hwloc
hwloc copied to clipboard
open-mpi
Hardware locality (hwloc)
See if there's any point in parallelizing the Linux or x86 discovery. In x86, we use CPUID on every PU, could be parallelized, but likely too short to be useful....
hwloc-bind uses exec() for launching the actual program. It's not clear whether thread/process binding is inherited in all cases (thread binding, process binding, with or without multiple processor groups, with...
We are running into an infinite loop when using hwloc in embedded mode in mpich with compiler Sun C 5.10 SunOS_i386 2009/06/03. I see it’s because of the loop in...
Treematch guys want to export the arities of a topology in a C array instead of a synthetic description string. Not clear whether they want the types too, but that's...
@jjhursey would like to get the major/minor numbers corresponding to a NVIDIA GPU (for cgroup config). It looks like /dev/nvidiaX corresponds to NVML device #X. However CUDA devices can be...
lstopo manually allows PU ignoring by skipping those objects during the output, but their children are also ignored. Those children are only Misc objects, not very important. Same issue applies...
Related to #252, we could extend the filtering API to filter I/O devices more precisely. If osdev/coproc are filtering-out but not osdev/openfabrics, opencl/cuda/... components don't need to run at all.
Factorize rules in something like this https://github.com/open-mpi/ompi/blob/master/Makefile.ompi-rules
The darwin backend reads everything from sysctlbyname(). It could read sysctl values from a file instead of from the actual OS. This way, we could regression test different Apple machines...
Berkeley likes hwloc but still thinks it is too hard for normal people such as physicists. So they would like a GUI combining lstopo and top to bind processes or...
