diff options
| author | Pjotr Prins | 2022-03-12 14:06:39 +0100 |
|---|---|---|
| committer | Pjotr Prins | 2022-03-12 14:06:39 +0100 |
| commit | ad14342a90f3b3487cc4ecc0e42d4c401f031565 (patch) | |
| tree | 5590b4c150cceb3985e1568ef9e9b4e9fc315e07 /issues | |
| parent | bf2e6d9c4b30b8bc58217a93f69889ffcb5d041f (diff) | |
| download | gn-gemtext-ad14342a90f3b3487cc4ecc0e42d4c401f031565.tar.gz | |
issues: gemma: fix OOM issue
Diffstat (limited to 'issues')
| -rw-r--r-- | issues/gemma/HS-Rat-crashes-gemma.gmi | 13 |
1 files changed, 8 insertions, 5 deletions
diff --git a/issues/gemma/HS-Rat-crashes-gemma.gmi b/issues/gemma/HS-Rat-crashes-gemma.gmi index 57bd451..be0af66 100644 --- a/issues/gemma/HS-Rat-crashes-gemma.gmi +++ b/issues/gemma/HS-Rat-crashes-gemma.gmi @@ -1,6 +1,8 @@ # Large datasets crash gemma -Running GEMMA on the NSNIH dataset in Genenetwork sends the server in a tail spin and logs `BUG: soft lockup CPU stuck` messages. This obviously is not great and appears to be a side effect of running openblas aggressively in parallel. +Running GEMMA on the NSNIH dataset in Genenetwork sends the server in a tail spin and logs `BUG: soft lockup CPU stuck` messages. This obviously is not great and appears to be a side effect of running openblas aggressively in parallel (I remember seeing some evidence of that, but I can no longer find that message). Or it may be GEMMA simply runs out of RAM and the kernel is busy cleaning up using the OOM reaper. See + +=> https://lkml.iu.edu/hypermail/linux/kernel/2003.2/01012.html ## Tags @@ -14,6 +16,7 @@ Running GEMMA on the NSNIH dataset in Genenetwork sends the server in a tail spi * [ ] tux02: reboot and reinstate services on tux02 * [ ] tux02: test GEMMA * [ ] tux02: try and optimize versions of openblas using -O2 +* [ ] tux02: deploy GEMMA latest And do the same on tux01 (production) @@ -21,7 +24,7 @@ And do the same on tux01 (production) A 'soft lockup' is defined as a bug that causes the kernel to loop in kernel mode for more than 20 seconds without giving other tasks a chance to run. The watchdog daemon will send an non-maskable interrupt (NMI) to all CPUs in the system who, in turn, print the stack traces of their currently running tasks. -We see +After a gemma lockup we see ``` [2512382.403215] watchdog: BUG: soft lockup - CPU#118 stuck for 22s! [migration/118:609] @@ -31,16 +34,16 @@ We see ``` It is clear parallel GEMMA is running out of RAM. -You can make softlocks messages relax by setting `/proc/sys/kernel/watchdog_thresh` higher. +We can make softlocks messages relax by setting `/proc/sys/kernel/watchdog_thresh` higher. Consider the message harmless. -Also overcommit is set to 0 on Tux01. We may want to change that to +Overcommit is set to 0 on Tux01. We may want to change that to ``` vm.overcommit_memory=2 vm.overcommit_ratio=90 ``` -That will make out-of-RAM problems less impactful. I have not set that before because it requires rebooting the production server. +That will make out-of-RAM problems less impactful. We have been running penguin2 like this for over a year with no more OOM problems. I have not set that before on tux01 because it requires rebooting the production server. From Zach I got the K and GWA commands: |