What do I do with a 1000 cores
summary
Framework developers report running on machines with 480 cores (960 threads with HT) and Windows test racks with ~1792 cores, and find nonuniform memory access dominates performance. Benchmarks show a cloud instance can be 6× slower than another; an older eBay Xeon beats a newer chip for one workload until the JVM flag -XX:+UseNUMA drops 620s to 320s. Advice: avoid JSON for bulk transfer, prefer binary serialization, use atomics/actor or event-sourcing instead of heavy locks, and pin threads with Java Thread Affinity / OpenHFT.
AI-written summary · about
chapters (9)
- 0:00Start
- 4:11Agent Based Simulations
- 7:57Microsoft Test Machine
- 8:24Supercomputers
- 9:06Non-Uniform Memory Access
- 9:32Typical Workstation
- 29:26Requirements
- 35:06Java Thread Affinity
- 47:08Summary
related talks
46mWhat do I do with 1000 cores
35mOpening the Box: Diagnosing Operating-System Task-Scheduler...
34mDesigning, Building and Testing Deterministic HFT Systems
17mP99 Pursuit: 8 Years of Battling P99
10mOn the Cloud We Can’t Wait: Asynchronous Actors Perform Even Better on the Cloud
49mResponding in a Timely Manner
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
39mAsynchronous Processing for Fun and Profity