💾 Archived View for thebird.nl › gn-gemtext-threads › issues › gemma › HS-Rat-crashes-gemma.gmi captured on 2023-04-19 at 23:21:50. Gemini links have been rewritten to link to archived content
⬅️ Previous capture (2022-04-28)
-=-=-=-=-=-=-
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
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.
After a gemma lockup we see
[2512382.403215] watchdog: BUG: soft lockup - CPU#118 stuck for 22s! [migration/118:609] [2512404.477219] Out of memory: Kill process 1723 (gemma) score 87 or sacrifice child [2512404.569158] Killed process 1723 (gemma) total-vm:44620288kB, anon-rss:25261688kB, file-rss:0kB, shmem-rss:0kB [2512405.788221] oom_reaper: reaped process 1723 (gemma), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
It is clear parallel GEMMA is running out of RAM.
We can make softlocks messages relax by setting `/proc/sys/kernel/watchdog_thresh` higher. Consider the message harmless.
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. 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:
/usr/local/guix-profiles/gn-latest-20220122/bin/gemma-wrapper --json --loco 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,X -- -debug -g /home/zas1024/gn2-zach/genotype_files/genotype/bimbam/HSNIH-Palmer_true_geno.txt -p /home/zas1024/gn2-zach/tmp/gn2/gn2/PHENO_2+FcfQiTVSC7FmmbsatUPg.txt -a /home/zas1024/gn2-zach/genotype_files/genotype/bimbam/HSNIH-Palmer_true_snps.txt -gk > /home/zas1024/gn2-zach/tmp/gn2/gn2/HSNIH-Palmer_K_TPTFHJ.json /usr/local/guix-profiles/gn-latest-20220122/bin/gemma-wrapper --json --loco --input /home/zas1024/gn2-zach/tmp/gn2/gn2/HSNIH-Palmer_K_TPTFHJ.json -- -debug -g /home/zas1024/gn2-zach/genotype_files/genotype/bimbam/HSNIH-Palmer_true_geno.txt -p /home/zas1024/gn2-zach/tmp/gn2/gn2/PHENO_2+FcfQiTVSC7FmmbsatUPg.txt -a /home/zas1024/gn2-zach/genotype_files/genotype/bimbam/HSNIH-Palmer_true_snps.txt -lmm 9 -maf 0.05 > /home/zas1024/gn2-zach/tmp/gn2/gn2/HSNIH-Palmer_GWA_MWKKYW.json
The geno file is massive:
3.7G Mar 12 11:56 HSNIH-Palmer_true_geno.txt 24K Mar 12 11:56 PHENO_2+FcfQiTVSC7FmmbsatUPg.txt 3.4M Mar 12 11:56 HSNIH-Palmer_true_snps.txt
Probably best to test on a different machine! Let's move to tux02. Running luna (a half year old version of GN2) gives `**** FAILED: number of columns in the kinship file does not match the number of individuals for row = 79`. So, that does not help! I think this is a known issue that got fixed later. Next up, try and run gemma by hand for (largest) chromosome 1 after installing gemma tools with a recent GNU Guix:
tux02:~/tmp/gemma-cpu-lockup$ /usr/bin/time -v gemma -loco 1 -gk -g HSNIH-Palmer_true_geno.txt -p PHENO_2+FcfQiTVSC7FmmbsatUPg.txt -a HSNIH-Palmer_true_snps.txt GEMMA 0.98.5-pre1 (2021-08-14) by Xiang Zhou, Pjotr Prins and team (C) 2012-2021 Reading Files ... ## number of total individuals = 6147 ## number of analyzed individuals = 1306 ## number of covariates = 1 ## number of phenotypes = 1 ## leave one chromosome out (LOCO) = 1 ## number of total SNPs/var = 134918 ## number of SNPS for K = 121094 ## number of SNPS for GWAS = 13824 ## number of analyzed SNPs = 132628 Calculating Relatedness Matrix ...
That is 4 minutes wall clock time using about 26GB of RAM. The machine has 256 GB RAM, so you can see where the problem is. Gemma-wrapper should handle this gracefully, but because of overcommit the machine goes in a tail-spin.
Next run lmm with
/usr/bin/time -v gemma -loco 1 -lmm 9 -g HSNIH-Palmer_true_geno.txt -p PHENO_2+FcfQiTVSC7FmmbsatUPg.txt -a HSNIH-Palmer_true_snps.txt -k output/result.cXX.txt GEMMA 0.98.5-pre1 (2021-08-14) by Xiang Zhou, Pjotr Prins and team (C) 2012-2021 Reading Files ... ## number of total individuals = 6147 ## number of analyzed individuals = 1306 ## number of covariates = 1 ## number of phenotypes = 1 ## leave one chromosome out (LOCO) = 1 ## number of total SNPs/var = 134918 ## number of SNPS for K = 121094 ## number of SNPS for GWAS = 13824 ## number of analyzed SNPs = 132628 Start Eigen-Decomposition... pve estimate =0.161934 se(pve) =0.0355033 ================================================== 100%
Running the LMM takes only 2.5 minutes and uses less than 1GB RAM.
Now the goal is to try and crash the server before setting overcommit.