Wasm-opt: A general purpose WASM optimizer available in dfx now
I’m excited to announce the integration of wasm-opt, a general purpose wasm optimizer, into dfx.
You can enable optimizations in your dfx.json like so:
{
"canisters": {
"my_canister": {
"optimize": "cycles"
}
}
}
You can check out the specific numbers here: Enable wasm optimizer from `dfx 0.14.0` by kentosugama · Pull Request #55 · dfinity/canister-profiling · GitHub
As a rough estimate, you can expect to see improvements in cycle usage of Motoko canisters by around 10% and Rust canisters by around 7%. Further, your binary sizes should be reduced by roughly 16%.
More Details
The optimizations are provided by wasm-opt, an open source wasm optimizer that is wrapped by ic-wasm to make it compatible with IC canisters.
ic-wasm exposes a couple options for optimization that you can enable from dfx.json. “cycles” is the recommended default, which maps to optimization level 3 in wasm-opt. If you want to aggressively reduce binary size instead, you can do so with the “size” option.
Optimization levels for cycle usage:
- O4
- O3 (equivalent to “cycles”)
- O2
- O1
- O0 (performs no optimizations)
Optimization levels for binary size:
- Oz (equivalent to “size”)
- Os
Our benchmarking indicates that O3 outperforms O4. This is most likely due to the fact that wasm-opt is optimizing for bare-metal execution speed (considering the cache hierarchy for example) while performance on the IC is calculated by the number of instructions executed.
These optimizations preserve the IC specific metadata sections of your canisters. Furthermore, if you want to preserve the name sections in your wasm module, you can directly invoke ic-wasm (instead of enabling optimizations in dfx) with the --keep-name-section flag.
Also, note that in certain cases the optimizations can increase the complexity of certain functions in your wasm module such that they are rejected by the replica. If you run into this issue, we recommend using a less aggressive optimization level such that you do not exceed the complexity limit.
Feedback welcome
We appreciate any feedback on the optimizer as you guys try it out! I hope that all that happens are some free performance gains, but if you encounter any issues please let us know.