I am looking to understand some server performance problems I am seeing with a (for us) heavily loaded web server. The environment is as follows:
- Debian Lenny (all stable packages + patched to security updates)
- Apache 2.2.9
- PHP 5.2.6
- Amazon EC2 large instance
The behavior we're seeing is that the web typically feels responsive, but with a slight delay to begin handling a request -- sometimes a fraction of a second, sometimes 2-3 seconds in our peak usage times. The actual load on the server is being reported as very high -- often 10.xx or 20.xx as reported by top. Further, running other things on the server during these times (even vi) is very slow, so the load is definitely up there. Oddly enough Apache remains very responsive, other than that initial delay.
We have Apache configured as follows, using prefork:
StartServers 5
MinSpareServers 5
MaxSpareServers 10
MaxClients 150
MaxRequestsPerChild 0
And KeepAlive as:
KeepAlive On
MaxKeepAliveRequests 100
KeepAliveTimeout 5
Looking at the server-status page, even at these times of heavy load we are rarely hitting the client cap, usually serving between 80-100 requests and many of those in the keepalive state. That tells me to rule out the initial request slowness as "waiting for a handler" but I may be wrong.
Amazon's CloudWatch monitoring tells me that even when our OS is reporting a load of > 15, our instance CPU utilization is between 75-80%.
Example output from top:
top - 15:47:06 up 31 days, 1:38, 8 users, load average: 11.46, 7.10, 6.56
Tasks: 221 total, 28 running, 193 sleeping, 0 stopped, 0 zombie
Cpu(s): 66.9%us, 22.1%sy, 0.0%ni, 2.6%id, 3.1%wa, 0.0%hi, 0.7%si, 4.5%st
Mem: 7871900k total, 7850624k used, 21276k free, 68728k buffers
Swap: 0k total, 0k used, 0k free, 3750664k cached
The majority of the processes look like:
24720 www-data 15 0 202m 26m 4412 S 9 0.3 0:02.97 apache2
24530 www-data 15 0 212m 35m 4544 S 7 0.5 0:03.05 apache2
24846 www-data 15 0 209m 33m 4420 S 7 0.4 0:01.03 apache2
24083 www-data 15 0 211m 35m 4484 S 7 0.5 0:07.14 apache2
24615 www-data 15 0 212m 35m 4404 S 7 0.5 0:02.89 apache2
Example output from vmstat at the same time as the above:
procs -----------memory---------- ---swap-- -----io---- -system-- ----cpu----
r b swpd free buff cache si so bi bo in cs us sy id wa
8 0 0 215084 68908 3774864 0 0 154 228 5 7 32 12 42 9
6 21 0 198948 68936 3775740 0 0 676 2363 4022 1047 56 16 9 15
23 0 0 169460 68936 3776356 0 0 432 1372 3762 835 76 21 0 0
23 1 0 140412 68936 3776648 0 0 280 0 3157 827 70 25 0 0
20 1 0 115892 68936 3776792 0 0 188 8 2802 532 68 24 0 0
6 1 0 133368 68936 3777780 0 0 752 71 3501 878 67 29 0 1
0 1 0 146656 68944 3778064 0 0 308 2052 3312 850 38 17 19 24
2 0 0 202104 68952 3778140 0 0 28 90 2617 700 44 13 33 5
9 0 0 188960 68956 3778200 0 0 8 0 2226 475 59 17 6 2
3 0 0 166364 68956 3778252 0 0 0 21 2288 386 65 19 1 0
And finally, output from Apache's server-status:
Server uptime: 31 days 2 hours 18 minutes 31 seconds
Total accesses: 60102946 - Total Traffic: 974.5 GB
CPU Usage: u209.62 s75.19 cu0 cs0 - .0106% CPU load
22.4 requests/sec - 380.3 kB/second - 17.0 kB/request
107 requests currently being processed, 6 idle workers
C.KKKW..KWWKKWKW.KKKCKK..KKK.KKKK.KK._WK.K.K.KKKKK.K.R.KK..C.C.K
K.C.K..WK_K..KKW_CK.WK..W.KKKWKCKCKW.W_KKKKK.KKWKKKW._KKK.CKK...
