Benchmark shuffle

[dapplion]

Motivation

Research performance improvements for the shuffle routine. See #2352

Description

• Adds multiple benchmarks to understand the cost of each component
• Inline closure. It's the only optimization that consistently showed significant improvements. It also matches the implementations in Go and Rust. It reduces the total cpu time (expect spent on hashing) by 50%

Next steps

• Hashing takes 80% of total CPU time so increasing the performance of our hashing library would have a big impact.
• Then we could consider paralleling this function. Given the overhead cost of message passing in Javascript this is not a route we should take until the validator set is really big.

---

Proto: Each round can be computed independently (or in groups, e.g. rounds 0-10, 10-20, etc.), and then you can merge them afterwards. If you use some webworker approach, that may be may improve things.

Proto: The parallelizing thing works like:

• create a list of 0...N for each worker
• assign a round start and end to each worker. E.g. worker 2 does rounds 20...30 of the total 90 rounds.
• every worker runs the list shuffle for their range of rounds
• a final merge step: apply the "mapping" of each worker to the input, in the order of the rounds

Memory-wise: Each worker basically requires as much memory as the original algo: N uint32 values.
Plus some small constant, to cover memory of the hash function and temporary variables.

[codeclimate]

Code Climate has analyzed commit b538e9b and detected 0 issues on this pull request.

View more on Code Climate.

[lgtm-com]

This pull request introduces 1 alert when merging 5d691aa into ec06563 - view on LGTM.com

new alerts:

• 1 for Useless assignment to local variable

[lgtm-com]

This pull request introduces 1 alert when merging b538e9b into ec06563 - view on LGTM.com

new alerts:

• 1 for Useless assignment to local variable