hwloc (Hardware Locality) 2.5.0rc1 is now available for download.
https://www.open-mpi.org/software/hwloc/v2.5/
v2.5.0 is a fairly big release with 3 major topics:
* NVIDIA NVLink and AMD XGMI bandwidth matrices between GPUs.
The distances API has been extended to make them convenient
to consult but we may add more helpers based user feedback.
* A new "levelzero" backend for Intel oneAPI L0 devices.
GPU distances will be added in the future.
* New topology flags to mitigate spurious binding changes
during hwloc discovery on Windows.
The following is a summary of the changes since v2.5.0.
Version 2.5.0
-------------
* API
+ Add hwloc/windows.h to query Windows processor groups.
+ Add hwloc_get_obj_with_same_locality() to convert between objects
with same locality, for instance NUMA nodes and Packages,
or OS devices within a PCI device.
+ Add hwloc_distances_transform() to modify distances structures.
- hwloc-annotate and lstopo have new distances-transform options.
+ hwloc_distances_add() is replaced with _add_create() followed by
_add_values() and _add_commit(). See hwloc/distances.h for details.
+ Add topology flags to mitigate binding modifications during
hwloc discovery, especially on Windows:
- HWLOC_TOPOLOGY_FLAG_RESTRICT_TO_CPUBINDING and _MEMBINDING
restrict discovery to PUs and NUMA nodes inside the binding.
- HWLOC_TOPOLOGY_FLAG_DONT_CHANGE_BINDING prevents from ever
changing the binding during discovery.
* Backends
+ Add a levelzero backend for oneAPI L0 devices, exposed as OS devices
of subtype "LevelZero" and name such as "ze0".
- Add hwloc/levelzero.h for interoperability between converting
between L0 API devices and hwloc cpusets or OS devices.
+ Expose NEC Vector Engine cards on Linux as OS devices of subtype
"VectorEngine" and name "ve0", etc.
Thanks to Anara Kozhokanova, Tim Cramer and Erich Focht for the help.
+ Add a NVLinkBandwidth distances structure between NVIDIA GPUs
(and POWER processor or NVSwitches) in the NVML backend,
and a XGMIBandwidth distances structure between AMD GPUs
in the RSMI backends.
- See "Topology Attributes: Distances, Memory Attributes and CPU Kinds"
in the documentation for details about these new distances.
+ Add support for NUMA node 0 being offline in Linux, thanks to Jirka Hladky.
* Build
+ Add --with-cuda-version=<version> or look at the CUDA_VERSION
environment variable to find the appropriate CUDA pkg-config files.
Thanks to Stephen Herbein for the suggestion.
- Also add --with-cuda=<dir> to specify the CUDA installation path
manually (and its NVML and OpenCL components).
Thanks to Andrea Bocci for the suggestion.
- See "How do I enable CUDA and select which CUDA version to use?"
in the FAQ for details.
* Tools
+ lstopo now has a --windows-processor-groups option on Windows.
+ hwloc-ps now has a --short-name option to avoid long/truncated
command path.
+ hwloc-ps now has a --single-ancestor option to return a single
(possibly too large) object where a process is bound.
+ hwloc-ps --pid-cmd may now query environment variables,
including MPI-specific variables to find out process ranks.
--
Brice
_______________________________________________
hwloc-announce mailing list
hwloc-announce@lists.open-mpi.org
https://lists.open-mpi.org/mailman/listinfo/hwloc-announce
