On 28/04/2012 9:38 a.m., Bruce Lysik wrote:
Hi guys,
Running latest 3.1 in a reverse proxy mode. 3 beefy servers with 96GB of ram. Seeing an odd problem:
Origin -> customer, equals fast speeds. (Tested by curling from a desktop to origin.)
Origin -> squid, equals fast speeds. (Tested by running curl on the squid server to the origin.)
Squid cache hit -> customer, equals fast speed. (Seen in browser.)
Squid cache miss -> customer, insanely slow. 36kB/sec, when origin to customer direct is like 50MB/sec.
Any ideas on what to look at here? It's so broken it feels like a misconfiguration somewhere.
These are on RHEL6u2, 96GB ram, 1.69TB RAID5 ext4 partition for disk cache, 4gb of bonded network interface. Machines are behind a load balancer operating in DSR mode.
The usual stuff is:
* disk I/O loading. Squid still cycles most objects through the disks
when caching and RAID does horrible things to the write cycle speed.
* forwarding loops. If the traffic is looping in an dout and back again
Squid impact can be huge.
* delay pools not being bypassed for the reverse-proxy traffic.
* QoS on the underlying system slowing things down.
* ECM or PMTU brokenness preventing the Squid box making fast
jumbo-packet connections.
Amos
