实现一台服务器的百万并发，服务器支撑百万连接会出现哪些问题，如何排查与解决这些问题 是本文的重点

服务器 采用 1台 centos7 12G 1核虚拟机

客户端 采用 2台 centos7 3G 1核虚拟机

服务器代码：单reactor单线程，IO多路复用使用epoll

客户端代码：IO多路复用使用epoll，每个客户端发51w个连接，每个连接发送一次数据，读取一次数据之后不再发送数据

由于fd的数量未知，这里设计ntyreactor 里面包含 eventblock ，eventblock 包含1024个fd。每个fd通过 fd/1024定位到在第几个eventblock，通过fd%1024定位到在eventblock第几个位置。

struct ntyevent { int fd; int events; void *arg; NCALLBACK callback; int status; char buffer[BUFFER_LENGTH]; int length;};struct eventblock { struct eventblock *next; struct ntyevent *events;};struct ntyreactor { int epfd; int blkcnt; struct eventblock *evblk;};

#include <stdio.h>#include <stdlib.h>#include <string.h>#include <sys/socket.h>#include <sys/epoll.h>#include <arpa/inet.h>#include <fcntl.h>#include <unistd.h>#include <errno.h>#define BUFFER_LENGTH 4096#define MAX_EPOLL_EVENTS 1024#define SERVER_PORT 8081#define PORT_COUNT 100typedef int (*NCALLBACK)(int, int, void *);struct ntyevent { int fd; int events; void *arg; NCALLBACK callback; int status; char buffer[BUFFER_LENGTH]; int length;};struct eventblock { struct eventblock *next; struct ntyevent *events;};struct ntyreactor { int epfd; int blkcnt; struct eventblock *evblk;};int recv_cb(int fd, int events, void *arg);int send_cb(int fd, int events, void *arg);struct ntyevent *ntyreactor_find_event_idx(struct ntyreactor *reactor, int sockfd);void nty_event_set(struct ntyevent *ev, int fd, NCALLBACK *callback, void *arg) { ev->fd = fd; ev->callback = callback; ev->events = 0; ev->arg = arg;}int nty_event_add(int epfd, int events, struct ntyevent *ev) { struct epoll_event ep_ev = {0, {0}}; ep_ev.data.ptr = ev; ep_ev.events = ev->events = events; int op; if (ev->status == 1) { op = EPOLL_CTL_MOD; } else { op = EPOLL_CTL_ADD; ev->status = 1; } if (epoll_ctl(epfd, op, ev->fd, &ep_ev) < 0) { printf("event add failed [fd=%d], events[%d]\n", ev->fd, events); return -1; } return 0;}int nty_event_del(int epfd, struct ntyevent *ev) { struct epoll_event ep_ev = {0, {0}}; if (ev->status != 1) { return -1; } ep_ev.data.ptr = ev; ev->status = 0; epoll_ctl(epfd, EPOLL_CTL_DEL, ev->fd, &ep_ev); return 0;}int recv_cb(int fd, int events, void *arg) { struct ntyreactor *reactor = (struct ntyreactor *) arg; struct ntyevent *ev = ntyreactor_find_event_idx(reactor, fd); int len = recv(fd, ev->buffer, BUFFER_LENGTH, 0); // nty_event_del(reactor->epfd, ev); if (len > 0) { ev->length = len; ev->buffer[len] = '\0';// printf("recv[%d]:%s\n", fd, ev->buffer); printf("recv fd=[%d\n", fd); nty_event_set(ev, fd, send_cb, reactor); nty_event_add(reactor->epfd, EPOLLOUT, ev); } else if (len == 0) { close(ev->fd); //printf("[fd=%d] pos[%ld], closed\n", fd, ev-reactor->events); } else { close(ev->fd);// printf("recv[fd=%d] error[%d]:%s\n", fd, errno, strerror(errno)); } return len;}int send_cb(int fd, int events, void *arg) { struct ntyreactor *reactor = (struct ntyreactor *) arg; struct ntyevent *ev = ntyreactor_find_event_idx(reactor, fd); int len = send(fd, ev->buffer, ev->length, 0); if (len > 0) {// printf("send[fd=%d], [%d]%s\n", fd, len, ev->buffer); printf("send fd=[%d\n]", fd); nty_event_del(reactor->epfd, ev); nty_event_set(ev, fd, recv_cb, reactor); nty_event_add(reactor->epfd, EPOLLIN, ev); } else { nty_event_del(reactor->epfd, ev); close(ev->fd); printf("send[fd=%d] error %s\n", fd, strerror(errno)); } return len;}int accept_cb(int fd, int events, void *arg) {//非阻塞 struct ntyreactor *reactor = (struct ntyreactor *) arg; if (reactor == NULL) return -1; struct sockaddr_in client_addr; socklen_t len = sizeof(client_addr); int clientfd; if ((clientfd = accept(fd, (struct sockaddr *) &client_addr, &len)) == -1) { printf("accept: %s\n", strerror(errno)); return -1; } if ((fcntl(clientfd, F_SETFL, O_NONBLOCK)) < 0) { printf("%s: fcntl nonblocking failed, %d\n", __func__, MAX_EPOLL_EVENTS); return -1; } struct ntyevent *event = ntyreactor_find_event_idx(reactor, clientfd); nty_event_set(event, clientfd, recv_cb, reactor); nty_event_add(reactor->epfd, EPOLLIN, event); printf("new connect [%s:%d], pos[%d]\n", inet_ntoa(client_addr.sin_addr), ntohs(client_addr.sin_port), clientfd); return 0;}int init_sock(short port) { int fd = socket(AF_INET, SOCK_STREAM, 