smallnest / 1m-go-tcp-server

benchmarks for implementation of servers which support 1 million connections

Benchmark for implementation of servers that support 1m connections

inspired by handling 1M websockets connections in Go

Servers

1. 1_simple_tcp_server： a 1m-connections server implemented based on goroutines per connection
2. 2_epoll_server: a 1m-connections server implemented based on epoll
3. 3_epoll_server_throughputs: add throughputs and latency test for 2_epoll_server
4. 4_epoll_client: implement the client based on epoll
5. 5_multiple_client: use multiple epoll to manage connections in client
6. 6_multiple_server: use multiple epoll to manage connections in server
7. 7_server_prefork: use prefork style of apache to implement server
8. 8_server_workerpool: use Reactor pattern to implement multiple event loops
9. 9_few_clients_high_throughputs: a simple goroutines per connection server for test throughtputs and latency
10. 10_io_intensive_epoll_server: an io-bound multiple epoll server
11. 11_io_intensive_goroutine: an io-bound goroutines per connection server
12. 12_cpu_intensive_epoll_server: a cpu-bound multiple epoll server
13. 13_cpu_intensive_goroutine: an cpu-bound goroutines per connection server

Test Environment

• two E5-2630 V4 cpus, total 20 cores, 40 logicial cores.
• 32G memory

tune the linux:

sysctl -w fs.file-max=2000500
sysctl -w fs.nr_open=2000500
sysctl -w net.nf_conntrack_max=2000500
ulimit -n 2000500

sysctl -w net.ipv4.tcp_tw_recycle=1
sysctl -w net.ipv4.tcp_tw_reuse=1

client sends the next request only when it has received the response. it has not used the pipeline style to test.

Benchmarks

1m connections

	throughputs (tps)	latency
goroutine-per-conn	202830	4.9s
single epoll(both server and client)	42495	23s
single epoll server	42402	0.8s
multiple epoll server	197814	0.9s
prefork	444415	1.5s
workerpool	190022	0.3s

中文介绍:

1. 百万 Go TCP 连接的思考: epoll方式减少资源占用
2. 百万 Go TCP 连接的思考2: 百万连接的服务器的性能
3. 百万 Go TCP 连接的思考3: 低连接场景下的服务器的吞吐和延迟