hwloc already has code (linux-libnuma.h) for hwloc_cpuset_t to libnuma data type conversion. Does it make sense to add functions to convert hwloc_cpuset_t types to the cpuset string format used by Linux's cgroups?
The cpuset format used by cgroups is not difficult to handle (in fact, the user interface is same as/very similar to the older cpuset implementation). Eg. create a cpuset that includes CPU 2 & 3 for OGS/Grid Engine: # echo 2-3 > /sys/fs/cgroup/cpuset/gridengine/cpuset.cpus Ref: http://www.kernel.org/doc/Documentation/cgroups/cpusets.txt We are pinging the list because we use hwloc for topology binding, and we also added cgroups support (for a lot of things) in OGS/Grid Engine: http://blogs.scalablelogic.com/2012/05/grid-engine-cgroups-integration.html We are planning to integrate hwloc & cgroups (and skip libnuma)... And we want to choose which method we use at runtime. While hwloc can be used by non-root applications, the applications can modify the affinity mask themselves and thus they can escape from the "cpuset" (as least with the current hwloc implementation that uses sched_setaffinity). The cgroups cpuset support is only available to root, but once binded, batch jobs running as non-root could not escape from the cpuset. Rayson ================================================== Open Grid Scheduler - The Official Open Source Grid Engine http://gridscheduler.sourceforge.net/
