php-fpm占用了太多内存

Question

我使用的是Nginx+ php和php 5.3.6，它工作了几个星期。几天前，我发现大多数PHP子程序的开销为3MB，但也有几个子程序占用了700 3MB。web应用程序忙于1000个请求/秒。PHP与APC一起使用(操作码使用)

有人能指点这个吗？

This script I used to get memory usage by child pid 
 http://www.pixelbeat.org/scripts/ps_mem.py.

It could be verified by 'top' method too

top output(limited view)::
------------------------------
------------------------------
28419 daemon    20   0  844m 757m 4200 S    0  6.4   0:14.27 php-fpm                                                          

16788 daemon    20   0  700m 614m 4632 S    0  5.2   0:28.34 php-fpm                                                          

29450 daemon    20   0  669m 581m 3548 S    0  4.9   0:08.31 php-fpm                                                          

17881 daemon    20   0  642m 556m 4108 S    0  4.7   0:14.83 php-fpm                                                          

19048 daemon    20   0  642m 555m 4108 S    0  4.7   0:08.86 php-fpm 
11956 daemon    20   0 97612  10m 5476 S    4  0.1   0:39.57 php-fpm                                                          
11993 daemon    20   0 97560  10m 5188 S    4  0.1   0:39.18 php-fpm                                                          
11925 daemon    20   0 97328  10m 5144 D    3  0.1   0:38.68 php-fpm                                                          
11953 daemon    20   0 97748  10m 5172 S    4  0.1   0:38.51 php-fpm   

php-fpm.conf：

/etc/php-fpm/php-fpm.conf
listen = 127.0.0.1:9000
user = daemon
group = daemon

pm = dynamic
pm.max_children = 2000
pm.start_servers = 50
pm.min_spare_servers = 40
pm.max_spare_servers = 90
pm.max_requests = 10000

以下是更多的调试输入：

pmap:：

pmap 28419
0000000000b52000     96K rw---    [ anon ]
0000000001a49000   1668K rw---    [ anon ]
0000000001bea000    208K rw---    [ anon ]
0000000001c1e000 770476K rw---    [ anon ]

斯特拉斯：

strace -p 28419
Process 28419 attached - interrupt to quit
restart_syscall(<... resuming interrupted call ...>) = 0
recvfrom(4, 0x1bda1d0, 8196, 64, 0, 0)  = -1 EAGAIN (Resource temporarily unavailable)
poll([{fd=4, events=POLLIN}], 1, 5000)  = 0 (Timeout)
recvfrom(4, 0x1bda1d0, 8196, 64, 0, 0)  = -1 EAGAIN (Resource temporarily unavailable)
poll([{fd=4, events=POLLIN}], 1, 5000^C <unfinished ...>

加载的PHP模块如下：

[PHP Modules]
apc
Core
date
ereg
hash
json
libxml
mbstring
mcrypt
memcache
memcached
mysql
mysqli
pcntl
pcre
Reflection
session
SimpleXML
SPL
standard
xml

Answer 1

尝试将最大请求降到类似于pm.max_requests = 100的地方。这将在100个请求之后重新启动进程，释放可能由于脚本中内存泄漏而丢失的内存。

Answer 2

您需要将pm = dynamic更改为pm = ondemand。它将把拉姆的使用量减少一半。