apache server reached MaxClients setting, consider raising the MaxClients setting - performance

Question

I am running centos 5.5 with 768mb ram. i keep getting server reached MaxClients setting, consider raising the MaxClients setting in the logs also apache runs really slow. when i look at cacti graphs it shows the server is not even using all the resources.. here is the current configuration
<IfModule prefork.c>
StartServers 8
MinSpareServers 5
MaxSpareServers 10
ServerLimit 1024
MaxClients 768
MaxRequestsPerChild 4000
</IfModule>
<IfModule worker.c>
StartServers 2
MaxClients 150
MinSpareThreads 25
MaxSpareThreads 75
ThreadsPerChild 25
MaxRequestsPerChild 0
</IfModule>
free -m
total used free shared buffers cached
Mem: 768 352 415 0 0 37
-/+ buffers/cache: 315 452
Swap: 0 0 0
top - 11:03:54 up 41 days, 11:53, 1 user, load average: 0.05, 0.03, 0.00
Tasks: 35 total, 1 running, 34 sleeping, 0 stopped, 0 zombie
Cpu(s): 0.0%us, 0.0%sy, 0.0%ni, 99.7%id, 0.0%wa, 0.0%hi, 0.0%si, 0.3%st
Mem: 786432k total, 389744k used, 396688k free, 0k buffers
Swap: 0k total, 0k used, 0k free, 38284k cached
I have tried the following but the server responds very slowly
<IfModule worker.c>
#StartServers 2
#MaxClients 150
#MinSpareThreads 25
#MaxSpareThreads 75
#ThreadsPerChild 25
#MaxRequestsPerChild 0
StartServers 20
MaxClients 1024
ServerLimit 1024
MinSpareThreads 128
MaxSpareThreads 768
ThreadsPerChild 64
MaxRequestsPerChild 0
</IfModule>
free -m
total used free shared buffers cached
Mem: 768 324 443 0 0 37
-/+ buffers/cache: 286 481
Swap: 0 0 0
#regilero
I have updated to
<IfModule prefork.c>
StartServers 12
MinSpareServers 12
MaxSpareServers 12
MaxClients 50
MaxRequestsPerChild 300
</IfModule>
using top i see
Tasks: 36 total, 1 running, 35 sleeping, 0 stopped, 0 zombie
Cpu(s): 0.0%us, 0.3%sy, 0.0%ni, 99.7%id, 0.0%wa, 0.0%hi, 0.0%si, 0.0%st
Mem: 786432k total, 613180k used, 173252k free, 0k buffers
Swap: 0k total, 0k used, 0k free, 76488k cached
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
1 root 20 0 10364 92 60 S 0.0 0.0 1:09.53 init
2 root 20 0 0 0 0 S 0.0 0.0 0:00.00 kthreadd/808
3 root 20 0 0 0 0 S 0.0 0.0 0:00.00 khelper/808
124 root 16 -4 12620 8 4 S 0.0 0.0 0:00.00 udevd
533 root 20 0 95504 5692 228 S 0.0 0.7 4:02.94 memcached
546 root 20 0 5924 332 276 S 0.0 0.0 6:54.51 syslogd
557 root 20 0 101m 1456 868 S 0.0 0.2 13:18.64 snmpd
570 root 20 0 62640 316 208 S 0.0 0.0 2:39.56 sshd
579 root 20 0 21656 24 20 S 0.0 0.0 0:00.00 xinetd
589 root 20 0 12072 12 8 S 0.0 0.0 0:00.05 mysqld_safe
940 mysql 20 0 559m 164m 3832 S 0.3 21.5 209:33.88 mysqld
1015 root 20 0 20880 200 132 S 0.0 0.0 0:10.48 crond
1023 root 20 0 46748 4 0 S 0.0 0.0 0:00.00 saslauthd
1024 root 20 0 46748 4 0 S 0.0 0.0 0:00.00 saslauthd
3605 root 20 0 62832 2168 636 S 0.0 0.3 0:02.58 sendmail
3613 smmsp 20 0 57712 1648 504 S 0.0 0.2 0:00.01 sendmail
17610 root 20 0 85932 3312 2600 S 0.0 0.4 0:00.02 sshd
17612 mcmap 20 0 86072 1760 1012 S 0.0 0.2 0:00.17 sshd
17613 mcmap 20 0 12076 1656 1292 S 0.0 0.2 0:00.01 bash
17637 root 20 0 45052 1432 1120 S 0.0 0.2 0:00.00 su
