Mailing List Archive: sluggish servers

sluggish servers

Jun 25, 1995, 8:43 PM

Post #1 of 3 (307 views)

I just traced a problem with the server at Cardiff to NFS.

I guess lots of people offload cron jobs onto other local machines to
keep the load on the web server low, well, if those jobs need to write
to the disk on the web server, then this can be more trouble that it's
worth.

The movie db on the Cardiff machine was keeping a big cache of CGI output,
and another local machine was flushing it every 6 hours. The simple job
of deleting files over NFS can suck the life out of a web server.. Ours
was responding immediately, but pausing 5-10 seconds to send output...
Why ?, a file delete needs to update a dorectory file which hold the
directory info... that file can get pretty big, so if you are
removing lots of files in rapid succession, the NFS overhead hits hard.

Even after the cache flush script had finished, the delays stayed for
at least 15 minutes (at which point I good tired of waiting and rebooted)..
The server responds and sends output immediately once more.

...just a lesson for you all, and something for the www documentation I
think.

rob
--
http://nqcd.lanl.gov/~hartill/

Re: sluggish servers [ In reply to ]

hartill at ooo

Jun 25, 1995, 8:47 PM

Post #2 of 3 (298 views)

Permalink

> at least 15 minutes (at which point I good tired of waiting and rebooted)..
^^^^
it's been a long day :-)

Re: sluggish servers [ In reply to ]

rst at ai

Jun 26, 1995, 9:50 AM

Post #3 of 3 (298 views)

Permalink

From: Rob Hartill <hartill@ooo.lanl.gov>
Date: Sun, 25 Jun 95 20:43:52 MDT

I just traced a problem with the server at Cardiff to NFS.

...Even after the cache flush script had finished, the delays stayed for
at least 15 minutes (at which point I good tired of waiting and rebooted)..
The server responds and sends output immediately once more.

Actually, this is a point that deserves emphasis and *re*-emphasis ---
Web server performance depends on the underlying file system at least
as much as anything else (barring only the bandwidth of the net
connection itself). Any recently released web server should be good
up to at least a few 100K hits/day on reasonable hardware, if the
underlying net connection and file system are up to it. However, the
file system is something which a whole lot of people don't think much
about, and some of them are probably tripping up on it.

I know, for instance, that when my server is getting to DocumentRoot
through the SunOS automounter, vs. through a hard mount point, it
becomes dramatically slower (at least 3x). AFS is probably at least
as bad, given that the way Web servers use the file system (a whole
lot of lookups, many for nonexistent files) is a usage pattern which
the AFS gurus regard as "low-incidence", and therefore don't bother to
cater to. (It's really amazing how many people wind up getting told
that their programs *crawl* on AFS because they are using a
"low-frequency usage pattern" which isn't worth the trouble to support
adequately).

rst