0); fcntl(fd, F_SETFL, O_NONBLOCK); struct sockaddr_in server_addr; memset(&server_addr, 0, sizeof(server_addr)); server_addr.sin_family = AF_INET; server_addr.sin_addr.s_addr = htonl(INADDR_ANY); server_addr.sin_port = htons(port); bind(fd, (struct sockaddr *) &server_addr, sizeof(server_addr)); if (listen(fd, 20) < 0) { printf("listen failed : %s\n", strerror(errno)); } return fd;}int ntyreactor_alloc(struct ntyreactor *reactor) { if (reactor == NULL) return -1; if (reactor->evblk == NULL) return -1; struct eventblock *blk = reactor->evblk; while (blk->next != NULL) { blk = blk->next; } struct ntyevent *evs = (struct ntyevent *) malloc((MAX_EPOLL_EVENTS) * sizeof(struct ntyevent)); if (evs == NULL) { printf("ntyreactor_alloc ntyevents failed\n"); return -2; } memset(evs, 0, (MAX_EPOLL_EVENTS) * sizeof(struct ntyevent)); struct eventblock *block = (struct eventblock *) malloc(sizeof(struct eventblock)); if (block == NULL) { printf("ntyreactor_alloc eventblock failed\n"); return -2; } memset(block, 0, sizeof(struct eventblock)); block->events = evs; block->next = NULL; blk->next = block; reactor->blkcnt++; // return 0;}struct ntyevent *ntyreactor_find_event_idx(struct ntyreactor *reactor, int sockfd) { int blkidx = sockfd / MAX_EPOLL_EVENTS; while (blkidx >= reactor->blkcnt) { ntyreactor_alloc(reactor); } int i = 0; struct eventblock *blk = reactor->evblk; while (i++ < blkidx && blk != NULL) { blk = blk->next; } return &blk->events[sockfd % MAX_EPOLL_EVENTS];}int ntyreactor_init(struct ntyreactor *reactor) { if (reactor == NULL) return -1; memset(reactor, 0, sizeof(struct ntyreactor)); reactor->epfd = epoll_create(1); if (reactor->epfd <= 0) { printf("create epfd in %s err %s\n", __func__, strerror(errno)); return -2; } struct ntyevent *evs = (struct ntyevent *) malloc((MAX_EPOLL_EVENTS) * sizeof(struct ntyevent)); if (evs == NULL) { printf("ntyreactor_alloc ntyevents failed\n"); return -2; } memset(evs, 0, (MAX_EPOLL_EVENTS) * sizeof(struct ntyevent)); struct eventblock *block = (struct eventblock *) malloc(sizeof(struct eventblock)); if (block == NULL) { printf("ntyreactor_alloc eventblock failed\n"); return -2; } memset(block, 0, sizeof(struct eventblock)); block->events = evs; block->next = NULL; reactor->evblk = block; reactor->blkcnt = 1; return 0;}int ntyreactor_destory(struct ntyreactor *reactor) { close(reactor->epfd); //free(reactor->events); struct eventblock *blk = reactor->evblk; struct eventblock *blk_next = NULL; while (blk != NULL) { blk_next = blk->next; free(blk->events); free(blk); blk = blk_next; } return 0;}int ntyreactor_addlistener(struct ntyreactor *reactor, int sockfd, NCALLBACK *acceptor) { if (reactor == NULL) return -1; if (reactor->evblk == NULL) return -1; struct ntyevent *event = ntyreactor_find_event_idx(reactor, sockfd); nty_event_set(event, sockfd, acceptor, reactor); nty_event_add(reactor->epfd, EPOLLIN, event); return 0;}_Noreturn int ntyreactor_run(struct ntyreactor *reactor) { if (reactor == NULL) return -1; if (reactor->epfd < 0) return -1; if (reactor->evblk == NULL) return -1; struct epoll_event events[MAX_EPOLL_EVENTS + 1]; int i; while (1) { int nready = epoll_wait(reactor->epfd, events, MAX_EPOLL_EVENTS, 1000); if (nready < 0) { printf("epoll_wait error, exit\n"); continue; } for (i = 0; i < nready; i++) { struct ntyevent *ev = (struct ntyevent *) events[i].data.ptr; if ((events[i].events & EPOLLIN) && (ev->events & EPOLLIN)) { ev->callback(ev->fd, events[i].events, ev->arg); } if ((events[i].events & EPOLLOUT) && (ev->events & EPOLLOUT)) { ev->callback(ev->fd, events[i].events, ev->arg); } } }}// <remoteip, remoteport, localip, localport,protocol>int main(int argc, char *argv[]) { unsigned short port = SERVER_PORT; // listen 8081 if (argc == 2) { port = atoi(argv[1]); } struct ntyreactor *reactor = (struct ntyreactor *) malloc(sizeof(struct ntyreactor)); ntyreactor_init(reactor); int i = 0; int sockfds[PORT_COUNT] = {0}; for (i = 0; i < PORT_COUNT; i++) { sockfds[i] = init_sock(port + i); ntyreactor_addlistener(reactor, sockfds[i], accept_cb); } ntyreactor_run(reactor); ntyreactor_destory(reactor); for (i = 0; i < PORT_COUNT; i++) { close(sockfds[i]); } free(reactor); return 0;}

