web 缓存服务器性能比较 nuster vs nginx vs varnish

[sajesemuy]

最近开发了一个基于 HAProxy 的高性能 web 缓存服务器 nuster。nuster 完全兼容 HAProxy，可以根据 url, path, query, header, cookie，请求速率等等来动态生成缓存，并设置有效期。支持 purge，支持前后端 HTTPS。

项目地址 https://github.com/jiangwenyuan/nuster 欢迎大家试用，提供宝贵意见并随手 star，谢谢啦。

本文简单比较了 nuster, nginx 和 varnish 的缓存性能，结果显示 nuster 的 RPS （每秒请求数）单进程模式下大概是 nginx 的 3 倍，多进程下是 nginx 的 2 倍，varnish 的 3 倍。

/helloworld url 包含 hello world文字的结果.

全部结果在 这里

测试环境

Server

两台 linux 服务器, server129 装有 origin web server, cache 服务器 nuster/nginx/varnish 装在 server130.

源服务器上设置 server_tokens off; 来保证 http header server 一样.

硬件

• Intel(R) Xeon(R) CPU X5650 @ 2.67GHz(12 cores)
• RAM 32GB
• 1Gbps ethernet card

软件

• CentOS: 7.4.1708 (Core)
• wrk: 4.0.2-2-g91655b5
• varnish: (varnish-4.1.8 revision d266ac5c6)
• nginx: nginx/1.12.2
• nuster: nuster/1.7.9.1

系统参数

/etc/sysctl.conf

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
net.ipv4.tcp_timestamps        = 1
vm.min_free_kbytes             = 65536
vm.overcommit_memory           = 1

/etc/security/limits.conf

* soft nofile 1000000
* hard nofile 1000000
* soft nproc  1000000
* hard nproc  1000000

配置文件

nuster, 1 core

global
    maxconn 1000000
    cache on data-size 1g
    daemon
    tune.maxaccept -1
defaults
    retries 3
    maxconn 1000000
    option redispatch
    option dontlognull
    timeout client  300s
    timeout connect 300s
    timeout server  300s
    http-reuse always
frontend web1
    bind *:8080
    mode http
    # haproxy removes connection header in HTTP/1.1 while nginx/varnish dont
    # add this to make headers same size
    http-response add-header Connectio1 keep-aliv1
    default_backend app1
backend app1
    balance roundrobin
    mode http
    filter cache on
    cache-rule all ttl 0
    server a2 10.0.0.129:8080

nuster, all cores

global
    maxconn 1000000
    cache on data-size 1g
    daemon
    nbproc 12
    tune.maxaccept -1
defaults
    retries 3
    maxconn 1000000
    option redispatch
    option dontlognull
    timeout client  300s
    timeout connect 300s
    timeout server  300s
    http-reuse always
frontend web1
    bind *:8081
    mode http
    default_backend app1
backend app1
    balance roundrobin
    mode http
    filter cache on
    cache-rule all ttl 0
    server a2 10.0.0.129:8080

nginx, 1 core

user  nginx;
worker_processes  1;
worker_rlimit_nofile 1000000;
error_log  /var/log/nginx/error1.log warn;
pid        /var/run/nginx1.pid;
events {
  worker_connections  1000000;
  use epoll;
  multi_accept on;
}
http {
  include                     /etc/nginx/mime.types;
  default_type                application/octet-stream;
  access_log                  off;
  sendfile                    on;
  server_tokens               off;
  keepalive_timeout           300;
  keepalive_requests          100000;
  tcp_nopush                  on;
  tcp_nodelay                 on;
  client_body_buffer_size     128k;
  client_header_buffer_size   1m;
  large_client_header_buffers 4 4k;
  output_buffers              1 32k;
  postpone_output             1460;
  open_file_cache             max=200000 inactive=20s;
  open_file_cache_valid       30s;
  open_file_cache_min_uses    2;
  open_file_cache_errors      on;
  proxy_cache_path /tmp/cache levels=1:2 keys_zone=STATIC:10m inactive=24h max_size=1g;
  server {
    listen 8082;
    location / {
      proxy_pass        http://10.0.0.129:8080/;
      proxy_cache       STATIC;
      proxy_cache_valid any 1d;
    }
  }
}

nginx, all cores

user  nginx;
worker_processes  auto;
worker_rlimit_nofile 1000000;
error_log  /var/log/nginx/errorall.log warn;
pid        /var/run/nginxall.pid;
events {
  worker_connections  1000000;
  use epoll;
  multi_accept on;
}
http {
  include                     /etc/nginx/mime.types;
  default_type                application/octet-stream;
  access_log                  off;
  sendfile                    on;
  server_tokens               off;
  keepalive_timeout           300;
  keepalive_requests          100000;
  tcp_nopush                  on;
  tcp_nodelay                 on;
  client_body_buffer_size     128k;
  client_header_buffer_size   1m;
  large_client_header_buffers 4 4k;
  output_buffers              1 32k;
  postpone_output             1460;
  open_file_cache             max=200000 inactive=20s;
  open_file_cache_valid       30s;
  open_file_cache_min_uses    2;
  open_file_cache_errors      on;
  proxy_cache_path /tmp/cache_all levels=1:2 keys_zone=STATIC:10m inactive=24h max_size=1g;
  server {
    listen 8083;
    location / {
      proxy_pass        http://10.0.0.129:8080/;
      proxy_cache       STATIC;
      proxy_cache_valid any 1d;
    }
  }
}

