 |
|
You are here: NRAO Public Wiki>HPC Web>HPCSupport>LustreBugABoo (2017-03-02, JessicaOtey)
Edit wiki text
Edit
Attach
Print version
Tracking Lustre Bugs
1 A table of bugs we care about
| Link | Name | Resolved in | Notes |
|---|---|---|---|
| lock callback not getting to client |
n/a |
[CRIT] This is the bug I saw on osiris while running the robinhood fs scan... I've reported my experience and am getting some response... |
|
| MDS not creating files on OSTs properly |
2.7 |
[CRIT] This is the round robin allocator bug... this is essential to get past |
|
| lfs getstripe in nested directories does not indicate pools exist |
not resolved |
[LOW] I think this is more annoying than it is problematic... |
|
| Permanently disabled OST causes clients to hang on df (statfs) |
2.6 |
[LOW] But at least the work-around works in 2.5.3... |
|
| High ldlm_poold load on client |
2.7 |
[LOW] We can see the high load and high CPU consumption by ldlm processes, but we don't see this trace in any of the logs. |
|
| don't print permanently deactivated OSTs in lfs df output |
2.10 |
[LOW] Not critical, but nice... |
|
| incorrect round robin object allocation |
2.8 |
[MED] This is an older problem than LU-5778... but any improvements to allocation are okay by me... |
|
| kernel BUG at fs/jbd2/transaction.c:1033 |
2.5.4, 2.7 |
[MED] We saw this on OSSes, not the MDT... |
|
| This server is not able to keep up with request traffic |
2.5.4 |
[LOW] Seems as if this bug is noise... |
|
| MGS: non-config logname received: params |
2.5.4 |
[LOW] Another noise bug... |
|
| OST pool inheritence is no longer persistent (no LOV EA for child directory) |
2.5.4, 2.7 |
[CRIT] If we want to continue to use pools, this matters. |
|
| lctl dl does not show inactive state for deactivated OST |
it's a "feature!" |
[LOW] Reported by Wolfgang... they are going to change the docs... now UP means configured and will not reflect inactive status. Bah!! |
|
| "lctl {get,set}_param" should also check in /sys/fs/{lnet,lustre} |
2.8 |
[MED] All this sort of rationalization is good news |
|
| OST label tokenizer changes from = to - after first mount |
2.x |
[LOW] But awareness is key. A writeconf will also send the volume back to a state of "first mount". Similar bug (LU-3167) reported about MDT label tokenizer. |
Edit | Attach | Print version | History: r6 < r5 < r4 < r3 | Backlinks | View wiki text | Edit wiki text | More topic actions
Topic revision: r6 - 2017-03-02, JessicaOtey
ALMA
Copyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors. Ideas, requests, problems regarding NRAO Public Wiki? Send feedback