hwloc
hwloc copied to clipboard
open-mpi
Hardware locality (hwloc)
v2.0 enforces detailed cache types in synthetic topology strings. You cannot pass "pack:2 cache:3 cache:1 core:1 pu:2" anymore (v1.x would add actual depth and type attributes to vague "cache" types)....
Our release tarballs are currently about 4MB, with about 700kB of tests/hwloc/linux tarballs. Not too bad. But things are expected to get worse with KNL tarballs for different MCDRAM/Clustering configs...
Provide functions like this (see latest numactl at http://halobates.de/blog/p/196): ``` hwloc_obj_t hwloc_get_path_obj(hwloc_topology_t topo, const char *path); hwloc_obj_t hwloc_get_fd_obj(hwloc_topology_t topo, int fd); ``` the latter may return a network device or...
Replace hwloc_get_closest_objs() with an iterator. Starting from object foo, return sibling, then grand-sibling, then grand-grand-siblings, etc. And add the matching --closest 3 to hwloc-calc.
The "cpuid" tool (www.etallen.com/cpuid.html) has a -r option that dumps something very similar to our hwloc-gather-cpuid tool in 2.0. If it's well maintained, we may want to just it and...
Reduce distance matrices so that parent objects get distances between them as well. We already do it druing recursive grouping anyway. Just do it at runtime if the application really...
nodesets are widely duplicated (all objects below a NUMA node have the same one). To reduce the memory footprint, we can add hwloc_bitmap_clone() with a refcount. comparison operations may be...
We get IB information (guid etc) from Linux sysfs but this could actually be generic using IB libs?
CCI interoperability to get cci_device and/or cci_device->name locality Use cci_device->pci.{domain,bus,dev,func} to retrieve the PCI device
If we have a libfabric interoperability, we may need to check if there's anything to do for UCX. It looks like it's going to replace MXM and verbs for Mellanox...