2014-05-06 10:05, Burakov, Anatoly: > David Marchand: > > Actually, if we don't care where memory is allocated, we can at least try > > to have the more common setup work properly (i.e. spread memory > > allocations based on used cores). > > I can see no usual setup where you > > want to use cores on a socket while having all memory on another socket > > but still expect performance to be good. > > So here is another approach for Didier's patch. > > We can try to spread memory on numa sockets, if this fails, then we > > default to previous behavior but leave a trace with a warning log "Could > > not spread memory on numa sockets". > > What do you think about this ? > > Sounds like an overcomplication to me. There could be cases where > performance doesn't matter, for example the -m switch could be used to run > various tests (unit tests, functional tests etc.). For anything > performance-related, the recommended option is to use --socket-mem, > especially if you have NICs on specific sockets. Presumably, when you're > setting up a coremask, you already know which sockets your cores are on, so > I don't see a problem with specifying which sockets you want memory from.
Having --socket-mem option to explicitly configure NUMA is OK. Having -m option for simple configuration is OK. Making -m option working for most use cases would be really nice. So I don't understand why we shouldn't do this enhancement. I don't know if "overcomplication" is a good argument. Maybe we should wait the patch to discuss it. -- Thomas
