[求助] nginx 服务器在并发达到 300 就出现响应慢，甚至到 8 秒的情况

服务器版本：centos
服务器配置：阿里云实例规格：ecs.c5.large 2 核 4G
服务器环境：nginx 1.14.1 + php 5.6.36 + mysql 5.7.22

问题描述：在并发量在 300 的时候，不知道为什么响应时间特别长达到了 8 秒甚至 10 秒以上

测试描述：使用 apache 的 ab 压力测试工具测试网站的页面，这个页面什么也没有就是单纯的用 php return 了一个 json 字符串。使用的命令是.\abs -c 300 -n 500 地址

测试结果
Server Port: 443
SSL/TLS Protocol: TLSv1.2,ECDHE-RSA-AES256-GCM-SHA384,2048,256

Document Path: /
Document Length: 166 bytes

Concurrency Level: 300
Time taken for tests: 12.325 seconds
Complete requests: 500
Failed requests: 257
(Connect: 0, Receive: 0, Length: 257, Exceptions: 0)
Non-2xx responses: 243
Total transferred: 124917 bytes
HTML transferred: 47791 bytes
Requests per second: 40.57 [#/sec] (mean)
Time per request: 7395.060 [ms] (mean)
Time per request: 24.650 [ms] (mean, across all concurrent requests)
Transfer rate: 9.90 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 557 1737 783.3 1517 3500
Processing: 30 2514 2053.8 1862 6267
Waiting: 28 2360 2152.5 1711 6267
Total: 1248 4251 2325.3 5031 8428

Percentage of the requests served within a certain time (ms)
50% 5031
66% 5784
75% 6284
80% 6574
90% 7450
95% 7755
98% 8107
99% 8338
100% 8428 (longest request)

配置信息：
nginx.conf

user www www;

worker_processes auto;

error_log /home/wwwlogs/nginx_error.log crit;

pid /usr/local/nginx/logs/nginx.pid;

#Specifies the value for maximum file descriptors that can be opened by this process.
worker_rlimit_nofile 51200;

events
{
use epoll;
worker_connections 51200;
accept_mutex off;
multi_accept on;
}

http
{
include mime.types;
default_type application/octet-stream;

server_names_hash_bucket_size 128;
client_header_buffer_size 32k;
large_client_header_buffers 4 32k;
client_max_body_size 50m;

sendfile on;
tcp_nopush on;

keepalive_timeout 0;

tcp_nodelay on;

fastcgi_connect_timeout 300;
fastcgi_send_timeout 300;
fastcgi_read_timeout 300;
fastcgi_buffer_size 64k;
fastcgi_buffers 4 64k;
fastcgi_busy_buffers_size 128k;
fastcgi_temp_file_write_size 256k;

gzip on;
gzip_min_length 1k;
gzip_buffers 4 16k;
gzip_http_version 1.1;
gzip_comp_level 2;
gzip_types text/plain application/javascript application/x-javascript text/javascript text/css application/xml application/xml+rss;
gzip_vary on;
gzip_proxied expired no-cache no-store private auth;
gzip_disable "MSIE [1-6]\.";

#limit_conn_zone $binary_remote_addr zone=perip:10m;
##If enable limit_conn_zone,add "limit_conn perip 10;" to server section.

server_tokens off;
access_log off;
}

php-fpm.conf

[global]
pid = /usr/local/php/var/run/php-fpm.pid
error_log = /home/wwwlogs/php-fpm.log
log_level = notice

[www]
listen = /tmp/php-cgi.sock
listen.backlog = -1
listen.allowed_clients = 127.0.0.1
listen.owner = www
listen.group = www
listen.mode = 0666
user = www
group = www
pm = static
pm.max_children = 32
pm.start_servers = 30
pm.min_spare_servers = 30
pm.max_spare_servers = 200
request_terminate_timeout = 100
request_slowlog_timeout = 0
slowlog = var/log/slow.log

dragonsunmoon

2019-05-28 22:25:55 +08:00

@q937298063 我推测，因为所有的公网过来的网络请求，都会经过阿里云的防火墙后，再转发到对应的 ECS 的内网 IP 上，而这个公网 IP 的连接数，转发速率是有限制的。
腾讯云也有类似的问题，所以腾讯云也一样有对应的负载均衡服务。
阿里云和腾讯云提供的负载均衡服务实例，就是用在需要对外（对公网）提供大量并发，大量连接（长连接，或同时指支持的连接数非常高）的应用场景。反过来想，如果阿里云，腾讯云不对 ECS 上的网络进行限制，这个负载均衡服务存在的意义就不大了（不好卖钱了嘛）