Linux Kernel Scheduling and split-LLC architectures: Overview, Ch...
summary
Measurements on split-LLC AMD EPYC systems show concrete misplacements: SCHbench tail latency becomes up to 2.5× worse when messenger and workers land on different LLCs; stream with 16 LLCs sees severe throughput loss when two bandwidth threads share one LLC because initial placement ties break poorly; t-bench gains ~20% (64 clients) and ~12% (128) when initial placement avoids a 10,000-migration uphill rebalancing. Experiments use futex/WF_SYNC behavior, the new idle load balancer, and a prototype using prctl hints (children inherit) that often helps but can regress other workloads; group-utilization heuristics also cause tradeoffs.
AI-written summary · about
chapters (9)
- 0:00Introduction and context
- 1:25Split vs unified LLC differences
- 2:52Consolidation challenge with SCH bench
- 5:25Spread challenge with STREAM workload
- 8:34Thundering herd problem with T-bench
- 11:19NUMA and LLC domain limitations
- 13:06Exploring user-space hints for scheduling
- 17:22Hints impact on different workloads
- 19:29Future directions and Q&A
related talks
45mUsing hardware hints for optimal page placement
38mMaximizing Heterogeneous Memory Bandwidth in Multi-Socket Systems
19mDo nothing fast: How to scale idle cpus?
22mLXDs: Towards Isolation of Kernel Subsystems
35mOpening the Box: Diagnosing Operating-System Task-Scheduler...
32mImprove CPU Utilization by Designing a New Schedule Algorithm
21mPreview: Operating Systems
41mLinux Systems Performance