[squid-users] squid 3.1 with https traffic and delay pools is flooding network with hundreds of thousands 65-70 bytes packets (and killing the routers, anyway)

[Amos Jeffries]

On 17/06/2015 10:11 p.m., Horváth Szabolcs wrote:
> Hello!
> 
> We're having serious problems with a squid proxy server. 
> 
> The good news is the problem can be reproduced at any time in our production squid system.
> 
> Environment:
> - CentOS release 6.5 (Final) with Linux kernel 2.6.32-431.29.2.el6.x86_64
> - squid-3.1.10-22.el6_5.x86_64 (a bit old, CentOS ships this version)
> 
> Problem description:
> - if we have a few mbytes/sec https traffic AND
> - delay_classes are in place AND
> - delay pools are full (I mean the available bandwidth for the customer are used)
> 
> -> then squid is trickling https traffic down to the clients in 65-70 byte packets.
> 
> Our WAN routers are not designed to handle thousands of 65-70 byte packets per seconds and therefore we have some network stability issues.
> 
> I tracked down the following:
> - if delay_pools are commented out (clients can go with full speed as they like) -> the problem eliminates, https traffic flows with ~1500 byte packets
> - if we use only http traffic, there is no problem: http traffic flows with ~1500 byte packets even if the delay pools are full
> 
> Our test URL is www.opengroup.org/infosrv/DCE/dce122.tar.gz, which is available both on http and https protocol.
> 
> Resources can be found at http://support.iqsys.hu/logs/
> 
> 1. squid.conf -> squid configuration file
> 2. http-delaypool.pcap: 
> 	- wget -c http://www.opengroup.org/infosrv/DCE/dce122.tar.gz, 
> 	- delay pools are active
> 	- http flows with 1500 byte packets
> 3. http-nodelaypool.pcap: 
> 	- wget -c http://www.opengroup.org/infosrv/DCE/dce122.tar.gz, 
> 	- delay pools are INACTIVE
> 	- http flows with 1500 byte packets
> 4. https-delaypool.pcap:
> 	- wget -c https://www.opengroup.org/infosrv/DCE/dce122.tar.gz, 
> 	- delay pools are active
> 	- http flows with 69 byte packets -> this is extremely bad
> 5. https-nodelaypool.pcap:
> 	- wget -c https://www.opengroup.org/infosrv/DCE/dce122.tar.gz, 
> 	- delay pools are INACTIVE
> 	- http flows with 1500 byte packets
> 
> My question is: is it a known bug?

Sounds like http://bugs.squid-cache.org/show_bug.cgi?id=2907,
 which was fixed in Squid-3.5.3.

see comment #16 in the bug report for a 3.1 workaround patch. Though if
your production server has high performance requirements the sleep(1)
workaround is not the best.

Amos