Update config_machines.xml for Feb 18 software update on Frontier.

[trey-ornl]

Various changes to config_machines.xml in preparation for system software updates on Frontier on 2025-02-18. Within the frontier-scream-gpu machine:

• Change lmod paths from /usr/share to /opt/cray/pe because the /usr/share software is not maintained and is not available on internal OLCF test computers.
• Update craygnuamdgpu to latest version of cpe available on Frontier, cpe/24.11.
• Add explicit versions to other modules in craygnuamdgpu. I used to think this wasn't needed, since module load cpe/24.11 should set all the appropriate defaults. I recently discovered that this setting of defaults only works if the module load cpe/24.11 is a separate command before the other module loads. Cime uses a single module load with a list of modules, which does not update the defaults. I selected the explicit versions based on the cpe/24.11 defaults, including the default for rocm, rocm/6.2.4.
• Remove the libfabric module from craygnuamdgpu. On Feb 18, the default libfabric will change from libfabric/1.20.1 to libfabric/1.22.0, but the latter is not yet available on Frontier. Only the default version on a given computer is officially supported by HPE. I removed the libfabric module from craygnuamdgpu so that it will stay with the default when it changes.
• Change from libfabric/1.15.2.0 to libfabric/1.20.1 for crayclang-scream. On Feb 18, libfabric/1.15.2.0 will disappear from Frontier. And the new default, libfabric/1.22.0, does not support Cray MPI versions before cray-mpich/8.1.28. Because crayclang-scream is stuck back at cpe/22.12 and cray-mpich/8.1.26, we have to use libfabric/1.20.1. We previously switched from libfabric/1.20.1 to libfabric/1.15.2.0 because the former had a serious performance regression. That is supposed to be fixed. And, yes, this does mean that crayclang-scream will be using a version of libfabric that isn't officially supported with the libfabric/1.22.0 drivers, mais c'est la vie.

I ran an 8-node test of NE30 on Frontier and on our internal system (for libfabric/1.22.0) to confirm that the changes work and probably don't seriously impact performance. Here are timing results as reported in e3sm_timing.t.*.

Compiler	cpe	rocm	libfabric	CPM:ATM_RUN
crayclang-scream	22.12	5.4.0	1.15.2.0	15.855: 16.571
			1.20.1	15.885: 16.683
craygnuamdgpu	24.07	6.2.0	1.15.2.0	14.850: 15.594
	24.11	6.2.4	1.15.2.0	14.727: 15.644
			1.20.1	14.729: 15.447
			1.22.0	14.990: 15.601

It would probably be good to confirm with NE1024 runs.

[grnydawn]

@jgfouca , Just to check, do you think crayclang-scream needs any changes due to this system software updates on Frontier?

[jgfouca]

@grnydawn my guess, looking at the module changes, is no.

[ndkeen]

Have you verified that tests are BFB? and similar performance? Including ne256 and ne1024?

[trey-ornl]

Have you verified that tests are BFB? and similar performance? Including ne256 and ne1024?

The chart above shows that the various configurations have similar performance for a short NE30 run using 8 nodes. If you have NE256 and NE1024 runs to suggest, then I'm happy to try them, though they may be too big to test libfabric/1.22.0 on the internal system.

I compared BFB using bfbhash> 234 on the NE30 runs.

• Changing crayclang-scream from libfabric/1.15.2.0 to libfabric/1.20.1 was BFB.
• Changing craygnuamdgpu from cpe/24.07+rocm/6.2.0 to cpe/24.11+rocm/6.2.4, both with libfabric/1.15.2.0, was not BFB.
• Changing craygnuamdgpu with cpe/24.11+rocm/6.2.4 from libfabric/1.15.20.0 to libfabric/1.20.1 was BFB.
• Changing craygnuamdgpu with cpe/24.11+rocm/6.2.4 from libfabric/1.20.1 on Frontier to libfabric/1.22.0 on our internal computer was not BFB.

[ndkeen]

I might suggest at least trying a day or two of ne1024 (with any recent production script) to ensure no issues before merging.