We launched mod_cluster in Production this week and CPU on our webservers has doubled as a result, compared with mod_proxy_balancer. There appears to be a direct relationship between the amount of CPU overhead and 1) the number of active nodes and 2) the Maxnode / Maxcontext setting regardless of how many nodes are actually active. For example, I was able to lower CPU simply by reducing Maxnode from 100 to 60 and Maxcontext from 500 to 300 even though the same number of nodes (60) were active before and after. It was having a hard time keeping up with traffic until we added a couple more web servers and reduced the number of nodes to 40. Previously we had plenty of headroom.
Is there something less-than-optimal about running mod_cluster on Solaris SPARC? Is it expected that a higher-than-needed Maxnode setting would increase CPU utilization? Or does this point to a possible problem that may be causing the high CPU?
Here's the environment:
Sun SPARC T1 x5 - 1.2Ghz, each: 8 cores, 32 hardware threads
Apache/2.2.21 32-bit
mod_cluster/1.2.0
40 nodes
200 contexts
16 virtualhosts
500 req./s per Apache instance
I've attached the Apache config file...only thing I left out is 15 additional vhosts.
