Squid 2.1 does not seem to scale too well on the CPU side. Let's compare
our old box, a 512MB 166 MHz Ultrasparc, with our new box, a 1GB 366 MHz
Ultrasparc with much larger L2 cache.
The old box ran Solaris 2.51, the new one runs Solaris 7. The old one
ran Disksuite logging, the new one runs with the logging and noatime
options of Solaris 7.
The new machine is 5 to 10 times as fast analyzing the logs, but Squid
performance has been raised by just a factor of 2.
Strangely enough, Squid still is CPU bound. Disks are rather quiet,
around 8% busy with service times of 7 milliseconds. The previous box
had much busier (and slower) disks.
We are running async IO. The new box goes up to around 7.5
megabits/second. The old box did up to 4 mb/s.
What surprises me is the numer of syscalls per second as shown by
mpstat. A busy cache seems to have a slightly lower number of syscalls
than a quiet cache. Measured over 15 minutes, at a time when the cache
is fairly quiet:
Requests/sec ~Mb/sec Syscalls/sec CPU idle
15.8 1.3 5490 49 %
29.5 1.7 3955 39 %
Both machines have exactly the same configuration.
Does this point to an inefficiency in Squid on Solaris?
Thomas
Received on Tue Mar 30 1999 - 01:59:22 MST
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:45:35 MST