最近给服务增加了一个cache_put_latency指标，加了之后，吓了一跳。发现往memcached put一个10KB左右的数据，latency居然有7ms左右，难于理解，于是花了一些精力找原因。我分别写了一个shell和C++的测试程序。

1、shell脚本使用nc发送set命令。

#/bin/env bash

let s=1

let i=0

let len=8*1024

while true

do

if (( i >= $len ))

then

break

fi

str=${str}1

let i++

done

let i=0

begin_time=`date +%s`

while true

do

if (( i >= 1000 ))

then

break

fi

printf "set $i 0 0 $len\r\n${str}\r\n" | nc 10.234.4.24 11211

if [[ $? -eq 0 ]]

then

echo "echo key: $i"

fi

let i++

done

end_time=`date +%s`

let use_time=end_time-begin_time

echo "set time consumed: $use_time"

let i=0

begin_time=`date +%s`

while true

do

if (( i >= 1000 ))

then

break

fi

printf "get $i\r\n" | nc 10.234.4.22 11211 > /dev/null 2>&1

let i++

done

end_time=`date +%s`

let use_time=end_time-begin_time

echo "get time consumed: $use_time"

2、C++程序则通过libmemcached set。

#include <iostream>

#include <map>

#include <string>

#include <sys/time.h>

#include <time.h>

#include <stdlib.h>

#include "libmemcached/memcached.h"

using namespace std;

uint32_t item_size = 0;

uint32_t loop_num = 0;

bool single_server = false;

std::string local_ip;

std::map<std::string, uint32_t> servers;

int64_t getCurrentTime()

{

struct timeval tval;

gettimeofday(&tval, NULL);

return (tval.tv_sec * 1000000LL + tval.tv_usec);

}

memcached_st* mc_init()

{

memcached_st * mc = memcached_create(NULL);

if (mc == NULL)

{

cout << "create mc error" << endl;

return NULL;

}

std::map<std::string, uint32_t>::iterator iter;

for (iter = servers.begin(); iter != servers.end(); ++iter)

{

if (single_server && iter->first != local_ip)

{

continue;

}

memcached_return rc = memcached_server_add(mc, iter->first.c_str(), iter->second);

if(rc != MEMCACHED_SUCCESS)

{

cout << "add server " << iter->first << " error" << endl;

return NULL;

}else

{

cout << "add server " << iter->first << " success" << endl;

}

}

return mc;

}

void test_put(memcached_st * mc)

{

char* ptr = new char[item_size];

memset(ptr, 'a', item_size);

char buf[32];

memset(buf, 0, 32);

struct iovec curkey, curval;

curval.iov_base = ptr;

curval.iov_len = item_size;

curkey.iov_base = buf;

curkey.iov_len = 32;

uint64_t begin_time = getCurrentTime();

for (uint32_t i=0; i < loop_num; ++i)

{

sprintf(buf, "%d", i);

memcached_return rc = memcached_set(mc,(const char *)curkey.iov_base, curkey.iov_len,

(const char *)curval.iov_base, curval.iov_len, 600, (uint32_t)0);

if (rc != MEMCACHED_SUCCESS)

{

cout << "set key error: " << buf << endl;

continue;

}else

{

cout << "set key: " << buf << endl;

}

}

uint64_t end_time = getCurrentTime();

cout << "put time comsumed: " <<  end_time - begin_time << endl;

}

void test_get(memcached_st * mc)

{

}

int main(int argc, const char* argv[])

{

//

if (strcmp(argv[1], "s") == 0)

{

single_server = true;

}else

{

single_server = false;

}

item_size = atoi(argv[2])*1024;

loop_num = atoi(argv[3]);

//

servers["10.232.42.91"] = 11211;

/*servers["10.234.4.22"] = 11211;

servers["10.234.4.23"] = 11211;

servers["10.234.4.24"] = 11211;*/

//

local_ip = "10.232.42.91";

//

memcached_st* mc = mc_init();

if (! mc)

{

cout << "mc_init error" << endl;

return -1;

}

test_put(mc);

test_get(mc);

}

测试发现二者的结果是相背的。shell脚本set 1000次8KB的item，只要3s左右，平均需要3ms。而C++版本则需要39s左右，平均耗时39ms。照理说shell脚本需要不断连接服务器和启动nc进程，应该更慢才对。我用ltrace跟踪了一下，发现8KB的数据需要发送两次，两次write都是非常快的，但是等memcached返回时用了很多时间，主要的时间就耗费在这个地方。

23:32:37.069922 [0x401609]

memcached_set(0x19076200, 0x7fffdad68560, 32, 0x1907a570, 8192 <unfinished ...>

23:32:37.070034 [0x3f280c5f80]

SYS_write(3, "set 29 0 600

8192\r\naaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"..., 8196) = 8196 <0.000022>

23:32:37.071657 [0x3f280c5f80]

SYS_write(3, "aaaaaaaaaaaaaaa\r\n", 17) = 17 <0.000012>

23:32:37.071741 [0x3f280c5f00]

SYS_read(3, "STORED\r\n", 8196) = 8 <0.039765>(39ms)

