Ryan Harper wrote: > * Cole Robinson <crobinso@xxxxxxxxxx> [2008-03-14 14:50]: >> The patch below adds an option to virt-install to auto choose >> physical cpu pinning based on available memory on a NUMA machine. >> Conceivably this could be expanded in the future to auto choose >> by some other metric as well, but for now only a NUMA setup is >> implemented. >> >> Using --cpuset=auto will create a traditional cpuset argument >> using all the cpus associated with the cell that best fits the >> requested guest memory. >> >> The only question I have for those interested: Is best fit an >> appropriate default choice? should there be a way to choose >> another scheme (i.e. worst fit)? > Cole, > > Thanks for the patch, looks good. I believe best fit is a good choice. > For our NUMA systems, we want to fully utilize a single node before > moving work to a second node as to minimize the amount of memory traffic > over the interconnects. Best-fit should do this by selecting cell with > the least amount of free ram that satisfies the request. > This is committed now: http://hg.et.redhat.com/virt/applications/virtinst--devel?cs=847863a0c075 The committed patch fixes what would have been a regression when no cpuset was specified, so be sure to use the upstream patch rather than the one I initially posted. Thanks, Cole _______________________________________________ et-mgmt-tools mailing list et-mgmt-tools@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/et-mgmt-tools