客户端代码

#include <stdio.h>#include <string.h>#include <stdlib.h>#include <sys/types.h>#include <sys/socket.h>#include <sys/epoll.h>#include <errno.h>#include <netinet/tcp.h>#include <arpa/inet.h>#include <netdb.h>#include <fcntl.h>#include <sys/time.h>#include <unistd.h>#define MAX_BUFFER128#define MAX_EPOLLSIZE(384*1024)#define MAX_PORT100#define TIME_SUB_MS(tv1, tv2) ((tv1.tv_sec - tv2.tv_sec) * 1000 + (tv1.tv_usec - tv2.tv_usec) / 1000)int isContinue = 0;static int ntySetNonblock(int fd) {int flags;flags = fcntl(fd, F_GETFL, 0);if (flags < 0) return flags;flags |= O_NONBLOCK;if (fcntl(fd, F_SETFL, flags) < 0) return -1;return 0;}static int ntySetReUseAddr(int fd) {int reuse = 1;return setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, (char *)&reuse, sizeof(reuse));}int main(int argc, char **argv) {if (argc <= 2) {printf("Usage: %s ip port\n", argv[0]);exit(0);}const char *ip = argv[1];int port = atoi(argv[2]);int connections = 0;char buffer[128] = {0};int i = 0, index = 0;struct epoll_event events[MAX_EPOLLSIZE];int epoll_fd = epoll_create(MAX_EPOLLSIZE);strcpy(buffer, " Data From MulClient\n");struct sockaddr_in addr;memset(&addr, 0, sizeof(struct sockaddr_in));addr.sin_family = AF_INET;addr.sin_addr.s_addr = inet_addr(ip);struct timeval tv_begin;gettimeofday(&tv_begin, NULL);while (1) {if (++index >= MAX_PORT) index = 0;struct epoll_event ev;int sockfd = 0;if (connections < 340000 && !isContinue) {sockfd = socket(AF_INET, SOCK_STREAM, 0);if (sockfd == -1) {perror("socket");goto err;}//ntySetReUseAddr(sockfd);addr.sin_port = htons(port+index);if (connect(sockfd, (struct sockaddr*)&addr, sizeof(struct sockaddr_in)) < 0) {perror("connect");goto err;}ntySetNonblock(sockfd);ntySetReUseAddr(sockfd);sprintf(buffer, "Hello Server: client --> %d\n", connections);send(sockfd, buffer, strlen(buffer), 0);ev.data.fd = sockfd;ev.events = EPOLLIN | EPOLLOUT;epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sockfd, &ev);connections ++;}//connections ++;if (connections % 1000 == 999 || connections >= 340000) {struct timeval tv_cur;memcpy(&tv_cur, &tv_begin, sizeof(struct timeval));gettimeofday(&tv_begin, NULL);int time_used = TIME_SUB_MS(tv_begin, tv_cur);printf("connections: %d, sockfd:%d, time_used:%d\n", connections, sockfd, time_used);int nfds = epoll_wait(epoll_fd, events, connections, 100);for (i = 0;i < nfds;i ++) {int clientfd = events[i].data.fd;if (events[i].events & EPOLLOUT) {sprintf(buffer, "data from %d\n", clientfd);send(sockfd, buffer, strlen(buffer), 0);} else if (events[i].events & EPOLLIN) {char rBuffer[MAX_BUFFER] = {0};ssize_t length = recv(sockfd, rBuffer, MAX_BUFFER, 0);if (length > 0) {printf(" RecvBuffer:%s\n", rBuffer);if (!strcmp(rBuffer, "quit")) {isContinue = 0;}} else if (length == 0) {printf(" Disconnect clientfd:%d\n", clientfd);connections --;close(clientfd);} else {if (errno == EINTR) continue;printf(" Error clientfd:%d, errno:%d\n", clientfd, errno);close(clientfd);}} else {printf(" clientfd:%d, errno:%d\n", clientfd, errno);close(clientfd);}}}usleep(1 * 1000);}return 0;err:printf("error : %s\n", strerror(errno));return 0;}

