httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From r..@ai.mit.edu (Robert S. Thau)
Subject Re: Real World Webserver Performance
Date Fri, 16 Feb 1996 20:47:22 GMT
  Our webserver is current I/O bound, and just about anything you can think 
  of affects I/O.  We found out today most of the I/O had nothing to do 
  with user-level disk accesses, but instead with file last-accessed time 
  updating.  Or so we think.  :)  Every 30 seconds, when the disk cache is 
  flushed, we see a huge spike in I/O, which constant running of "sync" 
  eliminates, so we'll see tomorrow at 11am PST (our heaviest time and the 
  only time we have a problem) if turning that off helps things.  

[Still catching up...]

FWIW, I've seen the exact same pattern [regular large spikes of
activity, on a web filesystem which is otherwise not very busy, all
writes] on my server; it would appear, at first blush, that most reads
(i.e., most of the activity that the server itself would be waiting on
--- that and directory lookups) are serviced out of the buffer cache.
Or am I missing something?

At any rate, it's probably much more common for web server processes
to be network-bound than disk-bound (both I/O, but with very different
properties, the former being quite a bit harder to control).

rst

[... or am I missing something?]

Mime
View raw message