Juraj Ahel

Ok, got the perf data from the actual compute node, see attached. [perf.zip](https://github.com/soedinglab/MMseqs2/files/10470588/perf.zip) 

Could this have anything to do with NUMA topology? Are your servers NUMA-less? Considering that all our high-mem nodes have multiple NUMA nodes, with at most ~370G per NUMA node,...

I am not finding the numbers now, but I had tested the effect of multiprocessing, both by giving it between 1 and 64 cpus (with fixed 2 threads per cpu),...

Htop reports all (uniref30 db idx file) is in memory: ``` CPU PID USER PRI NI VIRT RES SHR S CPU% MEM% TIME+ Command 4 65761 xxx.xxx 20 0 156G...

@milot-mirdita are you by any chance aware of a specific user that's running local mmseqs server with equivalent performance to what you have? I think you could be on the...

Thanks for the response! I had already done this via docker, which compiles all 3 versions (AVX2, SSE42, SSE2) - AVX2 was definitely the fastest of the three, and all...

@crisdarbellay no, sorry. I don't have access to a "bare metal" server to test this with, but my guess is that Milot guess that it has to do with virtualization...

I see, thank you. I was actually hoping the numactl wrapper would ensure this happens regardless of whether vmtouch is aware of it or not, i.e. that it would disallow...

Yes, it's a regular ~170G file. I was checking beforehand with vmtouch -v {file}, and it was saying 0/x pages are resident - would this have shown the "page cache"...