allow defining ucx device per partition for hpctests

Author: sjpb

ansible/roles/hpctests/README.md

@@ -26,7 +26,7 @@ Role Variables
 - `hpctests_rootdir`: Required. Path to root of test directory tree, which must be on a r/w filesystem shared to all cluster nodes under test. The last directory component will be created.
 - `hpctests_partition`: Optional. Name of partition to use, otherwise default partition is used.
 - `hpctests_nodes`: Optional. A Slurm node expression, e.g. `'compute-[0-15,19]'` defining the nodes to use. If not set all nodes in the selected partition are used.
-- `hpctests_ucx_net_devices`: Optional. Control which network device/interface to use, e.g. `mlx5_1:0`. The default of `all` (as per UCX) may not be appropriate for multi-rail nodes with different bandwidths on each device. See [here](https://openucx.readthedocs.io/en/master/faq.html#what-is-the-default-behavior-in-a-multi-rail-environment) and [here](https://github.com/openucx/ucx/wiki/UCX-environment-parameters#setting-the-devices-to-use).
+- `hpctests_ucx_net_devices`: Optional. Control which network device/interface to use, e.g. `mlx5_1:0`. The default of `all` (as per UCX) may not be appropriate for multi-rail nodes with different bandwidths on each device. See [here](https://openucx.readthedocs.io/en/master/faq.html#what-is-the-default-behavior-in-a-multi-rail-environment) and [here](https://github.com/openucx/ucx/wiki/UCX-environment-parameters#setting-the-devices-to-use). Alternatively a mapping of partition name (as `hpctests_partition`) to device/interface can be used. For partitions not defined in the mapping the default of `all` is used.
 - `hpctests_outdir`: Optional. Directory to use for test output on local host. Defaults to `$HOME/hpctests` (for local user).
 - `hpctests_hpl_NB`: Optional, default 192. The HPL block size "NB" - for Intel CPUs see [here](https://software.intel.com/content/www/us/en/develop/documentation/onemkl-linux-developer-guide/top/intel-oneapi-math-kernel-library-benchmarks/intel-distribution-for-linpack-benchmark/configuring-parameters.html).
 - `hpctests_hpl_mem_frac`: Optional, default 0.8. The HPL problem size "N" will be selected to target using this fraction of each node's memory.

ansible/roles/hpctests/tasks/setup.yml

@@ -28,3 +28,8 @@
     owner: "{{ ansible_user }}"
     group: "{{ ansible_user }}"
   become: true
+
+- name: Set fact for UCX_NET_DEVICES
+  set_fact:
+    hpctests_ucx_net_devices: "{{ hpctests_ucx_net_devices.get(hpctests_partition, 'all') }}"
+  when: hpctests_ucx_net_devices is mapping