Ok, I'm sending the summary now even though we haven't quite solved the
problem. First, thanks to:
James Mularadelis <james.mularadelis@bms.com>
Brian Scanlan <singer@redbrick.dcu.ie>
Paul Pescitelli <pbp@mindspring.com>
Joseph Chiu <josephc@systems.caltech.edu>
Second, a few websites for performance tuning info:
http://www.sunworld.com/sunworldonline/swol-03-1996/swol-03-perf.html
http://www.sunworld.com/sunworldonline/common/cockcroft.letters.html
http://www.apache.org/docs/misc/perf-tuning.html
http://www.rvs.uni-hannover.de/people/voeckler/tune/EN/tune.html
Third:
1) It was suggested to check the interfaces to make sure they are the same
speed and half/full duplex as the switch -- they are.
2) It was suggested to make sure we've got enough swap. The CPU, memory
and disk I/O are all nowhere near being throttled.
3) It was suggested that we check file caching:
> In /etc/system, ncsize and ufs_ninode are set to (maxusers *17) + 90.
> maxusers, if not set manually in /etc/system is set the the amount of mb
> free in ram, in your case, 1024, iirc. You can set ncsize (which controls the
> directory name lookup cache) as high as 34906. You can also set ufs_ninode as
> high as 34906. There is also a limit set on how many buffers can be allocated
> in one go, namely bufhwm, and is set to 2% of system memory in kilobytes. You
> may try setting this higher.
4) Of course, there's always "Sun Performance Tuning" by Cockroft.
Anyway, we haven't solved the problem, but it may not even be one.
Basically, we were on the phone with Sun support for a while today, and
they see nothing wrong with our systems or the tuning. Keynote's web
performance service is reporting this 4-8 second lag on the last few
images, and we're not seeing that lag on a different server with a
different ISP (same OS and web software, different hardware (Ultra 1)).
We're also only seeing the lag when more than three images are on the
page. We're working with the ISP to see if there's a routing issue or
some other problem there. We don't have any local way of getting a web
metric, only watching the page load, and locally it seems fast.
So, unfortunately, we've gotten a lot better at tuning servers, but we're
no closer to knowing if this is a real problem or if perhaps Keynote is
just goofing up somewhere.
Thanks to all the repliers.
-Adam
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:13:33 CDT