相关视频推荐

准备4台虚拟机，一起来实现服务器百万级并发

linux多线程之epoll原理剖析与reactor原理及应用

学习地址：C/C++Linux服务器开发/后台架构师【零声教育】-学习视频教程-腾讯课堂

需要C/C++ Linux服务器架构师学习资料加qun812855908（资料包括C/C++，Linux，golang技术，内核，Nginx，ZeroMQ，MySQL，Redis，fastdfs，MongoDB，ZK，流媒体，CDN，P2P，K8S，Docker，TCP/IP，协程，DPDK，ffmpeg，大厂面试题 等）

确定问题

程序执行到一半，创建了1023个连接后，报错Too many open files

//服务端new connect [192.168.109.101:36994], pos[1019]new connect [192.168.109.101:55832], pos[1020]new connect [192.168.109.101:43460], pos[1021]new connect [192.168.109.101:59938], pos[1022]new connect [192.168.109.101:46098], pos[1023]accept: Too many open filesaccept: Too many open files//客户端connect: Connection refusederror : Connection refused

怀疑是文件系统默认允许打开文件描述符数量个数（默认1024）的限制，使用ulimit -a查看open files的数量

open files：一个进程能够打开文件描述符的数量

[root@master temp]# ulimit -acore file size (blocks, -c) 0data seg size (kbytes, -d) unlimitedscheduling priority (-e) 0file size (blocks, -f) unlimitedpending signals (-i) 47748max locked memory (kbytes, -l) 64max memory size (kbytes, -m) unlimitedopen files (-n) 1024pipe size (512 bytes, -p) 8POSIX message queues (bytes, -q) 819200real-time priority (-r) 0stack size (kbytes, -s) 8192cpu time (seconds, -t) unlimitedmax user processes (-u) 47748virtual memory (kbytes, -v) unlimitedfile locks (-x) unlimited

那么我们把open files调大一点点，看是否会停在2047，如果是，则说明问题就是open files太小的问题，实验发现就是这个原因。

[root@master temp]# ulimit -n 2048[root@master temp]# ulimit -acore file size (blocks, -c) 0data seg size (kbytes, -d) unlimitedscheduling priority (-e) 0file size (blocks, -f) unlimitedpending signals (-i) 47748max locked memory (kbytes, -l) 64max memory size (kbytes, -m) unlimitedopen files (-n) 2048pipe size (512 bytes, -p) 8POSIX message queues (bytes, -q) 819200real-time priority (-r) 0stack size (kbytes, -s) 8192cpu time (seconds, -t) unlimitedmax user processes (-u) 47748virtual memory (kbytes, -v) unlimitedfile locks (-x) unlimitednew connect [192.168.109.101:53996], pos[2046]new connect [192.168.109.101:60742], pos[2047]accept: Too many open files

