Back in August I wrote about NVIDIA working on Linux scheduler improvements for helping the SMT performance on their new Vera CPU with Olympus cores. That work remains ongoing but is looking fairly beneficial for performance especially for knocking it out in less than one hundred lines of code.

The scheduler work is for enabling preferred SMT siblings on NVIDIA Olympus CPU cores. There is a change to honor asymmetric SMT priority in the idle selection code path and a new knob sched_smt_asym_packing= for forcing asymmetric packing at the SMT scheduling domain. This is for systems wanting to enforce specific policy or where the system firmware can't properly communicate the preference.

This work by NVIDIA engineer Andrea Righi is now up to its seventh revision as it works toward the mainline kernel. For those curious about the performance impact of this change, the latest patch series notes some nice 3~6% benefits:

"OpenBLAS throughput increased from 7.11876 +/- 0.06734 TFLOP/s on the baseline kernel to 7.34669 +/- 0.01936 TFLOP/s with this series (+3.20%). NVPL throughput increased from 9.64742 +/- 0.17311 TFLOP/s to 10.29695 +/- 0.01786 TFLOP/s (+6.73%). The lower standard deviation also shows that the results became more predictable. With the series applied, the workloads consistently settled on the lower-numbered sibling."

Hopefully this scheduler work manages to reach the mainline kernel soon (perhaps in time for Linux v7.4?). For now the work can be found out on the Linux kernel mailing list.

While not having these patches applied, recently I did conduct some NVIDIA Vera SMT performance benchmarks for those curious about the on/off impact of Vera's Spatial Multithreading. More NVIDIA Vera CPU benchmarks soon on Phoronix.