Building a Minimal Wasmtime embedding

Wasmtime embeddings may wish to optimize for binary size and runtime footprint to fit on a small system. This documentation is intended to guide some features of Wasmtime and how to best produce a minimal build of Wasmtime.

Building a minimal CLI

Note: the exact numbers in this section were last updated on 2023-10-18 on a macOS aarch64 host. For up-to-date numbers consult the artifacts in the dev release of Wasmtime where the wasmtime-min executable represents the culmination of these steps.

Many Wasmtime embeddings go through the wasmtime crate as opposed to the wasmtime CLI executable, but to start out let's take a look at minimizing the command line executable. By default the wasmtime command line executable is relatively large:

$ cargo build
$ ls -l ./target/debug/wasmtime
-rwxr-xr-x@ 1 root  root    140M Oct 18 08:33 target/debug/wasmtime

The easiest size optimization is to compile with optimizations. This will strip lots of dead code and additionally generate much less debug information by default

$ cargo build --release
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root     33M Oct 18 08:34 target/release/wasmtime

Much better, but still relatively large! The next thing that can be done is to disable the default features of the wasmtime-cli crate. This will remove all optional functionality from the crate and strip it down to the bare bones functionality.

$ cargo build --release --no-default-features
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    6.7M Oct 18 08:37 target/release/wasmtime

Note that this executable is stripped to the bare minimum of functionality which notably means it does not have a compiler for WebAssembly files. This means that wasmtime compile is no longer supported meaning that *.cwasm files must be fed to wasmtime run to execute files. Additionally error messages will be worse in this mode as less contextual information is provided.

The final Wasmtime-specific optimization you can apply is to disable logging statements. Wasmtime and its dependencies make use of the log crate and tracing crate for debugging and diagnosing. For a minimal build this isn't needed though so this can all be disabled through Cargo features to shave off a small amount of code. Note that for custom embeddings you'd need to replicate the disable-logging feature which sets the max_level_off feature for the log and tracing crate.

$ cargo build --release --no-default-features --features disable-logging
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    6.7M Oct 18 08:37 target/release/wasmtime

At this point the next line of tricks to apply to minimize binary size are general tricks-of-the-trade for Rust programs and are no longer specific to Wasmtime. For example the first thing that can be done is to optimize for size rather than speed via rustc's s optimization level. This uses Cargo's environment-variable based configuration via the CARGO_PROFILE_RELEASE_OPT_LEVEL=s environment variable to configure this.

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ cargo build --release --no-default-features --features disable-logging
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    6.8M Oct 18 08:40 target/release/wasmtime

Note that the size has increased here slightly instead of going down. Optimizing for speed-vs-size can affect a number of heuristics in LLVM so it's best to test out locally what's best for your embedding. Further examples below continue to pass this flag since by the end it will produce a smaller binary than the default optimization level of "3" for release mode. You may wish to also try an optimization level of "2" and see which produces a smaller build for you.

After optimizations levels the next compilation setting to configure is Rust's "panic=abort" mode where panics translate to process aborts rather than unwinding. This removes landing pads from code as well as unwind tables from the executable.

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ export CARGO_PROFILE_RELEASE_PANIC=abort
$ cargo build --release --no-default-features --features disable-logging
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    5.0M Oct 18 08:40 target/release/wasmtime

Next, if the compile time hit is acceptable, LTO can be enabled to provide deeper opportunities for compiler optimizations to remove dead code and deduplicate. Do note that this will take a significantly longer amount of time to compile than previously. Here LTO is configured with CARGO_PROFILE_RELEASE_LTO=true.

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ export CARGO_PROFILE_RELEASE_PANIC=abort
$ export CARGO_PROFILE_RELEASE_LTO=true
$ cargo build --release --no-default-features --features disable-logging
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    3.3M Oct 18 08:42 target/release/wasmtime

Similar to LTO above rustc can be further instructed to place all crates into their own single object file instead of multiple by default. This again increases compile times. Here that's done with CARGO_PROFILE_RELEASE_CODEGEN_UNITS=1.

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ export CARGO_PROFILE_RELEASE_PANIC=abort
$ export CARGO_PROFILE_RELEASE_LTO=true
$ export CARGO_PROFILE_RELEASE_CODEGEN_UNITS=1
$ cargo build --release --no-default-features --features disable-logging
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    3.3M Oct 18 08:43 target/release/wasmtime

Note that with LTO using a single codegen unit may only have marginal benefit. If not using LTO, however, a single codegen unit will likely provide benefit over the default 16 codegen units.

One final flag before getting to nightly features is to strip debug information from the standard library. In --release mode Cargo by default doesn't generate debug information for local crates, but the Rust standard library may have debug information still included with it. This is configured via CARGO_PROFILE_RELEASE_STRIP=debuginfo

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ export CARGO_PROFILE_RELEASE_PANIC=abort
$ export CARGO_PROFILE_RELEASE_LTO=true
$ export CARGO_PROFILE_RELEASE_CODEGEN_UNITS=1
$ export CARGO_PROFILE_RELEASE_STRIP=debuginfo
$ cargo build --release --no-default-features --features disable-logging
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    2.4M Oct 18 08:44 target/release/wasmtime

Next, if your use case allows it, the Nightly Rust toolchain provides a number of other options to minimize the size of binaries. Note the usage of +nightly here to the cargo command to use a Nightly toolchain (assuming your local toolchain is installed with rustup). Also note that due to the nature of nightly the exact flags here may not work in the future. Please open an issue with Wasmtime if these commands don't work and we'll update the documentation.