解决问题

注意这里修改的是：一个进程能够打开文件描述符的数量

[root@master temp]# vim /etc/security/limits.conf# 修改[root@master temp]# reboot# 重启生效

* soft nofile 1048576* hard nofile 1048576软限制：超出软限制会发出警告硬限制：绝对限制,在任何情况下都不允许用户超过这个限制

这里还需要注意一点：file-max: 系统一共可以打开的最大文件数（所有进程加起来）

[root@master temp]# cat /proc/sys/fs/file-max1202172

# 编辑内核参数配置文件vim /etc/sysctl.conf# 修改fs.file-max参数fs.file-max = 1048576# 重新加载配置文件sysctl -p

另外这里建议ulimit -n 和limits.conf里nofile 设定最好不要超过/proc/sys/fs/file-max的值（虽然我测试了超过也没关系），这个小问题仁者见仁智者见智了。

确定问题

现在的环境背景：服务器只开放一个端口，客户端不断的去请求去连接。然后客户端error : Cannot assign requested address

Cannot assign requested address这代表着客户端端口耗尽，我们先来看看如何确定一个fd，反过来说一个fd代表着什么

socket fd --- < 源IP地址 , 源端口 , 目的IP地址 , 目的端口 , 协议 > 一个fd就是一个五元组，在现在的环境中，五元组里面确定了四个，所以最多创建 1 * 源端口 * 1 * 1 * 1个fd

# 服务端new connect [192.168.109.101:57921], pos[28234]new connect [192.168.109.101:57923], pos[28235]send[fd=21003] error Connection reset by peersend[fd=22003] error Connection reset by peer# 客户端connections: 26999, sockfd:27002, time_used:2399connections: 27999, sockfd:28002, time_used:2404connect: Cannot assign requested addresserror : Cannot assign requested address

我们看到大概创建了2.8w的fd ， 可是我们知道端口一个有6w多个，也就是说有6w个端口，为什么我们只使用了2.8w个？

Linux中有限定端口的使用范围:60999 - 32768 = 2.8w ,与我们上面实验结果相符。

The /proc/sys/net/ipv4/ip_local_port_range defines the local port range that is used by TCP and UDP traffic to choose the local port. You will see in the parameters of this file two numbers: The first number is the first local port allowed for TCP and UDP traffic on the server, the second is the last local port number. For high-usage systems you may change its default parameters to 32768-61000 -first-last.

proc/sys/net/ipv4/ip_local_port_range范围定义TCP和UDP通信用于选择本地端口的本地端口范围。您将在该文件的参数中看到两个数字：第一个数字是服务器上允许TCP和UDP通信的第一个本地端口，第二个是最后一个本地端口号。对于高使用率的系统，您可以将其默认参数更改为32768-61000(first-last)。

[root@master temp]# sysctl net.ipv4.ip_local_port_rangenet.ipv4.ip_local_port_range = 3276860999

解决问题

修改net.ipv4.ip_local_port_range的范围，一般不这样做，我们这里研究的是服务器，怎么会去对客户端进行修改呢

之前已经说了这个问题的背景，就是只开放了一个端口，并且socket fd --- < 源IP地址 , 源端口, 目的IP地址 , 目的端口 , 运输层协议 >，在这个背景下才产生的这个问题，所以我们可以开放更多的端口，比如说100个，那么一个客户端就能连到280w了

确定问题

我们将服务器端口开100个，按理说客户端可以连280w，但是现在只连接到13w就error : Connection timed out，与我们的预期不符

//服务端new connect [192.168.109.101:54585], pos[131165]new connect [192.168.109.101:48265], pos[131166]new connect [192.168.109.101:51997], pos[131167]new connect [192.168.109.101:43239], pos[131168]send[fd=20102] error Connection reset by peersend[fd=21102] error Connection reset by peersend[fd=22102] error Connection reset by peer//客户端connections: 127999, sockfd:128002, time_used:7576connections: 128999, sockfd:129002, time_used:2683connections: 129999, sockfd:130002, time_used:2669connections: 130999, sockfd:131002, time_used:4610connect: Connection timed outerror : Connection timed out

