What do I do with 1000 cores
summary
A JVM-based agent‑based modeling SDK at Simudyne targets wall‑clock speed on 1,000+ cores and shows concrete tradeoffs: JDK 1.8 is common, NUMA matters so use numactl or CPU binding, AMD EPYC/ThreadRipper topology flips microbenchmark rankings, and cloud VMs show 1.3x–6.5x runtime variance. The implementation favors event sourcing and lock‑free patterns, uses JMH, vtune and profilers (with caution), and runs on bare‑metal testbeds; the result shortens runs enough to save hours and money.
AI-written summary · about
chapters (10)
- 0:00Introduction and context
- 3:02Hardware trends and scaling
- 8:06Supercomputers and cloud
- 12:41Cloud latency and performance
- 16:39Agent-based modeling
- 23:01Memory architecture and NUMA
- 27:02Processor comparisons
- 31:16Distributed systems challenges
- 36:56Optimization strategies
- 39:50Conclusions and best practices
related talks
47mWhat do I do with a 1000 cores
35mOpening the Box: Diagnosing Operating-System Task-Scheduler...
34mSimulating Our Apache Mesos Framework for Fun and Profit and...
18mGRANNY: Granular Management of Compute-Intensive Applications in the Cloud
11mRemote execution with 100 to 100,000 CPUs
23mBreaking the limits of x86
34mDesigning, Building and Testing Deterministic HFT Systems
35m100 machines do 100s work in 1s