This library provides bindings to google's cpuprofiler.
There are other profiling tools for Rust, cargo-profiler is particularly good! This library certainly doesn't replace those but adds a some different tools to the mix:
- Makes it easy to profile only sections of code
- Uses statistical sampling (like oprofiler) which means low overhead
- Works with pprof for a range of output formats
In order to use this library you will need to install gperftools. There are instructions in their repository but it's roughly the following:
- Download package from releases
- Run
./configure
- Run
make install
- Run
sudo ldconfig
There may be some other dependencies for your system - these are explained well in their INSTALL document. For example libunwind (> 0.99.0) is required for 64 bit systems.
Add cpuprofiler
to your Cargo.toml
manifest.
[dependencies]
cpuprofiler = "0.0.3"
Add the dependency to your root:
extern crate cpuprofiler;
Start and stop the profiler around the code you'd like to draw samples. This will save the profile to a file you specify.
use cpuprofiler::PROFILER;
PROFILER.lock().unwrap().start("./my-prof.profile").unwrap();
// Code you want to sample goes here!
PROFILER.lock().unwrap().stop().unwrap();
Now you can just run the code as you would normally. Once complete the profile will be saved to ./my-prof.profile
.
The final step is the fun part - analyzing the profile!
To analyze the profile we use google's pprof tool.
An old version of this tool is included with the gperftools package. This is the version I have been using but the newer Go version should work too! The usage of pprof is well documented in the cpuprofiler docs.
The output format is entirely dependent on pprof but here are some examples from a Rust program:
Total: 855 samples
207 24.2% 24.2% 207 24.2% matrixmultiply::gemm::masked_kernel::hfdb4f50027c4d91c
156 18.2% 42.5% 853 99.8% _$LT$rusty_machine..learning..optim..grad_desc..StochasticGD$u20$as$u20$rusty_machine..learning..optim..OptimAlgorithm$LT$M$GT$$GT$::optimize::h2cefcdfbe42a4db8
79 9.2% 51.7% 79 9.2% _$LT$$RF$$u27$a$u20$rulinalg..vector..Vector$LT$T$GT$$u20$as$u20$core..ops..Mul$LT$T$GT$$GT$::mul::h21ce4ecb4bbcb555
66 7.7% 59.4% 73 8.5% __ieee754_exp_sse2
61 7.1% 66.5% 95 11.1% _$LT$rusty_machine..learning..toolkit..regularization..Regularization$LT$T$GT$$GT$::l2_reg_grad::h4dff2e22567a587e
57 6.7% 73.2% 274 32.0% matrixmultiply::gemm::dgemm::h2d985771431fcfd4
41 4.8% 78.0% 42 4.9% _$LT$rulinalg..matrix..Matrix$LT$T$GT$$GT$::transpose::h736b18b122958bcd
31 3.6% 81.6% 32 3.7% sdallocx
The first column is the number of samples from each function. The second is the percentage of samples which were found directly in this function, and the third column is the percentage of samples which were in this function or it's children. I think...
Below is a snippet of an interactive graph output.
The above graph is produced by pprof and shows which functions the samples belong too.
In the above we see that there were 513 samples in the compute_grad
function and 119 of these were matrix multiplication.
- Better crate documentation
- Expose other functions from google's cpuprofiler. This allows more options, status checks and more.
- Can we write a sampling profiler in Rust?!
- Integration with cargo-profiler?
This project has a BSD license to match the gperftools license. Which makes sense, I think?