After a ton of experimentation with different ways to measure TPS, I settled upon having long-running tasks which all coordinate with the primary task via Atomics, and using ArcSwap for anything that could not be done via an Atomic. This seems to hit the sweet spot of accurate measurements and performance.
This is a major refactor which moves all Pin<Box>'s out of the hot path. I am planning to put together more documentation and benchmarks surrounding this change, since it surprised me at the cost of Pin<Box>. With the minimal benchmarking I've done, the performance difference is fairly massive!
Unfortunately, the cost is that I've broken the distributed runtime functionality. It should be a relatively simple fix to patch up the types, but I want to focus on stabalizing the various controller logic first.
- Added a
direct()scenario which is primarily useful for development. - Added some benchmarking code and a section for thorough benchmarking in the codebase.
- Various minor bug fixes.