网卡接收的数据，会发送到协议栈里面，通过sk_buff将数据传到协议栈，协议栈处理完再交给应用程序。由于操作系统在使用的时候，为防止被攻击，在数据发送给协议栈之前进行一个过滤，在协议栈前面加了一个小组件：过滤器，叫做netfilter。

netfilter主要是对网络数据包进行一个过滤，在netfilter的基础上我们就可以实现防火墙，在linux里面有一个就叫做iptables，iptables是基于netfilter做的，iptables分为两部分，一部分是内核实现的netfilter接口，一部分是应用程序提供给用户使用的。iptables真正实现的是netfilter提供的接口。

Connection timed out译为连接超时，也就是说，client发送的请求超时了，那么这个超时有两种情况，第一种：三次握手第一次的SYN没发出去，第二种：三次握手第二次ACK没收到。

netfilter不管对发送的数据，还是对接收的数据，都是可以过滤的。当连接数量达到一定数量的时候，netfilter就会不允许再对外发连接了。所以现在推测是情况1造成的，发送的SYN被netfilter拦截了。

事实是这样吗，我们来查看一下netfilter允许对外最大连接数量是多少。13w，与我们上面建立成功的数量一致，所以现在就可以确定是netfilter允许对外开放的最大连接数造成的了

[root@node1 temp]# cat /proc/sys/net/netfilter/nf_conntrack_max131072

解决问题

我们可以通过设置netfilter允许对外最大连接数量，来解决这个问题

# 查看允许对外最大连接数量[root@node1 temp]# cat /proc/sys/net/netfilter/nf_conntrack_max131072# 进行配置vim /etc/sysctl.conf# 在配置文件中把net.nf_conntrack_max参数修改为1048576（如果配置就自己添加一行）net.nf_conntrack_max = 1048576# 重新加载配置文件sysctl -p# 再次查看，发现生效了[root@node1 temp]# cat /proc/sys/net/netfilter/nf_conntrack_max1048576

确定问题

这里我们先给客户端虚拟机2G的内存，然后发现到24w的时候，客户端进程被杀死了

connections: 239999, sockfd:240002, time_used:9837connections: 240999, sockfd:241002, time_used:10608connections: 241999, sockfd:242002, time_used:13109connections: 242999, sockfd:243002, time_used:15112connections: 243999, sockfd:244002, time_used:12606已杀死

我们来看一下kill记录，发现是内存不足。

[root@node1 ~]# dmesg | egrep -i -B100 'killed process'[ 2310.265218] Out of memory: Kill process 7266 (C1000Kclient) score 1 or sacrifice child[ 2310.265962] Killed process 7266 (C1000Kclient) total-vm:8708kB, anon-rss:2960kB, file-rss:0kB, shmem-rss:0kB

这里直接说原因吧，是因为程序每个fd都有一个tcp接收缓冲区和tcp发送缓冲区。而默认的太大了，导致Linux内存不足，进程被杀死，所有我们需要适当的缩小。进程空间，代码段，堆栈都是要占用内存的。

解决问题

我们只需要对net.ipv4.tcp_mem，net.ipv4.tcp_wmem，net.ipv4.tcp_rmem进行适合的修改即可

# 编辑内核参数配置文件vim /etc/sysctl.conf# 添加以下内容# 最小值 默认值 最大值net.ipv4.tcp_mem = 252144 524288 786432# tcp协议栈的大小，单位为内存页（4K），分别是 1G 2G 3G，如果大于2G，tcp协议栈会进行一定的优化net.ipv4.tcp_wmem = 1024 1024 2048 # tcp接收缓存区(用于tcp接受滑动窗口)的最小值，默认值和最大值（单位byte）1k 1k 2k,每一个连接fd都有一个接收缓存区net.ipv4.tcp_rmem = 1024 1024 2048 # tcp发送缓存区(用于tcp发送滑动窗口)的最小值，默认值和最大值（单位byte）1k 1k 2k,每一个连接fd都有一个发送缓存区# 总缓存 = （每个fd发送缓存区 + 每个fd接收缓存区） * fd数量# （1024byte + 1024byte ） * 100w 约等于 2G

如果服务器是用来接收大文件，传输量很大的时候，就要把send buffer和read buffer调大。

如果服务器只是接收小数据字符的时候。把buffer调小是为了把fd的数量做到更多，并发数量能做到更大。如果buffer调大的话，内存会不够。

想要实现服务器百万并发：

原文地址：Linux服务器百万并发实现与问题排查