Skip to content

bench.md

Latest commit

 

History

History
102 lines (69 loc) · 3.37 KB

bench.md

File metadata and controls

102 lines (69 loc) · 3.37 KB

So far, only few benchmarks is done... Mainly focus on the overhead of the framework using C++20 coroutine.

Benchmark tools:

webbench, apache bench, rust_echo_bench

Baseline:

item baseline target
httpd webserver_liburing demo/http_server
echo server io_uring-echo-server demo/echo_server

baseline:

target:

Basic results:

Since the number is not matched with the one in early io_uring benchmarks with other kernel version, here only post a comparison instead of absolute numbers.

Without SQPoll, 100 concurrent clients:

utility QPS Throughput
webbench (120 bytes) 102% 99%
apache bench (120 bytes) 101% 102%
rust-echo-bench (512 bytes) 96% N/A
rust-echo-bench (2018 bytes) 102% N/A

More accurate data would be measured on native linux machine later.

Configuration:

Machine Spec:

  • Intel Core i7 6700h 2.60GHz 16GiB
  • WSL2 with kernel 5.17.1 (user-compiled wsl kernel, M$ haven't released the latest one)

Conventions:

  • In http bench, use a same 404 response packet on memory cache prevent irrelevant error. Run bench for 2 times each, gather the results and compute the average.
  • In echo bench, the setting refer to here, though, couldn't reproduce the 100K QPS on both baseline and target programs (around 30K~), if this is relevant to the machine spec or kernel version it remain to be found out.

webbench:

webbench -c 100 -t 60 http://127.0.0.1:PORT/404/

ab:

ab -c 100 -n 100000 http://127.0.0.1:PORT/404/

rust-echo-bench:

check here.

What about epoll:

Run two same configuration http server, both use HTTP/1.0 and no keepalive option, the http_server using io_uring has gain about 2X QPS than epoll based http server.

Factors on different framework may affect the result, more testings need to be done. A practical plan is to separate the io_context implementation and abstraction, then build the abstraction on top of both epoll, io_uring and even IORing in windows 11 insider channel.

A not-controlled-comparison HTTP/1.0 server bench result using apache bench:

#### without Connection: keepalive option ####
# epoll based, use some regular epoll-based framework 
Requests per second:    5517.08 [#/sec] (mean)
# io_uring based, and use libcoring (with SQPOLL off)
Requests per second:    15703.96 [#/sec] (mean)
# io_uring based, and use libcoring (with SQPOLL on)
Requests per second:    19503.34 [#/sec] (mean)