和剑豪讨论下之后，剑豪马上去grep了一把代码，发现原来libmemcached居然有MEMCACHED_MAX_BUFFER这样一个常量，其值为8196。并且它还没有对应的memcached_behavior_set函数。在memcached_constants.h中将其直接改成81960，然后就欣喜地发现cache_put_latency从7ms降低到1ms左右。

问题完美虽然地解决了，但是有点意犹未尽，于是想搞明白为什么会出现这种奇怪的现象。瓶颈貌似在服务器端，于是对memcached做了一些修改。在状态切换的时候加上一个精确到微秒的时间。

static int64_t getCurrentTime()

{

struct timeval tval;

gettimeofday(&tval, NULL);

return (tval.tv_sec * 1000000LL + tval.tv_usec);

}

static void conn_set_state(conn *c, enum conn_states state) {

assert(c != NULL);

assert(state >= conn_listening && state < conn_max_state);

if (state != c->state) {

if (settings.verbose > 2) {

fprintf(stderr, "%d: going from %s to %s, time: %lu\n",

c->sfd, state_text(c->state),

state_text(state), getCurrentTime());

}    

c->state = state;

if (state == conn_write || state == conn_mwrite) {

MEMCACHED_PROCESS_COMMAND_END(c->sfd, c->wbuf, c->wbytes);

}

}

}

从打印的时间戳可以看出来，时间主要花在conn_nread状态处理代码中。最后定位到第二次read花费的时间非常多。

15: going from conn_waiting to conn_read, time: 1348466584440118

15: going from conn_read to conn_parse_cmd, time: 1348466584440155

<15 set 98 0 600 8192

15: going from conn_parse_cmd to conn_nread, time: 1348466584440177

conn_nread: 17

> NOT FOUND 98

>15 STORED

15: going from conn_nread to conn_write, time: 1348466584480099(36ms)

15: going from conn_write to conn_new_cmd, time: 1348466584480145

15: going from conn_new_cmd to conn_waiting, time: 1348466584480152

value的数据可能在conn_read中读完了，这个时候只需要memmove一下就好了。如果没有在conn_read状态中读完，那么就需要conn_nread自己来一次read了(因为套接字被设置成了异步，所以还可能需要多次read)，关键就是这个read太慢了。

case conn_nread:

if (c->rlbytes == 0) {

complete_nread(c);

break;

}

/* first check if we have leftovers in the conn_read buffer */

if (c->rbytes > 0) {

int tocopy = c->rbytes > c->rlbytes ? c->rlbytes : c->rbytes;

if (c->ritem != c->rcurr) {

memmove(c->ritem, c->rcurr, tocopy);

}

c->ritem += tocopy;

c->rlbytes -= tocopy;

c->rcurr += tocopy;

c->rbytes -= tocopy;

if (c->rlbytes == 0) {

break;

}

}    

/*  now try reading from the socket */

res = read(c->sfd, c->ritem, c->rlbytes);

if (res > 0) {

pthread_mutex_lock(&c->thread->stats.mutex);

c->thread->stats.bytes_read += res;

pthread_mutex_unlock(&c->thread->stats.mutex);

if (c->rcurr == c->ritem) {

c->rcurr += res;

}

c->ritem += res;

c->rlbytes -= res;

break;

}

折腾了好久，在libmemcached的io_flush函数前后也打了不少时间戳，发现libmemcached发送数据是非常快的。突然灵感闪现，我想起来了TCP_NODELAY这个参数，于是在libmemcached memcached_connect.c文件中的set_socket_options函数中增加了这个参数(事实上set_socket_options函数里面可以设置TCP_NODELAY，没有仔细看)。

int flag = 1;

int error = setsockopt(ptr->fd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(flag) );

if (error == -1) {

 printf("Couldn't setsockopt(TCP_NODELAY)\n");

exit(-1);

}else

{

 printf("set setsockopt(TCP_NODELAY)\n");

}

在不改MEMCACHED_MAX_BUFFER的情况下，现在set 100KB的item也是一瞬间的事情了。不过新的困惑又出现了，Nagle算法什么情况会起作用呢？为什么第一个包没被缓存，第二个包一定会被缓存呢？

libmemcached发送一个set命令是分成三部分的，首先是header(set 0 0 600 8192\r\n，共18个字节)，然后是value(8192个字节)，最后是’\r\n’(两个字节)，一共是8212个字节。memcached在conn_read状态一共能读取2048+2048+4096+8196=16KB的数据，因此对于8KB的数据是完全可以在conn_read状态读完的。通过在conn_read状态处理的代码中增加下面的打印语句可以发现有些情况下，conn_read最后一次只读取了4个字节(正常情况应该是2048+2048+4096+20)，剩下的16个字节放到conn_nread中读了。

res = read(c->sfd, c->rbuf + c->rbytes, avail);

if (res > 0) {

char buf[10240] = {0};

sprintf(buf, "%.*s", res, c->rbuf + c->rbytes);

printf("avail=%d, read=%d, str=%s\n", avail, res, buf);

未设置TCP_NODELAY选项时，使用netstat可以看到客户端socket的Send-Q一直会维持在8214和8215之间。

tcp        0   8215 10.232.42.91:59836          10.232.42.91:11211          ESTABLISHED 25800/t

设置TCP_NODELAY选项时，客户端socket的Send-Q就一直为0了。

tcp        0      0 10.232.42.91:59890          10.232.42.91:11211          ESTABLISHE