Linux Cache Aware Scheduling Extended For Even Better Performance: Up To 360% In MySQL
Cache Aware Scheduling is one of the most exciting kernel innovations to land in Linux this year. While it was finally merged last week to Linux 7.2, a new patch series today is already working to extend Cache Aware Scheduling and is showing some exciting performance improvements.
Cache Aware Scheduling can help improve Linux performance on modern CPUs with multiple cache domains as the scheduler tries to help ensure that tasks sharing data are colocated to the same last level cache (LLC) domain. The net result with Cache Aware Scheduling is ensuring better cache locality and reducing cache misses/bouncing.
Cache Aware Scheduling currently relies on an LLC-centric task aggregation model that works well for workloads fitting within a single LLC domain. But it doesn't work well when the scheduler cannot extend aggregation to higher-level domains and the locality is not effectively preserved.
Hygon sent out patches today to extend Cache Aware Scheduling into topology-aware task aggregation that becomes hierarchical and dynamically expanding/contracting across scheduling domains dynamically based on the workload. The patches are looking quite good for enhancing Cache Aware Scheduling across varying workload sizes and system topologies.
Benchmark numbers on these patches show up to a 49% improvement in Hackbench, Schbench up to a 20% improvement, and in MySQL as much as a 360% improvement:
More details on the benchmark results and these new CAS patches via the kernel mailing list.
Cache Aware Scheduling can help improve Linux performance on modern CPUs with multiple cache domains as the scheduler tries to help ensure that tasks sharing data are colocated to the same last level cache (LLC) domain. The net result with Cache Aware Scheduling is ensuring better cache locality and reducing cache misses/bouncing.
Cache Aware Scheduling currently relies on an LLC-centric task aggregation model that works well for workloads fitting within a single LLC domain. But it doesn't work well when the scheduler cannot extend aggregation to higher-level domains and the locality is not effectively preserved.
Hygon sent out patches today to extend Cache Aware Scheduling into topology-aware task aggregation that becomes hierarchical and dynamically expanding/contracting across scheduling domains dynamically based on the workload. The patches are looking quite good for enhancing Cache Aware Scheduling across varying workload sizes and system topologies.
Benchmark numbers on these patches show up to a 49% improvement in Hackbench, Schbench up to a 20% improvement, and in MySQL as much as a 360% improvement:
More details on the benchmark results and these new CAS patches via the kernel mailing list.