17638 root 20 0 12180 1800 1324 S 0.0 0.2 0:00.08 bash
17740 root 20 0 246m 9264 4516 S 0.0 1.2 0:00.19 httpd
18264 apache 20 0 282m 43m 4940 S 0.0 5.7 0:00.56 httpd
18514 apache 20 0 279m 40m 4832 S 0.0 5.3 0:01.47 httpd
18518 apache 20 0 273m 36m 4396 S 0.0 4.7 0:00.45 httpd
18528 apache 20 0 251m 13m 3660 S 0.0 1.8 0:00.41 httpd
18529 apache 20 0 278m 40m 4340 S 0.0 5.3 0:00.99 httpd
18530 apache 20 0 278m 40m 4268 S 0.0 5.3 0:00.67 httpd
18548 apache 20 0 272m 33m 3516 S 0.0 4.4 0:00.28 httpd
18552 apache 20 0 280m 42m 3684 S 0.0 5.5 0:00.48 httpd
18553 apache 20 0 271m 33m 3768 S 0.0 4.3 0:00.45 httpd
18555 apache 20 0 274m 36m 3672 S 0.0 4.7 0:00.58 httpd
18572 apache 20 0 247m 9020 2856 S 0.0 1.1 0:00.01 httpd
18578 apache 20 0 280m 42m 3684 S 0.0 5.6 0:00.76 httpd
18589 apache 20 0 246m 5452 676 S 0.0 0.7 0:00.00 httpd
18588 root 20 0 12624 1216 932 R 0.0 0.2 0:00.06
free -m
total used free shared buffers cached
Mem: 768 578 189 0 0 74
-/+ buffers/cache: 504 263
Swap: 0 0 0
Just added current picture of cacti result last 4 hours. busy periods are monday tuesday. So i will wait till next week to see further results of the config change. but it looks like an improvement as before i only had max 10 threads available. Looking at this do you think i can make more improvment?
free -m
total used free shared buffers cached
Mem: 768 619 148 0 0 49
-/+ buffers/cache: 570 197
Swap: 0 0 0
NEW TEST
On a 2GB Ram VPS box i have now set prefork to
StartServers 20
MinSpareServers 20
MaxSpareServers 20
ServerLimit 256
MaxClients 256
MaxRequestsPerChild 4000
today morning my memcache server died from
Nov 20 09:28:40 vps22899094 kernel: Out of memory: Kill process 12517 (memcached) score 81 or sacrifice child
Nov 20 09:28:40 vps22899094 kernel: Killed process 12517, UID 497, (memcached) total-vm:565252kB, anon-rss:42940kB, file-rss:44kB
What should the optimal values be to set in apache?
#/etc/sysconfig/memcached
PORT="11211"
USER="memcached"
MAXCONN="1024"
CACHESIZE="1024"
OPTIONS="-l 127.0.0.1"
/etc/my.cnf
[mysqld]
datadir=/var/lib/mysql
socket=/var/lib/mysql/mysql.sock
user=mysql
# Disabling symbolic-links is recommended to prevent assorted security risks
symbolic-links=0
bind-address=127.0.0.1
#script
thread_concurrency=2
query_cache_size = 16M
query_cache_type=1
query_cache_limit=5M
# MyISAM #
#key-buffer-size = 32M
#myisam-recover = FORCE,BACKUP
# SAFETY #
#max-allowed-packet = 16M
#max-connect-errors = 1000000
# CACHES AND LIMITS #
tmp-table-size = 32M
max-heap-table-size = 32M
#query-cache-type = 0
#query-cache-size = 0
max-connections = 50
thread-cache-size = 16
#open-files-limit = 65535
#table-definition-cache = 1024
#table-open-cache = 2048
# INNODB #
#innodb-flush-method = O_DIRECT
#innodb-log-files-in-group = 2
#innodb-log-file-size = 5M
#innodb-flush-log-at-trx-commit = 1
#innodb-file-per-table = 1
#innodb-buffer-pool-size = 921M
# LOGGING #
log-error = /var/log/mysqld.log
log-queries-not-using-indexes = 1
slow-query-log = 1
slow-query-log-file = /var/log/mysqld-slow.log
[mysqld_safe]
log-error=/var/log/mysqld.log
pid-file=/var/run/mysqld/mysqld.pid

