Open qinsoon opened 4 years ago
https://github.com/mmtk/mmtk-core/pull/131 added a microbenchmark run that reports to the related PR by comparing master and PR. We should also run microbenchmarks and track its history.
We may also want to include this benchmark for our microbenchmark suite: https://github.com/dsrg-uoft/LangBench/tree/main/sort/java
It heavily tests write barrier performance.
We have some criterion benches, but those are using MockVM
right now which is slow. We need to introduce another VMBinding
in mmtk-core to run those benches at reasonable performance.
Two most important things we want to measure are: 1. allocation throughput, and 2. tracing throughput.
We should have simple kernel tests to drive critical part of mmtk-core code. This allows us to analyze performance without involving VM bindings. This also allows us to do thousands of iterations and invocations for stable results (1).
(1): https://stefan-marr.de/2020/07/is-this-noise-or-does-this-mean-something-benchmarking/