This program measures the time (ns) that it takes to send/receive a Compare-And-Swap (CAS) message among all the cores. In particular, it makes sure to allocate the message in the memory bound to the sending thread. In other words, it's NUMA aware.
Inspired by core-to-core-latency
USAGE: cpu-latency [options]
Measures average the time (ns) that it takes to send/receive a Compare-And-Swap (CAS) message among all the cores
The results are streamed to 'stdout' as a comma-separated values (CSV) format.
OPTIONS:
-rt, --round-trips <int> Number of times to send and receive messages from core A to core B (Default: 1000)
-r, --repeat <int> Number of times to repeat the experiment per core (Default: 15)
-s, --symmetric <bool> Whether to measure ping-pong latency from core A to core B but not the oposite (Default: true)
--randomize <bool> Whether to randomize the order of cores to measure (Default: true)
-h, --help Display available options
This program requires the hwloc API in order to bind threads and memory. Make sure to have it available on your system, for example:
# Ubuntu/Debian
apt-get install libhwloc-dev
To build, just configure cmake and build its targets:
mkdir cpu-latency/build && cd cpu-latency/build
CMAKE_BUILD_TYPE="Release" cmake ..
cmake --build .
Make sure to have numpy and matplotlib in yor system, for example:
python3 -m pip install -U matplotlib
python3 -m pip install -U numpy
Then you can run the script to measure the latency and plot it with the utility python script
./cpu-latency > cpu-latency.csv
./cpu-latency-plot -i cpu-latency.csv -o cpu-latency.png
Compared to other programs, this program can:
- Take into account NUMA domains (main motivation)
- Randomize order of CPUs to benchmark (it makes a difference!)
- Parameterize the number of round-trips and experiment repetitions
- Choose between symmetric and non-symmetric benchmarks
