2.4 KiB
Profiling
We use gperftools (Google
Performance Tools) for profiling. Note, we also considered using
llvm-xray but found it lacking in
comparison. This will not tell you how long (wall clock time) each function
took, but it will help you determine which functions are the most expensive.
Prequisities
On Linux (Debian), you need to install:
sudo apt install gperftools graphviz
On macOS, you need to install (via homebrew):
brew install gperftools ghostscript graphviz
How to run
Generating profiling graphs
There is a Makefile rule that should just auto-magically work:
make profile
For each profiled function, this will produce two files (a PROF and PDF file). The PROF file is the raw profiling data and the PDF is the human-friendly graph that generated from that profiling data.
Errors on macOS
Note, on macOS there may a lot of "errors" like:
otool-classic: can't open file: /usr/lib/libc++.1.dylib
In my experience, you can ignore these. It's somewhat a known issue and may be resolved later. The PDFs should still generate successfully. I think it's the reason some function names are a hexadecimal address though.
Viewing profiling graphs
On Linux, you can open an individual PDF file like:
xdg-open blob_to_kzg_commitment.pdf
On macOS, you can open an individual PDF file like:
open blob_to_kzg_commitment.pdf
Or, you can open all the PDF files like:
open *.pdf
Interpreting the profiling graphs
These might not make much sense without guidance. From a high-level, this works by polling the instruction pointer (what's being executed) at a specific rate (like once every 5 nanoseconds) and tracking this information. From this, you can infer the relative time each function uses by counting the number of samples that are in each function.
Given a box containing:
my_func 189 (0.6%) of 28758 (96.8%)
- Each box is a unique function.
- Bigger boxes are more expensive.
- Lines between boxes are function calls.
- 189 is the number of profiling samples in this function.
- 0.6% is the percentage of profiling samples in the functions.
- 28758 is the number of profiling samples in this function and its callees.
- 96.8% is the percentage of profiling samples in this function and its callees.