Hi Greg, I like the idea to have a macro-benchmarking suite to exactly test the points you've mentioned. If we don't have reliable performance numbers, then it will always be hard to tell whether an improvement makes sense or not (performance-wise).
I think we already undertook a first attempt to do solve the problem with Yoka [1]. The idea was to run a set of algorithms continuously on a machine in the cloud. Yoka was running for some time, but I'm not sure whether this is still the case. Another tool I know of and which people use to run benchmark suites with Flink is Peel [2]. Researcher of Dima are using it to benchmark different distributed engines against each other. But I have never really worked with it. [1] https://github.com/mxm/yoka [2] https://github.com/stratosphere/peel Cheers, Till On Wed, Apr 6, 2016 at 6:56 PM, Greg Hogan <c...@greghogan.com> wrote: > I'd like to discuss the creation of a macro-benchmarking module for Flink. > This could be run during pre-release testing to detect performance > regressions and during development when refactoring or performance tuning > code on the hot path. > > Many users have published benchmarks and the Flink libraries already > contain a modest selection of algorithms. Some benefits of creating a > consolidated collection of macro-benchmarks include: > > - comprehensive code coverage: a diverse set of algorithms can stress every > aspect of Flink (streaming, batch, sorts, joins, spilling, cluster, ...) > > - codify best practices: benchmarks should be relatively stable and > repeatable > > - efficient: an automated system can run many more tests and generate more > accurate results > > Macro-benchmarks would be useful in analyzing improved performance with the > proposed specialized serializes and comparators [FLINK-3599] or making > Flink NUMA-aware [FLINK-3163]. > > I've also been looking recently at some of the hot code and see about a > ~12-14% total improvement when modifying NormalizedKeySorter.compare/swap > to bitshift and bitmask rather than divide and modulo. The trade-off is > that to align on a power-of-2 we have holes in and require additional > MemoryBuffers. And I'm testing on a single data type, IntValue, and there > may be different results for LongValue or StringValue or custom types or > with different algorithms. And replacing multiply with a left shift reduces > performance, demonstrating the need to test changes in isolation. > > There are many more ideas, i.e. NormalizedKeySorter writing keys before the > pointer so that the offset computation is performed outside of the compare > and sort methods. Also, SpanningRecordSerializer could skip to the next > buffer rather than writing length across buffers. These changes might each > be worth a few percent. Other changes might be less than a 1% speedup, but > taken in aggregate will yield a noticeable performance increase. > > I like the idea of profile first, measure second, then create and discuss > the pull request. > > As for the actual macro-benchmarking framework, it would be nice if the > algorithms would also verify correctness alongside performance. The > algorithm interface would be warmup (run only once) and execute, which > would be run multiple times in an interleaved manner. There benchmarking > duration should be tunable. > > The framework would be responsible for configuration of as well as starting > and stopping the cluster, executing algorithms and recording performance, > and comparing and analyzing results. > > Greg >
