This is an experimental cpu profiler aimed at sampling stack traces and outputting them to pprof format.
Experimental kernel assisted sampling via zx_sampler_create can be enabled via the GN flag ‘experimental_thread_sampler_enabled=true’ which improves sampling times to single digit us per sample.
If not enabled, the sampler will fall back to userspace based sampling which uses the root resource to suspend the target threads periodically, uses zx_process_read_memory and the fuchsia unwinder to read stack traces from the target, then exfiltrates the stack data. In this mode taking a sample using frame pointers takes roughly 300us[^1] per sample. In this fallback mode, it's recommended to set a relatively low sample rate to not overly perturb the profiled program. For reference, 50 samples per second would be a 1.5% sampling overhead.
Note: As experimental_thread_sampler_enabled=true isn't enabled in CI/CQ yet, integration tests need to be run locally with the build flag enabled -- CI/CQ results will reflect the state of the zx_process_read_memory based implementation.
[^1]: Numbers measured on core.x64-qemu
You'll need to add the host tool to interpret the output file to your build. In addition, we recommend several additional arguments to improve the quality of the profiles you get:
fx set <product>.<board> \ --with-host //src/performance/experimental/profiler/samples_to_pprof:install \ --args='debuginfo="backtrace"' \ --args='enable_frame_pointers=true' \ --args='experimental_thread_sampler_enabled=true'
‘debuginfo=“backtrace”’ and ‘enabled_frame_pointers=true’ will give enough information to retrieve and symbolize stacks in a release build.
--release may result in more difficult to follow stacks due to inlining and optimization passes, but will give overall more representative samples. Especially for Rust and C++ based targets, the zero cost abstractions in their standard libraries are not zero cost unless some level of optimization is applied.
The easiest way to interact with the profiler is through the ffx plugin:
ffx profiler start --pid <target pid> --duration 5
This will place a profile.pb file in your current directory which can be handed to pprof.
$ pprof -top profile.out.pb
Main binary filename not available.
Type: location
Showing nodes accounting for 272, 100% of 272 total
flat flat% sum% cum cum%
243 89.34% 89.34% 243 89.34% count(int)
17 6.25% 95.59% 157 57.72% main()
4 1.47% 97.06% 4 1.47% collatz(uint64_t*)
3 1.10% 98.16% 3 1.10% add(uint64_t*)
3 1.10% 99.26% 3 1.10% sub(uint64_t*)
1 0.37% 99.63% 1 0.37% rand()
1 0.37% 100% 1 0.37% <unknown>
0 0% 100% 157 57.72% __libc_start_main(zx_handle_t, int (*)(int, char**, char**))
0 0% 100% 154 56.62% _start(zx_handle_t)
0 0% 100% 160 58.82% start_main(const start_params*)