canbench
canbench is a tool for benchmarking canisters on the Internet Computer.
Background
Canister smart contracts on the Internet Computer consume compute and memory resources. Given that resources are finite, there are bounds in place when canisters execute a message (transaction):
- Instructions: a monotonically increasing counter that's corelated with the amount of computation and memory accesses.
- Dirty Pages: the number of memory pages that are written to.
A single message execution must stay within the allowed bounds, otherwise it's terminated.
canbench provides developers the tools and insights to understand how their code is using instructions and memory.
Support for reporting dirty pages will be available once there's a way to retrieve dirty page information from the IC.
Use Cases
- Understanding how a canister consumes instructions, heap memory, and stable memory.
- Detecting performance regressions locally or on CI.
- Analyzing where performance bottlenecks are present.
Features
-
Metrics that are relevant
Typically benchmarking tools run a benchmark multiple times and return the average time. On the Internet Computer, where instrumentation is deterministic, this approach is neither ideal nor insightful. Instead,
canbenchreports the number of instructions consumed, as well as changes to both heap and stable memories. -
Easy detection of regressions
canbenchallows you to persist the benchmarking results in your canister's repository. Storing the benchmarking results allowscanbenchto determine how the performance has changed relative to the past to detect regressions. -
Generous instruction limit
While messages on the Internet Computer are bound to a few billion instructions,
canbenchcan run benchmarks that are up to 10 trillion instructions, giving you the freedom to write resource-intensive benchmarks as needed. -
Language Agnostic
canbenchcan, in theory, benchmark canisters written in any language. Initially support for only Rust exists, but support for additional languages can easily be introduced.
Installation
Quickstart (Rust)
See the crate's documentation.
Github CI Support
canbench can be included in Github CI to automatically detect performance changes.
Have a look at the workflows in this repository for working examples.
You'll need the following:
- Scripts for executing the benchmark, which can be found in the
scriptsdirectory. - A workflow for
canbenchto post a comment with the benchmarking results. Seecanbench-post-comment.yml. - A job for uploading the PR number to the artifacts. This is necessary for step #2 to work correctly. See
upload-pr-numberinci.yml. - The benchmarking job in CI. See
benchmark-fibonacci-exampleinci.yml.
Once you have the CI workflow set up, the job will pass if there are no significant performance changes detected and fail otherwise. A comment is added to the PR to show the results. See this PR for an example.