Skip to content

Latest commit

 

History

History
62 lines (46 loc) · 3.44 KB

Benchmark.md

File metadata and controls

62 lines (46 loc) · 3.44 KB
Raw

GO echo-server benchmark

requirements to run the benchmarks

Linux 5.11 or higher

programs under test

  • echo-server using go-uring net-reactor as I/O backend
  • echo-server using go-uring net-reactor as I/O backend with SQ_POLL enabled
  • echo-server built with amd64_atomic tag using go-uring net-reactor as I/O backend
  • echo-server using net package (net.TCPListener and net.TCPConn) - with netpoller - default GO I/O backend

Check main.go for familiarization.

system specs

  • Intel(R) Core(TM) i5-7500 CPU @ 3.40GHz, 16GB RAM, 4 physical cores
  • Virtual box with Ubuntu 20.04 (8GB RAM, 4 physical cores)
  • Linux 5.16rc3

benchmark tool

  • echo server and benchmarks run at virtual machine as is
  • Rust echo bench: https://github.com/haraldh/rust_echo_bench
  • cargo run --release -- --address "localhost:8080" --number {number of clients} --duration 60 --length {msg size}
  • 5 runs for each combination of 128 and 1024 bytes message size with 100, 500 and 1000 clients
  • bench.sh script using for benchmarking

Results

Run echo server and benchmarks on 4 CPU cores

c: 100 bytes: 128 c: 100 bytes: 1024 c: 500 bytes: 128 c: 500 bytes: 1024 c: 1000 bytes: 128 c: 1000 bytes: 1024
net/http 132664 139206 133039 139171 133480 139617
go-uring 34202 33159 147362 139313 158483 154194
go-uring + amd64_atomic tag 34043 33328 146213 142255 158054 154687
go-uring SQ_POLL mode 24406 22847 134863 130668 127896 122601

Q: why go-uring so bad for 100 connections?

As you can see in benchmarks there is a bad performance for go-uring echo-servers on 100-connections test cases. This is not a problem of io_uring, there is a problem of current net-reactor implementation. For 100-conn's test case echo-server's with go-uring does not load the processor completely, maybe I fix this in the future. Anyway, goal of this benchmark is not a win GO netpoller, it's a show that io_uring can be an alternative of netpoller, and in some cases, it also can over-perform netpoller.

By the way, there is a simple fix for bad performance on 100-conn's - change startup parameters for creating less io_uring(-ring-count flag) instances at echo-server.

Benchmark script

Echo-server's startup parameters:

  • net/http - ./main -mode default
  • go-uring - ./main -mode uring -ring-count 6 -wp-count 2
  • go-uring + amd64_atomic tag - ./main -mode uring -ring-count 6 -wp-count 2
  • go-uring SQ_POLL mode + amd64_atomic tag - ./main -mode uring-sq-poll -ring-count 2 -wp-count 2

Source code

  • $1 - path to executable (with the necessary options)
  • $2 - path to rust_echo_bench Cargo.toml

Run example:

  ./bench.sh "./main -mode uring -ring-count 4 -wp-count 4" ~/rust/rust_echo_bench/Cargo.toml