Vmem is real or nonsense on Diskless?

Users/Site Administrators
1

our cluster is modest, but was testing out running comsol for a researcher, heavy on the MPI rather than the OMP and running a short running model. I was watching the mem on node (ours are diskless btw) and basically mucking about.. because comsol is all “it depends” so mpi or omp over a couple of runs, nothing dire but then I caught the vmem number… . oh my. 600+GB??

The real ram is NO MORE than 256. I am pretty sure I did not see htop go above 30 while watching the run.

Searching through the forum, vmem is sorta swap plus “other”

Would love to have a deeper understanding here.

#PBS -l select=1:ncpus=45:mpiprocs=45:ompthreads=1:mem=60GB

resources_used.cpupercent=3970
resources_used.cput=01:25:18
resources_used.mem=29258184kb
resources_used.ncpus=45
resources_used.vmem=608075460kb
resources_used.walltime=00:02:24

2

https://unix.stackexchange.com/questions/375280/is-virtual-memory-always-available#:~:text=So%20if%20there%20is%20no,doesn’t%20require%20any%20swap.

If you use cgroups hook with memsw configuration that might lead to some undestanding and conclusions.

3

Using memsoftlimit, but not memsw (fyi this is also under bright cluster, so enabled/diabled via their interface). It is basically 20x higher than mem.

1 Like