And Bogdan Costescu writes: > Your description fits *light* swap usage. However, in quite some years > of combined experience on both sides (user and sysadmin) I have never > seen such swap usage on a compute node: [...]
You know all those system daemons that aren't doing anything while a compute job is running? Those get pushed out to swap if the thing doing real work needs the memory. They are swapped back into memory between jobs (ideally not during). It'd be nice simply not to have those daemons or to force them into out-of-memory hibernation, but that's much more effort. (Also, it'd be nice for boot times on 1+TiB nodes to be less than 15 minutes before even touching the OS. ugh. If only there were a fast scan of a small portion where the OS will be loaded, then the OS could scan the rest...) > Overcommitting is unfortunately still needed for the older Fortran > applications which statically allocate large arrays. Also for modern runtime systems that allocate a large heap they manage themselves. Or C code that does similar things to avoid copying pages while remaining portable (e.g. UMFPACK). _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
