Allright so I'm about to upgrade my box (with 2k external ips) to the latest squid (3.3.x) since my old version do not take any advantage of the multi-core processors.
While researching on the workers I found a discussion where someone says:
Use cpu_affinity_map in squid.conf.
Leave Cpu0 "for the OS". Be careful not to put two busy workers on
sibling hyper-cores. This is just a sketch of an optimization
algorithm. There are many details that depend on your setup.
I also understand that it's best to use as many workers as you have CPUs:
FWIW, we usually see best performance results when using
cpu_affinity_map with 1:1 mapping between workers and cores (which
effectively disables those complex algorithms as far as Squid workers
are concerned).
My question is…it's that physical CPUs? My box reports 8 cpus but only 2 physical shown when I run cat /proc/cpuinfo | grep "physical id" | sort | uniq | wc -l so that means 2 workers?
If I put 2 workers on 2 physical CPUs how does Leave Cpu0 "for the OS" fits? This means I set only 1 worker for cpu1 while OS uses cpu0?
P.S. I know about cpu afinity. I just need a clarification about how to find out exactly how many workers I can use before I lose poerformance.
Best Answer
You have two Xeon E5506 CPUs with 8 cores total. The E5506 has no hyper-threading. You could run seven workers and still have an free core for the OS.