. |
| Comment by Andreas Dilger [ 26/Oct/14 ] |
|
Yu Jian wrote in http://review.whamcloud.com/12185:
|
| Comment by Andreas Dilger [ 27/Oct/14 ] |
|
The failing test run shows about 32KB of space used per inode on the MDT both before and after the test is run. This is more than I would have expected, which is about 4KB per ZFS inode. It is expected that the free space and free inodes would run out at the same time on a ZFS filesystem, since the ZFS inodes are not preallocated as they are on ldiskfs. The total number of inodes and the number of free inodes is an estimate that is based on the space used and number of inodes used (average bytes per inode) and the number of free blocks. It would be interesting to see what the average space used per inode is for other ZFS filesystems. |
| Comment by Peter Jones [ 04/Nov/14 ] |
|
Isaac What do you suggest here? Peter |
| Comment by Andreas Dilger [ 12/Nov/14 ] |
|
Is there any way that we can figure out why the MDS is consuming 32KB per inode? I now recall a patch landing: commit 47c0b97421b21dab686b05d6bf829ebcaf62d5db
Author: Isaac Huang <he.huang@intel.com>
Date: Tue Jul 22 16:42:03 2014 -0600
LU-5391 osd-zfs: ZAP object block sizes too small
Currently osd-zfs ZAP objects use 4K for both leaf
and indirect blocks. This patch increases:
- leaf block to 16K, which equals ZFS fzap_default_block_shift
- indirect block to 16K, the default used by ZPL directories
Signed-off-by: Isaac Huang <he.huang@intel.com>
Change-Id: I5b476414d27822a14afb25e1307991fbd2e3a59e
Reviewed-on: http://review.whamcloud.com/11182
which might account for some of this. If the ZAP leaf and indirect blocks are being updated randomly due to many file creations, it may be that the indirect blocks are taking up a lot of space in the metadnode and saved snapshots? |
| Comment by Gerrit Updater [ 17/Nov/14 ] |
|
Oleg Drokin (oleg.drokin@intel.com) merged in patch http://review.whamcloud.com/12185/ |
| Comment by Jodi Levi (Inactive) [ 20/Nov/14 ] |
|
Patch landed to Master. |
| Comment by Andreas Dilger [ 08/Jan/15 ] |
|
Only a debug patch was landed, and the test was skipped. The core problem is not fixed. |
| Comment by Peter Jones [ 13/Jul/15 ] |
|
Nathaniel Isaac thinks that this failure is due to a flaw in the test script but does not have the bandwidth to dig into to it atm. Are you able to investigate? Thanks Peter |
| Comment by Andreas Dilger [ 28/Aug/15 ] |
|
The test is skipped (SLOW) on review-zfs, but it is still being run during full test runs, and is consistently passing on ZFS. The average space usage before the test run is about 11KB/inode on the MDT, and it reports plenty of free inodes before the test is passing, but the MDT filesystem is 3GB in size so it would have enough space for 95k inodes even at 32KB/inode that we were seeing before. I think 11KB/inode is reasonable for the number of actual files created in the filesystem at this point, since this includes all of the other filesystem overhead. It would be more useful to know the average per-inode space usage once all 70k files were created to get a better average and/or the differential usage for just those inodes. https://testing.hpdd.intel.com/sub_tests/7b8bcf2a-4c80-11e5-b77f-5254006e85c2 == sanity test 51b: exceed 64k subdirectory nlink limit == 05:28:37 (1440480517) UUID 1K-blocks Used Available Use% Mounted on lustre-MDT0000_UUID 3063424 12416 3048960 0% /mnt/lustre[MDT:0] lustre-OST0000_UUID 2031360 5632 2023680 0% /mnt/lustre[OST:0] lustre-OST0001_UUID 2031360 4608 2024704 0% /mnt/lustre[OST:1] lustre-OST0002_UUID 2031104 3968 2025088 0% /mnt/lustre[OST:2] lustre-OST0003_UUID 2031360 4224 2025088 0% /mnt/lustre[OST:3] lustre-OST0004_UUID 2031360 5760 2023552 0% /mnt/lustre[OST:4] lustre-OST0005_UUID 2031104 5888 2023168 0% /mnt/lustre[OST:5] lustre-OST0006_UUID 2031360 6912 2022400 0% /mnt/lustre[OST:6] filesystem summary: 14219008 36992 14167680 0% /mnt/lustre UUID Inodes IUsed IFree IUse% Mounted on lustre-MDT0000_UUID 162442 1363 161079 1% /mnt/lustre[MDT:0] lustre-OST0000_UUID 76956 423 76533 1% /mnt/lustre[OST:0] lustre-OST0001_UUID 73555 323 73232 0% /mnt/lustre[OST:1] lustre-OST0002_UUID 74679 320 74359 0% /mnt/lustre[OST:2] lustre-OST0003_UUID 74436 324 74112 0% /mnt/lustre[OST:3] lustre-OST0004_UUID 71288 323 70965 0% /mnt/lustre[OST:4] lustre-OST0005_UUID 70901 321 70580 0% /mnt/lustre[OST:5] lustre-OST0006_UUID 69146 322 68824 0% /mnt/lustre[OST:6] filesystem summary: 162442 1363 161079 1% /mnt/lustre The osd_statfs->osd_objs_count_estimate() information that is being computed for ZFS is using about 21KB/inode for its free inodes estimate, which is conservative but reasonable given how few inodes are actually in use at this point. It wouldn't be terrible to print another lfs df and lfs df -i after the test, regardless of pass/fail result, to see what the average space usage is on ZFS, and then if it is still reasonable (i.e. going down from 11KB/inode) this bug could be closed. |
| Comment by Nathaniel Clark [ 08/Aug/16 ] |
|
This test hasn't failed since 2015-06-15: Same MDS error: 15:41:02:LustreError: 19303:0:(osd_handler.c:209:osd_trans_start()) lustre-MDT0000: failed to start transaction due to ENOSPC. Metadata overhead is underestimated or grant_ratio is too low. |
| Comment by Gerrit Updater [ 08/Aug/16 ] |
|
Nathaniel Clark (nathaniel.l.clark@intel.com) uploaded a new patch: http://review.whamcloud.com/21821 |
| Comment by Nathaniel Clark [ 23/Aug/16 ] |
|
11KB/inode does seem to be a good estimate for ZFS based custom runs from 21821 |
| Comment by Gerrit Updater [ 21/Sep/16 ] |
|
Oleg Drokin (oleg.drokin@intel.com) merged in patch http://review.whamcloud.com/21821/ |
| Comment by Peter Jones [ 21/Sep/16 ] |
|
Landed for 2.9 |