MLB: A Memory-aware Load Balancing for Mitigating Memory Contention

on YouTube ↗34m

" summary

MLB identifies memory contention with a simple contention model built from retired memory traffic and retry rates, predicts how many memory‑intensive threads run simultaneously, and uses inter-core task migration to concentrate or spread those threads so simultaneous memory requests fall below the predicted threshold. Evaluated on Intel i7‑2600 and Xeon E5‑2690, MLB raises throughput for memory‑heavy tasks, improves CPU↔GPU transfer times, outperforms VBSC in multitasking workloads and works without timeslice modification, explicitly handles the multitasking case.

AI-written summary · about

" chapters (13)

1. 0:00Intro
2. 1:28Trend in the CPU architecture
3. 2:24Performance bottleneck on multicore CPU
4. 5:10Cluster imbalance
5. 6:20Memory-aware Load Balancing(MLB)
6. 10:33Memory Contention Modeling (Cont'd)
7. 13:55Design and implementation of MLB (Cont'd)
8. 18:30Evaluation (Model applicability)
9. 21:33Evaluation (System performance)
10. 23:35Evaluation (Additional benefit)
11. 28:28Worst case
12. 29:06Conclusion
13. 30:54Extension of MLB for multi-threaded applications