(转)测试Lighttpd accept的惊群现象

转载自:http://www.javaeye.com/topic/382107

lighttpd里面采用的是prefork的模型,在fork进程之前就已经创建好了listen socket
那么fork了进程池之后,所有进程都有一份自己独立的listen socket fd,

但实际上这个独立的fd 对应的确是一个文件表项,即实际上任然是一个共享的文件描述符

在阻塞模型中,各进程分别通过accept阻塞,等待连接到达,当一个连接到达时,所有的进程都会被唤醒,但只有其中一个进程可以成功accept该连接,其余的则继续投入睡眠,这就是所谓的惊群现象

lighttpd使用的是非阻塞IO复用模型,测试一下是否会有惊群现象呢?

先把结论给出:

1.比如有20个进程注册了listen socket的请求连接事件,当一个连接到达确实会有多个进程 被通知有事件要处理(但不是全部,大约只有5,6个进程)
2.被唤醒的这几个进程会调用accept函数,其中只有一个成功返回连接fd,其余进程均返回EAGAIN或者 EWOULDBLOCK错误(因为是非阻塞的)

测试方法,自己写了一个prefork进程 + epoll的非阻塞server,启动20个进程,client telnet,打印服务器日志

try to accept new connection,pid=29879
try to accept new connection,pid=29876
try to accept new connection,pid=29880

process 29879 accept connection

accept EAGAIN error pid=29876
try to accept new connection,pid=29875
accept EAGAIN error pid=29880
accept EAGAIN error pid=29875

四个进程被通知有事件处理,1个成功accept,3个返回EAGAIN

在lighttpd中,server当被通知有连接要处理时,server会通过循环执行
accept,直到返回错误,或者超过一个上限值

这样,当海量请求连接到达时,似乎惊群不会带来太多的性能损耗。

[cpp]
enum conn_states {
conn_listening, /** the socket which listens for connections */
conn_read, /** reading in a command line */
conn_write, /** writing out a simple response */
conn_nread, /** reading in a fixed number of bytes */
conn_swallow, /** swallowing unnecessary bytes w/o storing */
conn_closing, /** closing this connection */
conn_mwrite, /** writing out many items sequentially */
};
typedef struct{
int fd;
int state;
}conn;
[/cpp]
[cpp]
#include<stdio.h>
#include<unistd.h>
#include<sys/types.h>
#include<stdlib.h>
#include<sys/socket.h>
#include<errno.h>
#include<fcntl.h>
#include<netinet/in.h>
#include<sys/resource.h>
#include "event.h"
#include "base.h"

//forward declaration

static fdevents *ev;
static conn **conns;
static int freetotal;
static int freecurr;

static int create_listen_fd(char *addr,int port);
static int conn_init();
static conn *get_conn_from_freelist();
static int add_conn_to_freelist(conn *c);
conn *conn_new(int fd,int state);

static int conn_init(){
freetotal=200;
freecurr=0;
conns=(conn **)malloc(freetotal * sizeof(*conns));
if(!conns){
return -1;
}
return 0;
}
static conn *get_conn_from_freelist(){
conn *con;
if(freecurr > 0){
con=conns[–freecurr];
conns[freecurr]=NULL;
return con;
}
return NULL;
}
static int add_conn_to_freelist(conn *c){
if(freecurr<freetotal){
conns[freecurr++]=c;
return 0;
}else{
conn **new_conns=(conn **)realloc(conns,sizeof(*new_conns)*2*freetotal);
if(new_conns){
freetotal*=2;
conns=new_conns;
conns[freecurr++]=c;
return 0;
}
}
return -1;
}
conn *conn_new(int fd,int state){
conn *c;
c=get_conn_from_freelist();
if(!c){
c=(conn *)malloc(sizeof(*c));
}

c->fd=fd;
c->state=state;

return c;

