使用 Go 对 Nginx 进行性能测试
jopen
10年前
目前有很多提供Go语言HTTP应用服务的方法，但其中最好的选择取决于每个应用的实际情况。目前，Nginx看起来是每个新项目的标准Web服务器，即使在有其他许多不错Web服务器的情况下。然而，在Nginx上提供Go应用服务的开销是多少呢？我们需要一些nginx的特性参数（vhosts，负载均衡，缓存，等等）或者直接使用Go提供服务？如果你需要nginx，最快的连接机制是什么？这就是在这我试图回答的问题。该基准测试的目的不是要验证Go比nginx的快或慢。那将会很愚蠢。
下面是我们要比较不同的设置：
- Go HTTP standalone (as the control group)
- Nginx proxy to Go HTTP
- Nginx fastcgi to Go TCP FastCGI
- Nginx fastcgi to Go Unix Socket FastCGI
硬件
因为我们将在相同的硬件下比较所有设置，硬件选择的是廉价的一个。这不应该是一个大问题。
- Samsung 笔记本 NP550P5C-AD1BR
- Intel Core i7 3630QM @2.4GHz (quad core, 8 threads)
- CPU caches: (L1: 256KiB, L2: 1MiB, L3: 6MiB)
- RAM 8GiB DDR3 1600MHz
软件
- Ubuntu 13.10 amd64 Saucy Salamander (updated)
- Nginx 1.4.4 (1.4.4-1~saucy0 amd64)
- Go 1.2 (linux/amd64)
- wrk 3.0.4
设置
内核
只需很小的一点调整，将内核的limits调高。如果你对这一变量有更好的想法，请在写在下面评论处：
fs.file-max 9999999 fs.nr_open 9999999 net.core.netdev_max_backlog 4096 net.core.rmem_max 16777216 net.core.somaxconn 65535 net.core.wmem_max 16777216 net.ipv4.ip_forward 0 net.ipv4.ip_local_port_range 1025 65535 net.ipv4.tcp_fin_timeout 30 net.ipv4.tcp_keepalive_time 30 net.ipv4.tcp_max_syn_backlog 20480 net.ipv4.tcp_max_tw_buckets 400000 net.ipv4.tcp_no_metrics_save 1 net.ipv4.tcp_syn_retries 2 net.ipv4.tcp_synack_retries 2 net.ipv4.tcp_tw_recycle 1 net.ipv4.tcp_tw_reuse 1 vm.min_free_kbytes 65536 vm.overcommit_memory 1
Limits
供root和www-data打开的最大文件数限制被配置为200000。
Nginx
有几个必需得Nginx调整。有人跟我说过，我禁用了gzip以保证比较公平。下面是它的配置文件/etc/nginx/nginx.conf：
user www-data; worker_processes auto; worker_rlimit_nofile 200000; pid /var/run/nginx.pid; events { worker_connections 10000; use epoll; multi_accept on; } http { sendfile on; tcp_nopush on; tcp_nodelay on; keepalive_timeout 300; keepalive_requests 10000; types_hash_max_size 2048; open_file_cache max=200000 inactive=300s; open_file_cache_valid 300s; open_file_cache_min_uses 2; open_file_cache_errors on; server_tokens off; dav_methods off; include /etc/nginx/mime.types; default_type application/octet-stream; access_log /var/log/nginx/access.log combined; error_log /var/log/nginx/error.log warn; gzip off; gzip_vary off; include /etc/nginx/conf.d/*.conf; include /etc/nginx/sites-enabled/*.conf; } Nginx vhosts
upstream go_http { server 127.0.0.1:8080; keepalive 300; } server { listen 80; server_name go.http; access_log off; error_log /dev/null crit; location / { proxy_pass http://go_http; proxy_http_version 1.1; proxy_set_header Connection ""; } } upstream go_fcgi_tcp { server 127.0.0.1:9001; keepalive 300; } server { listen 80; server_name go.fcgi.tcp; access_log off; error_log /dev/null crit; location / { include fastcgi_params; fastcgi_keep_conn on; fastcgi_pass go_fcgi_tcp; } } upstream go_fcgi_unix { server unix:/tmp/go.sock; keepalive 300; } server { listen 80; server_name go.fcgi.unix; access_log off; error_log /dev/null crit; location / { include fastcgi_params; fastcgi_keep_conn on; fastcgi_pass go_fcgi_unix; } } Go源码
package main import ( "fmt" "log" "net" "net/http" "net/http/fcgi" "os" "os/signal" "syscall" ) var ( abort bool ) const ( SOCK = "/tmp/go.sock" ) type Server struct { } func (s Server) ServeHTTP(w http.ResponseWriter, r *http.Request) { body := "Hello World\n" // Try to keep the same amount of headers w.Header().Set("Server", "gophr") w.Header().Set("Connection", "keep-alive") w.Header().Set("Content-Type", "text/plain") w.Header().Set("Content-Length", fmt.Sprint(len(body))) fmt.Fprint(w, body) } func main() { sigchan := make(chan os.Signal, 1) signal.Notify(sigchan, os.Interrupt) signal.Notify(sigchan, syscall.SIGTERM) server := Server{} go func() { http.Handle("/", server) if err := http.ListenAndServe(":8080", nil); err != nil { log.Fatal(err) } }() go func() { tcp, err := net.Listen("tcp", ":9001") if err != nil { log.Fatal(err) } fcgi.Serve(tcp, server) }() go func() { unix, err := net.Listen("unix", SOCK) if err != nil { log.Fatal(err) } fcgi.Serve(unix, server) }() <-sigchan if err := os.Remove(SOCK); err != nil { log.Fatal(err) } } 检查HTTP header
