One performance-specific 2.7 question - I recall hearing mention of an
issue with 2.6 where larger objects held in mem_cache required
exponentially more CPU cycles to serve up (i.e. n cycles for a 4KB
object, n*2 for an 8KB object, n*4 for a 12KB object, etc). Does
anyone know if this issue is still present in 2.7 code?
Thanks,
-C
On Apr 1, 2009, at 12:42 AM, Amos Jeffries wrote:
Hello people, I'm having some bottlenecks on our squid deployement
and I
was
wondering if anyone had any recommendations because I'm near out of
ideas.
Would someone change my architecture or does anyone have any
experience
with
a squid deployment of this size? Basically we are pushing 600mb at
240,000k
connections. When we reach speeds on our around that that number,
we start
seeing slow performance and getting alot of page timeouts. We are
running
20
squid 2.6 boxes running dsr behind a single foundry gte load
balancer. I
recently had 18 squid boxes and thought we had a squid bottleneck
but no
change. I was kinda leaning in the direction of splitting it in
half and
doing multiple load balancers. Does anyone have any experience
pushing
this
much traffic?
I'd go to 2.7 if it was squid being slow. It has a number of small
performance boosters missing in 2.6.
But, it does sound like a load balancer bottleneck if you had zero
change
from adding two more Squid. The Yahoo and Wikimedia guys are the
poster
installs for these types of Squid deployment and they use CARP
meshing.
PS: We are very interested in getting some real-world benchmarks from
high-load systems like yours. Can you grab the data needed for
http://wiki.squid-cache.org/KnowledgeBase/Benchmarks ?
Amos
