Sequence db size != result db size

[jpecar]

Hi,

My users reported hitting this error on MMseqs2-17 on our HPC beegfs parallel storage. I worked with them to isolate a small and quick way to reproduce it consistently.

We run mmseqs easy-linclust input.faa ccl_clus tmp --min-seq-id 1.0 -c 1.0 --cov-mode 0 --threads 1 -v 3 --remove-tmp-files 1 and this works as expected. As soon as we increase --threads to 2 we hit db size inconsistency.

I did some digging in the code and suspect that issue is timing related and comes from OpenMP scheduling in Util::ompCountLines. I see similar case was already worked around in issue #210. Can you comment on this and suggest a workaround or patch?

[milot-mirdita]

That is a very surprising error. Could you upload the input set so I can try to reproduce this locally?

[milot-mirdita]

Does this only happen with tmp on beegfs?

[jpecar]

input.faa.gz

So far yes, we only hit this on beegfs.

We also have another scenario hitting this error from foldseek but I don't know what version of MMseqs2 is embedded into that binary.

[milot-mirdita]

Could you please upload the full terminal output of one of the runs that crashes with this issue?

[jpecar]

output_fail.txt

This is stdout only, but stderr only has the title error.