Answer 1

When you use Apache with mod_php apache is enforced in prefork mode, and not worker. As, even if php5 is known to support multi-thread, it is also known that some php5 libraries are not behaving very well in multithreaded environments (so you would have a locale call on one thread altering locale on other php threads, for example).
So, if php is not running in cgi way like with php-fpm you have mod_php inside apache and apache in prefork mode.
On your tests you have simply commented the prefork settings and increased the worker settings, what you now have is default values for prefork settings and some altered values for the shared ones :
StartServers 20
MinSpareServers 5
MaxSpareServers 10
MaxClients 1024
MaxRequestsPerChild 0
This means you ask apache to start with 20 process, but you tell it that, if there is more than 10 process doing nothing it should reduce this number of children, to stay between 5 and 10 process available. The increase/decrease speed of apache is 1 per minute. So soon you will fall back to the classical situation where you have a fairly low number of free available apache processes (average 2). The average is low because usually you have something like 5 available process, but as soon as the traffic grows they're all used, so there's no process available as apache is very slow in creating new forks. This is certainly increased by the fact your PHP requests seems to be quite long, they do not finish early and the apache forks are not released soon enough to treat another request.
See on the last graphic the small amount of green before the red peak? If you could graph this on a 1 minute basis instead of 5 minutes you would see that this green amount was not big enough to take the incoming traffic without any error message.
Now you set 1024 MaxClients. I guess the cacti graph are not taken after this configuration modification, because with such modification, when no more process are available, apache would continue to fork new children, with a limit of 1024 busy children. Take something like 20MB of RAM per child (or maybe you have a big memory_limit in PHP and allows something like 64MB or 256MB and theses PHP requests are really using more RAM), maybe a DB server... your server is now slowing down because you have only 768MB of RAM. Maybe when apache is trying to initiate the first 20 children you already reach the available RAM limit.
So. a classical way of handling that is to check the amount of memory used by an apache fork (make some top commands while it is running), then find how many parallel request you can handle with this amount of RAM (that mean parallel apache children in prefork mode). Let's say it's 12, for example. Put this number in apache mpm settings this way:
<IfModule prefork.c>
StartServers 12
MinSpareServers 12
MaxSpareServers 12
MaxClients 12
MaxRequestsPerChild 300
</IfModule>
That means you do not move the number of fork while traffic increase or decrease, because you always want to use all the RAM and be ready for traffic peaks. The 300 means you recyclate each fork after 300 requests, it's better than 0, it means you will not have potential memory leaks issues. MaxClients is set to 12 25 or 50 which is more than 12 to handle the ListenBacklog queue, which can enqueue some requests, you may take a bigger queue, but you would get some timeouts maybe (removed this strange sentende, I can't remember why I said that, if more than 12 requests are incoming the next one will be pushed in the Backlog queue, but you should set MaxClient to your targeted number of processes).
And yes, that means you cannot handle more than 12 parallel requests.
If you want to handle more requests:
buy some more RAM
try to use apache in worker mode, but remove mod_php and use php as a parallel daemon with his own pooler settings (this is called php-fpm), connect it with fastcgi. Note that you will certainly need to buy some RAM to allow a big number of parallel php-fpm process, but maybe less than with mod_php
Reduce the time spent in your php process. From your cacti graphs you have to potential problems: a real traffic peak around 11:25-11:30 or some php code getting very slow. Fast requests will reduce the number of parallel requests.
If your problem is really traffic peaks, solutions could be available with caches, like a proxy-cache server. If the problem is a random slowness in PHP then... it's an application problem, do you do some HTTP query to another site from PHP, for example?
And finally, as stated by #Jan Vlcinsky you could try nginx, where php will only be available as php-fpm. If you cannot buy RAM and must handle a big traffic that's definitively desserve a test.
Update: About internal dummy connections (if it's your problem, but maybe not).
Check this link and this previous answer. This is 'normal', but if you do not have a simple virtualhost theses requests are maybe hitting your main heavy application, generating slow http queries and preventing regular users to acces your apache processes. They are generated on graceful reload or children managment.
If you do not have a simple basic "It works" default Virtualhost prevent theses requests on your application by some rewrites:
RewriteCond %{HTTP_USER_AGENT} ^.*internal\ dummy\ connection.*$ [NC]
RewriteRule .* - [F,L]
Update:
Having only one Virtualhost does not protect you from internal dummy connections, it is worst, you are sure now that theses connections are made on your unique Virtualhost. So you should really avoid side effects on your application by using the rewrite rules.
Reading your cacti graphics, it seems your apache is not in prefork mode bug in worker mode. Run httpd -l or apache2 -l on debian, and check if you have worker.c or prefork.c. If you are in worker mode you may encounter some PHP problems in your application, but you should check the worker settings, here is an example:
<IfModule worker.c>
StartServers 3
MaxClients 500
MinSpareThreads 75
MaxSpareThreads 250
ThreadsPerChild 25
MaxRequestsPerChild 300
</IfModule>
You start 3 processes, each containing 25 threads (so 3*25=75 parallel requests available by default), you allow 75 threads doing nothing, as soon as one thread is used a new process is forked, adding 25 more threads. And when you have more than 250 threads doing nothing (10 processes) some process are killed. You must adjust theses settings with your memory. Here you allow 500 parallel process (that's 20 process of 25 threads). Your usage is maybe more:
<IfModule worker.c>
StartServers 2
MaxClients 250
MinSpareThreads 50
MaxSpareThreads 150
ThreadsPerChild 25
MaxRequestsPerChild 300
</IfModule>

