> On Nov. 19, 2012, 12:21 a.m., Ben Mahler wrote: > > src/slave/cgroups_isolation_module.cpp, line 593 > > <https://reviews.apache.org/r/8108/diff/1/?file=191733#file191733line593> > > > > Looks like setting cpuset.mems is also mandatory here.. > > > > https://access.redhat.com/knowledge/docs/en-US/Red_Hat_Enterprise_Linux/6/html/Resource_Management_Guide/sec-cpuset.html > > Jie Yu wrote: > And seems that cpuset.mems is different from cpuset.cpus. > > For example, it's possible that > cpuset.mems = 0-1 > cpuset.cpus = 0-15 > > Probably it means that there are two chips, each chip has a memory node > associated with, and each chip has 8 hardware thread (probably 4-cores with > hyper threading). > > Jie Yu wrote: > No sure whether you need to configure cpuset.mems at all if you don't > care about memory affinity. Probably you can just set cpuset.mems to be the > value of all the available memory nodes.
Agreed, for this first pass, I think we shouldn't consider memory affinity. We do need to be careful in that we must use a (possibly equal) subset of the hierarchy's cpuset.cpus and cpuset.mems. - Ben ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/8108/#review13550 ----------------------------------------------------------- On Nov. 17, 2012, 11:13 p.m., Ben Mahler wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/8108/ > ----------------------------------------------------------- > > (Updated Nov. 17, 2012, 11:13 p.m.) > > > Review request for mesos, Benjamin Hindman, Vinod Kone, and Jie Yu. > > > Description > ------- > > This is the first pass at adding cpuset isolation for pinning cgroups to cpus. > > We decided to start with a simplistic grow/shrink allocation technique, as > such this initial technique: > -Does not take cache locality into account. > -Does not actively fight fragmentation*, but does a good job at preventing > it in many cases, given it's simplicity. > -Note that when cpus resource requests are integral (non-fractional), then > fragmentation does not occur. > > *By fragmentation, I'm referring to the case where we've spread a cgroup over > more cpus than necessary, due to other cgroups sharing the same cpus. > High fragmentation would mean a lot of shared cpus across cgroups. > No fragmentation would mean each cgroup has a unique set of cpus. > > I've punted on documenting the pitfalls of this technique, wiring up the > handler, and adding tests for now. > > Note that this is diffed off of benh's changes: > https://reviews.apache.org/r/8058/ > https://reviews.apache.org/r/8059/ > > > Diffs > ----- > > src/linux/proc.hpp 27e15bf8695aa694b0d5bdb6881b9fa55a447528 > src/slave/cgroups_isolation_module.hpp > 9f80fc5a969b959b34eaea4cac40700662d7f8b2 > src/slave/cgroups_isolation_module.cpp > 8211618d7729350654e2d17946c5b912ed9dda6a > third_party/libprocess/include/stout/strings.hpp > 914c280a994733764957d19f37b48d151bb93778 > > Diff: https://reviews.apache.org/r/8108/diff/ > > > Testing > ------- > > None as of yet. > > > Thanks, > > Ben Mahler > >
