Apache 1.3 + PHP 5 on FreeBSD 6 -- bad performance under load

[BSD]

Nope:

Mem: 921M Active, 699M Inact, 290M Wired, 96K Cache, 199M Buf, 2041M Free
Swap: 4096M Total, 4096M Free


Mark J. Nernberg wrote:

>Is it swapping?
>
>
>On 2/15/06 11:15 AM, "BSD" <bsdlists at celeritystorm.com> wrote:
>
>  
>
>>Hello list,
>>
>>I've been having trouble identifying what's causing my apache server to
>>take about 5 seconds to reply to new http requests.
>>
>>The machine in question is a Dual Xeon 2.8GHz with 4GB RAM, running
>>apache 1.3.33_4 and PHP5 on 6.0-RELEASE-p4. It is a very busy PHP-based
>>site.
>>
>>I have done a lot of tweaking on this server, but when the load goes a
>>bit higher, it'll take about 5 seconds to answer to a standard GET /
>>HTTP/1.1 via telnet (no PHP involved, since I'm not sending valid HTTP
>>headers and it will reply with a 400 bad request).
>>
>>It is a problem with apache, as telnetting to the sshd port and hitting
>>ENTER yields an instant "Protocol mismatch". -- no delay whatsoever.
>>
>>Both CPUs are around 65% idle after some PHP tweaks, so it's not CPU
>>starvation either.
>>
>>I have accept filters enabled on the kernel, and apache is using them; also:
>>
>>options         KVA_PAGES=768
>>options         PMAP_SHPGPERPROC=5120
>>
>>- to avoid that "colleting PV entries - suggest increasing
>>PMAP_SHPGPERPROC".
>>
>>So far I've been unable to pinpoint what's causing apache a ~5 sec delay
>>to process the GET. A timestamped tcpdump shows the following:
>>
>>Establishing the connection:
>>02/15/2006 16:06:44.232973 192.168.0.1.53578 > APACHE_SERVER.80: S
>>3958791824:3958791824(0) win 65535 <mss 1460,nop,wscale
>>1,nop,nop,timestamp 7596496 0> (DF) [tos 0x10]
>>02/15/2006 16:06:44.387141 APACHE_SERVER.80 > 192.168.0.1.53578: S
>>2261546731:2261546731(0) ack 3958791825 win 65535 <mss 1460,nop,wscale
>>2,nop,nop,timestamp 1124627170 7596496> (DF)
>>02/15/2006 16:06:44.387348 192.168.0.1.53578 > APACHE_SERVER.80: . ack 1
>>win 33304 <nop,nop,timestamp 7596512 1124627170> (DF) [tos 0x10]
>>
>>02/15/2006 16:06:44.536819 APACHE_SERVER.80 > 192.168.0.1.53578: . ack 1
>>win 32942 <nop,nop,timestamp 1124627318 7596512> (DF)
>>
>>
>>GET: (note: delay from previous step was induced by manually switching
>>terminals and typing)
>>02/15/2006 16:06:49.148143 192.168.0.1.53578 > APACHE_SERVER.80: P
>>1:17(16) ack 1 win 33304 <nop,nop,timestamp 7596988 1124627318> (DF)
>>[tos 0x10]
>>02/15/2006 16:06:49.395079 APACHE_SERVER.80 > 192.168.0.1.53578: . ack
>>17 win 32938 <nop,nop,timestamp 1124632177 7596988> (DF)
>>02/15/2006 16:06:49.395284 192.168.0.1.53578 > APACHE_SERVER.80: P
>>17:19(2) ack 1 win 33304 <nop,nop,timestamp 7597013 1124632177> (DF)
>>[tos 0x10]
>>02/15/2006 16:06:49.642127 APACHE_SERVER.80 > 192.168.0.1.53578: . ack
>>19 win 32937 <nop,nop,timestamp 1124632423 7597013> (DF)
>>
>>
>>"400 bad request" reply:
>>02/15/2006 16:06:59.769277 APACHE_SERVER.80 > 192.168.0.1.53578: P
>>1:481(480) ack 19 win 32942 <nop,nop,timestamp 1124642548 7597013> (DF)
>>02/15/2006 16:06:59.769453 APACHE_SERVER.80 > 192.168.0.1.53578: F
>>481:481(0) ack 19 win 32942 <nop,nop,timestamp 1124642548 7597013> (DF)
>>02/15/2006 16:06:59.769543 192.168.0.1.53578 > APACHE_SERVER.80: . ack
>>482 win 33064 <nop,nop,timestamp 7598050 1124642548> (DF) [tos 0x10]
>>02/15/2006 16:06:59.770548 192.168.0.1.53578 > APACHE_SERVER.80: F
>>19:19(0) ack 482 win 33304 <nop,nop,timestamp 7598050 1124642548> (DF)
>>[tos 0x10]
>>02/15/2006 16:06:59.923720 APACHE_SERVER.80 > 192.168.0.1.53578: . ack
>>20 win 32941 <nop,nop,timestamp 1124642703 7598050> (DF)
>>
>>
>>As you can see, it took 10 seconds to reply to the GET, yet the
>>accepting of the connection was instant.
>>I have more than enough maxfiles, maxproc, maxfilesperproc, nmbclusters
>>etc. I guess the connection is so promptly answered because of the http
>>accept filters built in the kernel.
>>
>>As a final note, the MaxClients I'm using are not nearly enough (around
>>200 connections/second, according to pfstat graphics vs MaxClients 98).
>>The reason for this is that trying to match MaxClients with the real
>>load resulted in two 0.0% idle processors, since the PHP code is a bit
>>heavy. I've opted to disable KeepAlive and lower MaxClients instead.
>>Could this be the delay I'm seeing ? I think so, but I need more
>>opinions, so feel free to reply.
>>
>>
>>_______________________________________________
>>freebsd-apache at freebsd.org mailing list
>>http://lists.freebsd.org/mailman/listinfo/freebsd-apache
>>To unsubscribe, send any mail to "freebsd-apache-unsubscribe at freebsd.org"
>>
>>
>>    
>>
>
>  
>