asio-async-model-perftest

Performance test to measure the relative performance between asio's different async handlers

This project implements a simple async server work-flow that "reads" a number of input buffers, process the data and then "writes" a number of output buffers.

In other words, we aim to simulate typical memory usage of a server application - draining the CPU cache.

The same logic is implemented using:

• Asio async methods with completion callbacks
• Asio stackful coroutines (via spawn())
• Asio legacy stackless coroutines
• Asio stackless C++ 20 coroutines (via co_spawn())

The motivation is to measure the performance regression using the more convenient methods, like stateful coroutines or C++ 20 coroutines.

In addition, the test provides implementations of layered C++20 generators, using lazy fetching/flushing of data at the lowest layer. This test is implemented for:

• Asio stackful coroutines (via spawn())
• Asio stackless C++ 20 coroutines (via co_spawn())

Requirements

• gcc 11 or newer
• Boost 1.76.0 or newer (1.81 is broken, 1.82 works).

Status

In progress. Once done, I will post results and graphs with results from a number of hardware configurations.

Run all tests:

Example on how to run all the performance-tests.

From the build/bin directory with boost in /opt/boost/boost_1_82_0:

LD_LIBRARY_PATH=/opt/boost/boost_1_82_0/stage/lib PATH="$PATH:."  bash ~/src/asio-async-model-perftest/run-all.sh