Preliminary Study on Message Aggregation Optimizations for Energy Savings in PGAS(…)
summary
Compares OpenSHMEM conveyor aggregation vs atomic/get/put on Frontier, measuring CPU and memory energy with HPM and visualizing with Scorp and Vampier using Cray openmex 11.7; runs five bale apps (histogram, index gather, sparse matrix transpose, triangle counting, topological sort) on two nodes. Finds conveyor raises instantaneous power but cuts total energy massively — reporting 41× memory improvement, 39× for topological sort, and clear shifts from memory-bound to CPU-bound regimes.
AI-written summary · about
chapters (10)
- 0:00Message aggregation research
- 1:50Energy efficiency in PGAS
- 3:13Frontier supercomputer setup
- 5:41Time improvements with conveyor
- 6:23Energy savings across apps
- 8:32Power consumption analysis
- 10:49Power ratio vs process count
- 12:57Memory vs CPU power trade-off
- 13:41Conclusion and future work
- 14:32Q&A
related talks
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
01h 07mIndustry Presentations
16mDisaggregation and the Application
18mGRANNY: Granular Management of Compute-Intensive Applications in the Cloud
16mA Tale of Two Paths: Toward a Hybrid Data Plane for Efficient Far-Memory Applications
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
18mEthane: An Asymmetric File System for Disaggregated Persistent Memory
01h 00mSystems @Scale Summer Q&A Week 4