On the Cloud We Can’t Wait: Asynchronous Actors Perform Even Better on the Cloud
summary
Comparison of OpenSHMEM (synchronous PGAS) and HClib-actor (asynchronous) on Bale graph microkernels shows asynchronous code runs far faster on cloud VMs with Ethernet than synchronous code on Azure, with up to 1,000x speedups versus 10–30x on the InfiniBand HPC cluster (PACE). Tests use single‑core VMs, avoid Azure accelerated networking, and reveal highly variable execution times (e.g., 68s spread on histogram); response includes a runtime profiler of logical topology and routing/atomics tuning that improves actor performance.
AI-written summary · about
chapters (10)
- 0:00Cloud performance challenges
- 1:04Synchronous vs asynchronous PGAS
- 2:27Experimental setup PACE Azure
- 4:17Bale graph micro-kernels
- 5:02Speedup of async actors
- 6:27Network noise in cloud
- 6:50Node locality limits Azure
- 7:54Research on general VMs
- 8:15Profiling network noise
- 8:36Summary
related talks
18mDistributed Actors in OCaml
01h 07mThe Asynchronous C++ Parallel Programming Model
32mMaximizing Throughput and Scalability for Akka Streams
23mImproving Ray for Large-scale Applications
20mAdvancing HPC The Hybrid Scheduler Approach at the D.E.Shaw Group
16mPreview: Cloud Computing
01h 14mAkka.NET: The Future of Distributed Programming in .NET
47mWhat do I do with a 1000 cores