Thanks for continuing to help me trouble-shoot here. Wasn't able to upgrade
the kernel today as we had issues with failover. Will try again next
saturday morning.
You asked:
What's the approximative session rate and data rate ?
We're peaking at about 30Mbps. There are two types of data going through
haproxy:
- HTTP - 90% of requests have maybe 30 bytes of data outside of the HTTP
overhead. These 90% requests are AJAX polling calls that are *held open on
the server for 20 seconds* before returning a response that effectively says
"no messages waiting for you".
- TCP - there's a lot of data going through here at high rates. I'd
estimate that this makes up 20Mbps of the 30. Yesterday for about 60
minutes we didn't have the TCP traffic going through here. We were
bypassing the load balancer during that period. The CPU usage of haproxy
remained at the same 100% level during that period. These TCP sessions are
persistent, btw. There are about 100 of them and once the sockets are
established they stick around for days at a clip.
You asked about the size of the machine: Dual Core Pentium E2160, 1.8 GHz, 1MB Cache, 800MHz Front Side Bus 2GB RAM Glad to hear the configuration sounds ok and that health-checks are not an issue at all.
Here's the stats output from strace:
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ---------------- 93.05 3.276951 1146 2859 poll 1.56 0.054937 12 4532 8 send 1.29 0.045360 5 8370 40 recv 1.04 0.036664 9 3941 close 0.68 0.023890 7 3235 1314 accept 0.67 0.023423 12 1938 1926 connect 0.47 0.016556 9 1926 socket 0.37 0.013019 7 1877 shutdown 0.36 0.012678 3 3847 setsockopt 0.30 0.010513 3 3847 fcntl64 0.21 0.007548 3 2860 gettimeofday 0.00 0.000103 4 28 getsockopt ------ ----------- ----------- --------- --------- ---------------- 100.00 3.521642 39260 3288 totalReceived on 2008/02/16 19:20
This archive was generated by hypermail 2.2.0 : 2008/02/16 19:30 CET