Measuring one cache line delivery latency

This sample measures single-trip delivery latency of one cache line-sized block between two CPUs of the same core.

How it works

The samples writes a cacheline to memory by a core, which is read it by another core. The cacheline is duplicated and is read by an original core. The roundtrip time is measured and divided by two.

Requirements

• x86

• linux

• multicore CPU

• a C++14-compatible compiler

• cmake 3.18 or later

Build

1. Create build folder

mkdir build && cd build

2. Create cmake auxiliary files

cmake ..

3. Build

cmake --build .

Run

Prerequisite

CPU frequency of the cores involved needs to be precalculated before starting.

It can be done by launching

cat /proc/cpuinfo | grep Hz
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 800.000
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 2800.000
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 800.000
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 800.000
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 2794.449
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 800.000
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 800.000
model name	: Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
cpu MHz		: 800.000

Core selection

Select two cores on which you'll be launching the benchmark (ex, cores 2, 3).

Launch the benchmark

./measure_cache_line_delivery_ns [<num_messages> (default 1M)] [<consumer cpu id> (default 0)] [<producer cpu id> (default 1)] [<cpu freq (HGz)> (default 3.0)]

Examples:

./measure_cache_line_delivery_ns 1000 2 3 4.0
A tool for measuring latency of a single cache line-length message delivery between different cores of a CPU
Usage: ./measure_cache_line_delivery_ns [<num_messages> (default 1M)] [<consumer cpu id> (default 0)] [<producer cpu id> (default 1)] [<cpu freq (HGz)> (default 3.0)]
example: ./measure_cache_line_delivery_ns 1000 5 6 4.0
Running the benchmark with parameters: num_messages=1000, core1=2, core2=3, freq=4GHz...
Single trip is 82ns 

./measure_cache_line_delivery_ns
A tool for measuring latency of a single cache line-length message delivery between different cores of a CPU
Usage: ./measure_cache_line_delivery_ns [<num_messages> (default 1M)] [<consumer cpu id> (default 0)] [<producer cpu id> (default 1)] [<cpu freq (HGz)> (default 3.0)]
example: ./measure_cache_line_delivery_ns 1000 5 6 4.0
Running the benchmark with parameters: num_messages=1000000, core1=0, core2=1, freq=3GHz...
Single trip is 80ns