KK_KWKKKWKCKCWKK.KKKCK..........................................
................................................................
From my limited experience I draw the following conclusions/questions:
We may be allowing far too many
KeepAliverequestsI do see some time spent waiting for IO in the vmstat although not consistently and not a lot (I think?) so I am not sure this is a big concern or not, I am less experienced with vmstat
Also in vmstat, I see in some iterations a number of processes waiting to be served, which is what I am attributing the initial page load delay on our web server to, possibly erroneously
We serve a mixture of static content (75% or higher) and script content, and the script content is often fairly processor intensive, so finding the right balance between the two is important; long term we want to move statics elsewhere to optimize both servers but our software is not ready for that today
I am happy to provide additional information if anybody has any ideas, the other note is that this is a high-availability production installation so I am wary of making tweak after tweak, and is why I haven't played with things like the KeepAlive value myself yet.
I'll start by admitting that I don't much about running stuff in clouds - but based on my experience elsewhere, I'd say that this webserver config reflects a fairly low volume of traffic. That the runqueue is so large suggests that there just isn't enough CPU available to deal with it. What else is in the runqueue?
No - keeplive still improves performance, modern browsers are very smart about knowing when to pipeline and when to run requests in parallel, although a timeout of 5 seconds is still rather high, and you've got a LOT of servers waiting - unless you've got HUGE latency problems I'd recommend cranking this down to 2-3. This should shorten the runqueue a little.
If you've not already got mod_deflate installed on the webserver - then I'd recommend you do so - and add the ob_gzhandler() to your PHP scripts. You can do this as an auto-prepend:
(yes, copression uses more CPU - but you should save CPU overall by getting servers out of the runqueue faster / handling fewer TCP packets - and as a bonus, your site is also faster).
I'd recommend setting an upper limit on MaxRequestsPerChild - say something like 500. This just allows some turnover on processes in case you've got a memory leak somewhere. Your httpd processes look to be HUGE - make sure you've removed any apache modules you don't need and make sure you're serving up static content with good caching information.
If you're still seeing problems, then the problem is probably within the PHP code (if you switch to using fastCGI, this should be evident without any major performance penalty).
update
If the static content doesn't vary much across pages, then it might also be worth experimenting with:
on the PHP scripts too.
You should consider installing an asynchronous reverse proxy, because a number of processes in W state is quite high too. Your Apache processes seem to spend a lot of time sending content to slow clients over network being blocked on that. Nginx or lighttpd as a frontend to your Apache server can reduce a number of processes in W state dramatically. And yes, you should limit a number of keepalive requests. Probably it is worth trying to turn keepalive off.
BTW, 107 Apache processes are too high for 22 rps, I was able to serve 100-120 rps using only 5 Apache processes. Probably, the next step is to profile your application.
You have two rows in your vmstat that show your CPU wait time is fairly high, and around those, you do a fair number of writes (io - bo) and context switching. I would look at what's writing blocks, and how to eliminate that wait. I think the most improvement could be found in improving your disk IO. Check syslog - set it to write async. Make sure your controller's write cache is working (check it -- you might have a bad battery).
Keepalive isn't causing your perf problem, it saves you time on connection setup if you're not running a cache in front. You might bump MaxSpareServers a bit so that in a crunch you're not waiting for all the forks.
you should consider turning keepalive off as a first try...
with 107 request processed I would keep MaxSpareServers higher then what you set...
IMHO in the long-term nginx as reverse proxy for static content should be take into consideration
First suggestion: disable keepalives. I've only ever needed it when I could identify a specific situation that the performance increased, but in general requests/sec decreased with Keepalive enabled.
Second suggestion: Set a MaxRequestsPerChild. I echo symcbean here, it will help with process rollover in the case of a memory leak. 500 is a good starting point.
Third Suggestion: Increase MaxClients. A ballpark calculation for this is (physical memory - memory used by non-httpd process)/size of each httpd process. Depending on how httpd was compiled, this number maxes at 255. I use 250 for my public servers to deal with google/yahoo/MS crawling the systems.
Forth Suggestion: Increase MaxSpareServers: something like 4-5x MinSpareServers.
Baring those suggestions failing, I would look at load-balancing with reverse-proxy or memcache for DB.