为公平起见，所有的请求必需大小相同。
$ curl -sI http://127.0.0.1:8080/ HTTP/1.1 200 OK Connection: keep-alive Content-Length: 12 Content-Type: text/plain Server: gophr Date: Sun, 15 Dec 2013 14:59:14 GMT $ curl -sI http://127.0.0.1:8080/ | wc -c 141
$ curl -sI http://go.http/ HTTP/1.1 200 OK Server: nginx Date: Sun, 15 Dec 2013 14:59:31 GMT Content-Type: text/plain Content-Length: 12 Connection: keep-alive $ curl -sI http://go.http/ | wc -c 141
$ curl -sI http://go.fcgi.tcp/ HTTP/1.1 200 OK Content-Type: text/plain Content-Length: 12 Connection: keep-alive Date: Sun, 15 Dec 2013 14:59:40 GMT Server: gophr $ curl -sI http://go.fcgi.tcp/ | wc -c 141
$ curl -sI http://go.fcgi.unix/ HTTP/1.1 200 OK Content-Type: text/plain Content-Length: 12 Connection: keep-alive Date: Sun, 15 Dec 2013 15:00:15 GMT Server: gophr $ curl -sI http://go.fcgi.unix/ | wc -c 141
启动引擎
- 使用sysctl配置内核
- 配置Nginx
- 配置Nginx vhosts
- 用www-data启动服务
- 运行基准测试
基准测试
GOMAXPROCS = 1
Go standalone
# wrk -t100 -c5000 -d30s http://127.0.0.1:8080/ Running 30s test @ http://127.0.0.1:8080/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 116.96ms 17.76ms 173.96ms 85.31% Req/Sec 429.16 49.20 589.00 69.44% 1281567 requests in 29.98s, 215.11MB read Requests/sec: 42745.15 Transfer/sec: 7.17MB
Nginx + Go through HTTP
# wrk -t100 -c5000 -d30s http://go.http/ Running 30s test @ http://go.http/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 124.57ms 18.26ms 209.70ms 80.17% Req/Sec 406.29 56.94 0.87k 89.41% 1198450 requests in 29.97s, 201.16MB read Requests/sec: 39991.57 Transfer/sec: 6.71MB
Nginx + Go through FastCGI TCP
# wrk -t100 -c5000 -d30s http://go.fcgi.tcp/ Running 30s test @ http://go.fcgi.tcp/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 514.57ms 119.80ms 1.21s 71.85% Req/Sec 97.18 22.56 263.00 79.59% 287416 requests in 30.00s, 48.24MB read Socket errors: connect 0, read 0, write 0, timeout 661 Requests/sec: 9580.75 Transfer/sec: 1.61MB
Nginx + Go through FastCGI Unix Socket
# wrk -t100 -c5000 -d30s http://go.fcgi.unix/ Running 30s test @ http://go.fcgi.unix/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 425.64ms 80.53ms 925.03ms 76.88% Req/Sec 117.03 22.13 255.00 81.30% 350162 requests in 30.00s, 58.77MB read Socket errors: connect 0, read 0, write 0, timeout 210 Requests/sec: 11670.72 Transfer/sec: 1.96MB
GOMAXPROCS = 8
Go standalone
# wrk -t100 -c5000 -d30s http://127.0.0.1:8080/ Running 30s test @ http://127.0.0.1:8080/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 39.25ms 8.49ms 86.45ms 81.39% Req/Sec 1.29k 129.27 1.79k 69.23% 3837995 requests in 29.89s, 644.19MB read Requests/sec: 128402.88 Transfer/sec: 21.55MB
Nginx + Go through HTTP
# wrk -t100 -c5000 -d30s http://go.http/ Running 30s test @ http://go.http/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 336.77ms 297.88ms 632.52ms 60.16% Req/Sec 2.36k 2.99k 19.11k 84.83% 2232068 requests in 29.98s, 374.64MB read Requests/sec: 74442.91 Transfer/sec: 12.49MB
Nginx + Go through FastCGI TCP
# wrk -t100 -c5000 -d30s http://go.fcgi.tcp/ Running 30s test @ http://go.fcgi.tcp/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 217.69ms 121.22ms 1.80s 75.14% Req/Sec 263.09 102.78 629.00 62.54% 721027 requests in 30.01s, 121.02MB read Socket errors: connect 0, read 0, write 176, timeout 1343 Requests/sec: 24026.50 Transfer/sec: 4.03MB
Nginx + Go through FastCGI Unix Socket
# wrk -t100 -c5000 -d30s http://go.fcgi.unix/ Running 30s test @ http://go.fcgi.unix/ 100 threads and 5000 connections Thread Stats Avg Stdev Max +/- Stdev Latency 694.32ms 332.27ms 1.79s 62.13% Req/Sec 646.86 669.65 6.11k 87.80% 909836 requests in 30.00s, 152.71MB read Requests/sec: 30324.77 Transfer/sec: 5.09MB
结论
第一组基准测试时一些Nginx的设置还没有很好的优化（启用gzip，Go的后端没有使用keep-alive连接）。当改为wrk以及按建议优化Nginx后结果有较大差异。
当GOMAXPROCS=1时，Nginx的开销不是那么大，但当OMAXPROCS=8时差异就很大了。以后可能会再试一下其他设置。如果你需要使用Nginx像虚拟主机，负载均衡，缓存等特性，使用HTTP proxy，别使用FastCGI。有些人说Go的FastCGI还没有被很好优化，这也许就是测试结果中巨大差异的原因。