}
static int create_listen_fd(char *addr,int port){
int fd,val,flags;
struct sockaddr_in sockaddr;
fd=socket(AF_INET,SOCK_STREAM,0);
if(fd==-1){
fprintf(stderr,"socket()\n");
return -1;
}
val=1;
if(setsockopt(fd,SOL_SOCKET,SO_REUSEADDR,&val,sizeof(val))<0){
fprintf(stderr,"reuseaddr\n");
return -1;
}
if((flags=fcntl(fd,F_GETFL,0)<0) || fcntl(fd,F_SETFL,flags | O_NONBLOCK) < 0){
fprintf(stderr,"nonblocking\n");
return -1;
}

bzero(&sockaddr,sizeof(sockaddr));
sockaddr.sin_family=AF_INET;
sockaddr.sin_port=htons(port);
inet_pton(AF_INET,addr,&sockaddr.sin_addr);

if(bind(fd,(struct sockaddr *)&sockaddr,sizeof(sockaddr))<0){
fprintf(stderr,"bind error %s",strerror(errno));
return -1;
}
if(listen(fd,2048)<0){
fprintf(stderr,"listen %s",strerror(errno));
return -1;
}
return fd;
}
void event_handler(int fd,void *ctx, int revents){
struct sockaddr_in addr;
socklen_t sock_len;
int done=0,connfd;
conn *c;
c=(conn *)ctx;
while(!done){
switch(c->state){
case conn_listening:
printf("try to accept new connection,pid=%d\n",getpid());
sock_len=sizeof(addr);
connfd=accept(fd,(struct sockaddr *)&addr,&sock_len);
if(connfd>0){
printf("process %d accept connection\n",getpid());
c = conn_new(connfd,conn_read);
fdevent_register(ev,connfd,event_handler,c);
fdevent_event_add(ev,connfd,FDEVENT_IN);
}else{
if(errno== EAGAIN || errno == EWOULDBLOCK){
printf("accept EAGAIN error pid=%d\n",getpid());
}
if(errno==EINTR){
printf("accept EINTR error pid=%d\n",getpid());
}
if(errno==ECONNABORTED){ /* this is a FreeBSD thingy */
printf("accept EABORTED error pid=%d\n",getpid());
}
if(errno==EMFILE){
printf("accept EMFILE error pid=%d\n",getpid());
}
}
done=1;
break;
case conn_read:
printf("on read");
break;
}
}
}
int main(int argc,char **argv){
int fd,o;
char *listen_addr;
int port,num_childs,max_fds;
struct rlimit rlim;
conn *c;
port=0;
num_childs=5;
while(-1!=(o=getopt(argc,argv,"l:p:f:h"))){
switch(o){
case ‘l’:
listen_addr=strdup(optarg);
break;
case ‘p’:
port=atoi(optarg);
break;
case ‘f’:
num_childs=atoi(optarg);
break;
case ‘h’:
printf("Usage -l listen addr\n");
printf("Usage -p listen port \n");
printf("Usage -f fork num\n");
exit(1);
}
}
if(!listen_addr){
listen_addr=strdup("127.0.0.1");
}
if(!port){
printf("port is unknown\n");
exit(1);
}
if(0 != getrlimit(RLIMIT_NOFILE,&rlim)){
fprintf(stderr,"getrlimit failed.reason %s\n",strerror(errno));
exit(1);
}

max_fds=rlim.rlim_cur;

//create listen socket
if(-1==(fd=create_listen_fd(listen_addr,port))){
fprintf(stderr,"create listen fd failed\n");
exit(1);
}
//prefork child
if(num_childs > 0){
int child=0;
while(!child){
if(num_childs >0){
switch(fork()){
case -1:
return -1;
case 0:
child=1;
break;
default:
num_childs–;
break;
}
}else{
int status;
if(-1 !=wait(&status)){
num_childs++;
}else{
//ignore
}
}
}
}
//child process event
conn_init();
c=conn_new(fd,conn_listening);
ev=fdevent_init(max_fds);
if(!ev){
fprintf(stderr,"fdevent_init()\n");
exit(1);
}
fdevent_register(ev,fd,event_handler,c);
fdevent_event_add(ev,fd,FDEVENT_IN);
fdevent_poll(ev,1000);
}
[/cpp]

Leave a Reply

Your email address will not be published.