CppNet/doc/efficiency/apache_ab_bench.md

Apache ab Benchmark

Apache ab is a commonly used tool for HTTP server pressure testing. I press test the http test service in the cppnet test directory on different platforms, The command to be executed is:

ab -kc[1-10000] -n100000 http://127.0.0.1:8921/hello

Execute the above command three times each time, take the highest value of QPS among the three times, and draw the test result as a line chart.
In the line chart, the ordinate is QPS and the abscissa is concurrent value.

Linux

environment：

• the system environment is WSL2 in windows
• the operating system is Ubuntu20.04
• system memory 16G DDR3 2666Mhz
• CPU is i7-6700 @ 3.40GHz four cores
• compile optimized -O2
• all tests have 2 threads

Compared with Muduo on Linux，because Linux support REUSE_PORT and EPOLLEXCLUSIVE, so I added a comparison of two different options, The results are shown in the figure below

In the picture cppnet_1 used REUSE_PORT，cppnet_2 used EPOLLEXCLUSIVE.
When use REUSE_PORT, efficiency is slightly better than muduo.
When use EPOLLEXCLUSIVE, the overall performance is better than muduo(muduo does not support the EPOLLEXCLUSIVE setting).

Windows

environment：

• The system environment is physical real machine
• the operating system is Windows 10 @ 19042.985
• system memory 16G DDR3 2666HZ
• CPU is i7-6700 @ 3.40GHz four cores
• compile optimized /O2

Because IOCP is used in the Windows, the system kernel is responsible for managing the thread pool, and each socket wake-up operation will be executed on different threads, which is quite different from Linux and MacOS. Therefore, the pressure test performance of different number of threads is specially listed, The results are shown in the figure below

MacOS

environment：

• The system environment is physical real machine
• the operating system is MacOS @ 10.15.4
• system memory 8G DDR3 2133Mhz
• CPU is i5-8257U @ 1.40GHz four cores
• compile optimized -O2
• 2 threads

The results are shown in the figure below

Result

It is difficult to have both cross platform and high performance. Due to the Linux platform developed first, the windows platform adapted later, the usage of IOCP is very different from epoll and kqueue. There are many thread safe locks on Windows, The performance of IOCP is not displayed, so the performance on Windows is normal.
On Linux and MacOS, most QPS can exceed 100000 under normal concurrency conditions, the performance is excellent.