Answer 2

Did you consider using nginx (or other event based web server) instead of apache?
nginx shall allow higher number of connections and consume much less resources (as it is event based and does not create separate process per connection). Anyway, you will need some processes, doing real work (like WSGI servers or so) and if they stay on the same server as the front end web server, you only shift the performance problem to a bit different place.
Latest apache version shall allow similar solution (configure it in event based manner), but this is not my area of expertise.

Answer 3

Here's an approach that could resolve your problem, and if not would help with troubleshooting.
Create a second Apache virtual server identical to the current one
Send all "normal" user traffic to the original virtual server
Send special or long-running traffic to the new virtual server
Special or long-running traffic could be report-generation, maintenance ops or anything else you don't expect to complete in <<1 second. This can happen serving APIs, not just web pages.
If your resource utilization is low but you still exceed MaxClients, the most likely answer is you have new connections arriving faster than they can be serviced. Putting any slow operations on a second virtual server will help prove if this is the case. Use the Apache access logs to quantify the effect.

Answer 4

I recommend to use bellow formula suggested on Apache:
MaxClients = (total RAM - RAM for OS - RAM for external programs) / (RAM per httpd process)
Find my script here which is running on Rhel 6.7. you can made change according to your OS.
#!/bin/bash
echo "HostName=`hostname`"
#Formula
#MaxClients . (RAM - size_all_other_processes)/(size_apache_process)
total_httpd_processes_size=`ps -ylC httpd --sort:rss | awk '{ sum += $9 } END { print sum }'`
#echo "total_httpd_processes_size=$total_httpd_processes_size"
total_http_processes_count=`ps -ylC httpd --sort:rss | wc -l`
echo "total_http_processes_count=$total_http_processes_count"
AVG_httpd_process_size=$(expr $total_httpd_processes_size / $total_http_processes_count)
echo "AVG_httpd_process_size=$AVG_httpd_process_size"
total_httpd_process_size_MB=$(expr $AVG_httpd_process_size / 1024)
echo "total_httpd_process_size_MB=$total_httpd_process_size_MB"
total_pttpd_used_size=$(expr $total_httpd_processes_size / 1024)
echo "total_pttpd_used_size=$total_pttpd_used_size"
total_RAM_size=`free -m |grep Mem |awk '{print $2}'`
echo "total_RAM_size=$total_RAM_size"
total_used_size=`free -m |grep Mem |awk '{print $3}'`
echo "total_used_size=$total_used_size"
size_all_other_processes=$(expr $total_used_size - $total_pttpd_used_size)
echo "size_all_other_processes=$size_all_other_processes"
remaining_memory=$(($total_RAM_size - $size_all_other_processes))
echo "remaining_memory=$remaining_memory"
MaxClients=$((($total_RAM_size - $size_all_other_processes) / $total_httpd_process_size_MB))
echo "MaxClients=$MaxClients"
exit