NVIDIA Posts Linux Scheduler Patches To Further Boost SMT Performance On NVIDIA Vera
Hitting the Linux kernel mailing list today is a new kernel scheduler patch series to enable preferred SMT siblings for NVIDIA Vera with its Olympus cores...
NVIDIA has released patches for the Linux scheduler to enhance symmetric multi-threading (SMT) performance on its NVIDIA Vera processors. SMT allows a single core to run multiple threads simultaneously, utilizing the full core resources when both processing elements (PEs) are active. However, brief activation of an idle sibling can degrade performance, as the sibling remains in low-power mode (WFI) even after the active sibling returns to single-thread mode.
The patches address this issue by making the first processing element (PE0) the preferred sibling for Olympus cores using SD_ASYM_PACKING SMT domain. The scheduler first selects an idle core based on existing placement and capacity rules, then chooses the highest-priority available sibling within that core. This change prevents unnecessary waking of idle siblings and improves overall SMT performance.
The modified scheduler behavior was tested on a two-node Vera system, where an 88-thread single-precision GEMM workload was run on 88 physical cores in NUMA node 0. With the workload free to choose either sibling of every core, observed throughput increased from approximately 9.4 TFLOP/s to approximately 10.1 TFLOP/s with the patches applied. The workload also became more predictable, consistently settling on PE0 while PE1 remained idle.
Written by urgent.news from Phoronix's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
