dfinity / canister-profiling

Collection of canister performance benchmarks
Apache License 2.0
21 stars 8 forks source link

canister-profiling

This repository contains code and scripts for collecting performance data for different canisters running on the IC.

Community contributions are strongly encouraged.

Performance report

Performance reports are generated in gh-pages branch. The reported Wasm binary size is after the instrumentation.

How to reproduce performance report

Prerequisites:

Running locally:

How to create a new benchmark

Each benchmark usually contains multiple implementations written in different languages, e.g., Motoko and Rust. The folder follows the following structure:

Benchmark_name/
  Makefile
  README.md // Perf result will be appended to this markdown file.
  perf.sh   // ic-repl script that generates perf result. If the candid interface is different, we can use multiple scripts.
  motoko/
    dfx.json
    src/
      benchmark1.mo
      benchmark2.mo
  rust/
    dfx.json
    benchmark1/
      Cargo.toml
      benchmark1.did
      src/
        lib.rs
    benchmark2/
      Cargo.toml
      benchmark2.did
      src/
        lib.rs

Wasm Optimizer

A Wasm optimizer is applied to each Wasm binary before instrumentation. The optimizer can be found in ic-wasm, which wraps wasm-opt.

The following optimizations are applied:

ic-wasm -o <wasm> <wasm> shrink --optimize O3 --keep-name-section

Note that the name section is preserved in the optimization process. This is because the name section is used by the profiler to produce the flame graphs.

For users who wish to use the optimizer, the easiest way is to enable it via a field in dfx.json:

{
  "canisters": {
    "my_canister": {
      "optimize": "cycles"
    }
  }
}

This, as in most real world uses, removes the name section to minimize the binary size.