Performance Improvements Strategy & Tracking Issue

[holgerd77]

This issue is to keep track of the ongoing activities around improving the performance of the various EthereumJS libraries - especially the VM - and serve as a general entry point to the topic.

Focus

Improvement Targets

The following library parts are especially suited for performance improvement work (feel free to add):

Library	Improvement Ideas	Tracking Issue
VM	Execution, Cache Improvements (StateManager), Stack & Memory Management, Precompiles	ethereumjs/ethereumjs-monorepo#865
Merkle Patricia Tree	Checkpointing, Storage Backend, Tree Search, Cache Improvements (Tree Search, Storage)	https://github.com/ethereumjs/merkle-patricia-tree/issues/132
Ethash	Overall Execution
Util	Hashing, Signatures
RLP	Encoding / decoding, highly used in e.g. MPT

Improvement Areas

Possible improvements areas:

• Algorithmic Improvements on performance sensitive code parts (iteration strategy, data structures,...)
• More sophisticated caching
  • Cache Structure & Setup (what to store? how long to store? where to store?)
  • Cache Application (when to use? when to flush?)
• Storage Backends
• Build Improvements (build target, dependencies, compiler flags,...)
• Environment & Language (WASM)

Benchmarks / Tooling

Performance Measurement Setups

Helpful setups for performance analysis:

• Dedicated benchmarking tool
• CI Runs
• Behaviour (optimally on reproducible tasks) in a productive environment (e.g. the VM within Buidler)

Benchmark Tooling

Helpful JavaScript/CI Tooling:

• Benchmark.js: benchmarking library
• github-actions-benchmark: GitHub Action for CI

Progress Tracking

It would be nice if we collect some comprehensive results on progress being made over time here.

Note: this can be challenging if the benchmarking suite evolves/changes significantly over time. We'll see.

[s1na]

Great initiative!

On the VM side, I'd like to add investigating the performance penalty of the AsyncEventEmitter, I think it's draining a lot of time. Something else we discussed was how MPT doesn't clear the db nodes after the trie changes.