varnish

/etc/varnish/default.vcl

vcl 4.0;
backend default {
    .host = "10.0.0.129";
    .port = "8080";
}
sub vcl_recv {
}
sub vcl_backend_response {
    set beresp.ttl = 1d;
}
sub vcl_deliver {
    # remove these headers to make headers same
    unset resp.http.Via;
    unset resp.http.Age;
    unset resp.http.X-Varnish;
}

/etc/varnish/varnish.params

RELOAD_VCL=1
VARNISH_VCL_CONF=/etc/varnish/default.vcl
VARNISH_LISTEN_PORT=8084
VARNISH_ADMIN_LISTEN_ADDRESS=127.0.0.1
VARNISH_ADMIN_LISTEN_PORT=6082
VARNISH_SECRET_FILE=/etc/varnish/secret
VARNISH_STORAGE="malloc,1024M"
VARNISH_USER=varnish
VARNISH_GROUP=varnish

检查 http 头大小

所有 http 头都是一样的

Note that HAProxy removes Connection: Keep-Alive header when its HTTP/1.1 while nginx/varnish do not, so I added Connectio1: keep-aliv1 to make the size same.

See nuster config file above

# curl -is http://10.0.0.130:8080/helloworld
HTTP/1.1 200 OK
Server: nginx
Date: Sun, 05 Nov 2017 07:58:02 GMT
Content-Type: application/octet-stream
Content-Length: 12
Last-Modified: Thu, 26 Oct 2017 08:56:57 GMT
ETag: "59f1a359-c"
Accept-Ranges: bytes
Connectio1: keep-aliv1

Hello World
# curl -is http://10.0.0.130:8080/helloworld | wc -c
255

# curl -is http://10.0.0.130:8081/helloworld
HTTP/1.1 200 OK
Server: nginx
Date: Sun, 05 Nov 2017 07:58:48 GMT
Content-Type: application/octet-stream
Content-Length: 12
Last-Modified: Thu, 26 Oct 2017 08:56:57 GMT
ETag: "59f1a359-c"
Accept-Ranges: bytes
Connectio1: keep-aliv1

Hello World
# curl -is http://10.0.0.130:8081/helloworld | wc -c
255

# curl -is http://10.0.0.130:8082/helloworld
HTTP/1.1 200 OK
Server: nginx
Date: Sun, 05 Nov 2017 07:59:24 GMT
Content-Type: application/octet-stream
Content-Length: 12
Connection: keep-alive
Last-Modified: Thu, 26 Oct 2017 08:56:57 GMT
ETag: "59f1a359-c"
Accept-Ranges: bytes

Hello World
# curl -is http://10.0.0.130:8082/helloworld | wc -c
255

# curl -is http://10.0.0.130:8083/helloworld
HTTP/1.1 200 OK
Server: nginx
Date: Sun, 05 Nov 2017 07:59:31 GMT
Content-Type: application/octet-stream
Content-Length: 12
Connection: keep-alive
Last-Modified: Thu, 26 Oct 2017 08:56:57 GMT
ETag: "59f1a359-c"
Accept-Ranges: bytes

Hello World
# curl -is http://10.0.0.130:8083/helloworld | wc -c
255

# curl -is http://10.0.0.130:8084/helloworld
HTTP/1.1 200 OK
Server: nginx
Date: Sun, 05 Nov 2017 08:00:05 GMT
Content-Type: application/octet-stream
Content-Length: 12
Last-Modified: Thu, 26 Oct 2017 08:56:57 GMT
ETag: "59f1a359-c"
Accept-Ranges: bytes
Connection: keep-alive

Hello World
# curl -is http://10.0.0.130:8084/helloworld | wc -c
255

Benchmark

wrk -c CONN -d 30 -t 100 http://HOST:PORT/FILE

结果

wrk on server129, cache servers on server130, 1Gbps bandwidth

• 1 core
  • 没有用满所有带宽
  • nuster 差不多是 nginx 的 3 倍
• 12 cores
  • 占用所有带宽(see Raw output)
  • 没用满前 nuster 是 nginx 的 2 倍，varnish 的 3 倍
  • 用满时基本差不多

I did the test again with wrk on server130 using 127.0.0.1 since I do not have a 10Gbps network

wrk and cache servers on same host, server130, use 127.0.0.1

• nuster is almost 2 times faster than nginx and varnish
• error occurs with nginx-1core when the connections is 5000

测试输出日志

比较长，详见这里https://github.com/jiangwenyuan/nuster/wiki/Performance-benchmark:-nuster-vs-nginx-vs-varnish

[autogen]

很好的帖子，我来第一个回复，楼主辛苦了