Heterogeneity and characteristic testing
summary
Cloud infrastructure moves toward wider heterogeneity with ARM and many accelerators like TPUs DPUs GPUs appearing, creating up to 100× power difference between generic and accelerated implementations. Characteristic testing must cover many CPU/accelerator/software combinations — potentially hundreds of combinations — but those tests are slow expensive and resource demanding. Benchmarks often do not match application footprint so sizing for SLAs fails. The claim: existing benchmarking methods do not scale and new tools and methods are required.
AI-written summary · about
related talks
11mBuilding a Test for Diverse Fleet of Platforms, How Hard That Could Be?
44mBenchmark testing
01h 08mPerformance testing in the cloud
49mPerformance Testing
06mIGNITE: How to benchmark … poorly
35mBenchmarking: You're Doing It Wrong
54mPerformance tests and other lies
02h 44mMeasure, don’t guess: benchmarking stories from the trenches