Run weak scaling tests

[tkoskela]

The weak scaling test in benchmarks can now be run automatically (tested on Archer2), using ReFrame and Spack. The spack package is upstreamed to spack, the ReFrame test is currently in a branch of excalibur-tests that I'll try to merge soon.

To run the weak scaling test (on Archer2), follow the setup instructions on excalibur-tests, and launch the test with reframe using

reframe -c ../excalibur-tests/benchmarks/apps/conquest/ -r -n SiWeakScaling --system=archer2 -J'--qos=standard'

Current result is below.

Questions raised by the results

• Why are we not seeing much performance degradation from going from one full node to two or more full nodes? Performance actually improves from 256 cores (2 nodes) to 512 cores (4 nodes)
• Why does thread scaling get so much worse going to 4 and 8 threads?
• Why are we not seeing much performance degradation from filling up a node with MPI ranks?

[tkoskela]

Briefly discussed this on Dec 5th, trying to write down what I remember

• [ ] Try scaling up further to larger system size to better see the multi-node behaviour. @davidbowler needs to generate more input files
• [x] Try running with more work per MPI rank. @tkoskela running with 16 atoms per rank

[tkoskela]

Comparison of 16 atoms per rank (left) vs 8 atoms per rank (right).

[tkoskela]

Archer2 budget has been increased by 3000 CU