The first nightly feature we can leverage is to remove filename and line number information in panics with -Zlocation-detail=none

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ export CARGO_PROFILE_RELEASE_PANIC=abort
$ export CARGO_PROFILE_RELEASE_LTO=true
$ export CARGO_PROFILE_RELEASE_CODEGEN_UNITS=1
$ export CARGO_PROFILE_RELEASE_STRIP=debuginfo
$ export RUSTFLAGS="-Zlocation-detail=none"
$ cargo +nightly build --release --no-default-features --features disable-logging
$ ls -l ./target/release/wasmtime
-rwxr-xr-x@ 1 root  root    2.4M Oct 18 08:43 target/release/wasmtime

Further along the line of nightly features the next optimization will recompile the standard library without unwinding information, trimming out a bit more from the standard library. This uses the -Zbuild-std flag to Cargo. Note that this additionally requires --target as well which will need to be configured for your particular platform.

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ export CARGO_PROFILE_RELEASE_PANIC=abort
$ export CARGO_PROFILE_RELEASE_LTO=true
$ export CARGO_PROFILE_RELEASE_CODEGEN_UNITS=1
$ export CARGO_PROFILE_RELEASE_STRIP=debuginfo
$ export RUSTFLAGS="-Zlocation-detail=none"
$ cargo +nightly build --release --no-default-features --features disable-logging \
    -Z build-std=std,panic_abort --target aarch64-apple-darwin
$ ls -l ./target/aarch64-apple-darwin/release/wasmtime
-rwxr-xr-x@ 1 root  root    2.3M Oct 18 09:39 target/aarch64-apple-darwin/release/wasmtime

Next the Rust standard library has some optional features in addition to Wasmtime, such as printing of backtraces. This may not be required in minimal environments so the features of the standard library can be disabled with the -Zbuild-std-features= flag which configures the set of enabled features to be empty.

$ export CARGO_PROFILE_RELEASE_OPT_LEVEL=s
$ export CARGO_PROFILE_RELEASE_PANIC=abort
$ export CARGO_PROFILE_RELEASE_LTO=true
$ export CARGO_PROFILE_RELEASE_CODEGEN_UNITS=1
$ export CARGO_PROFILE_RELEASE_STRIP=debuginfo
$ export RUSTFLAGS="-Zlocation-detail=none"
$ cargo +nightly build --release --no-default-features --features disable-logging \
    -Z build-std=std,panic_abort --target aarch64-apple-darwin \
    -Z build-std-features=
$ ls -l ./target/aarch64-apple-darwin/release/wasmtime
-rwxr-xr-x@ 1 root  root    2.1M Oct 18 09:39 target/aarch64-apple-darwin/release/wasmtime

Minimizing further

Above shows an example of taking the default cargo build result of 130M down to a 2.1M binary for the wasmtime executable. Similar steps can be done to reduce the size of the C API binary artifact as well which currently produces a ~2.8M dynamic library. This is currently the smallest size with the source code as-is, but there are more size reductions which haven't been implemented yet.

This is a listing of some example sources of binary size. Some sources of binary size may not apply to custom embeddings since, for example, your custom embedding might already not use WASI and might already not be included.

  • WASI in the Wasmtime CLI - currently the CLI includes all of WASI. This includes two separate implementations of WASI - one for preview2 and one for preview1. This accounts for 1M+ of space which is a significant chunk of the remaining 2.1M. While removing just preview2 or preview1 would be easy enough with a Cargo feature, the resulting executable wouldn't be able to do anything. Something like a plugin feature for the CLI, however, would enable removing WASI while still being a usable executable.

  • Argument parsing in the Wasmtime CLI - as a command line executable wasmtime contains parsing of command line arguments which currently uses the clap crate. This contributes ~200k of binary size to the final executable which would likely not be present in a custom embedding of Wasmtime. While this can't be removed from Wasmtime it's something to consider when evaluating the size of CI artifacts.

  • Cranelift in the C API - one of the features of Wasmtime is the ability to have a runtime without Cranelift that only supports precompiled (AOT) wasm modules. It's not possible to build the C API without Cranelift though because defining host functions requires Cranelift at this time to emit some stubs. This means that the C API is significantly larger than a custom Rust embedding which doesn't suffer from the same restriction. This means that while it's still possible to build an embedding of Wasmtime which doesn't have Cranelift it's not easy to see what it might look like size-wise from looking at the C API artifacts.

  • Formatting strings in Wasmtime - Wasmtime makes extensive use of formatting strings for error messages and other purposes throughout the implementation. Most of this is intended for debugging and understanding more when something goes wrong, but much of this is not necessary for a truly minimal embedding. In theory much of this could be conditionally compiled out of the Wasmtime project to produce a smaller executable. Just how much of the final binary size is accounted for by formatting string is unknown, but it's well known in Rust that std::fmt is not the slimmest of modules.

  • Cranelift vs Winch - the "min" builds on CI try to exclude Cranelift from their binary footprint (e.g. the CLI excludes it) but this comes at a cost of the final executable not supporting compilation of wasm modules. If this is required then no effort has yet been put into minimizing the code size of Cranelift itself. One possible tradeoff that can be made though is to choose between the Winch baseline compiler vs Cranelift. Winch should be much smaller from a compiled footprint point of view while not sacrificing everything in terms of performance. Note though that Winch is still under development.

Above are some future avenues to take in terms of reducing the binary size of Wasmtime and various tradeoffs that can be made. The Wasmtime project is eager to hear embedder use cases/profiles if Wasmtime is not suitable for binary size reasons today. Please feel free to open an issue and let us know and we'd be happy to discuss more how best to handle a particular use case.