So long as you only use it as an indication. As you said, this is really looking at things the wrong way around.
It is easy to construct situations where several tasks are "waiting" for I/O, but only one is contributing almost all the load. That can only really be checked by something like blktrace which is at the VFS layer.
Something else you might be interested in is iotop - very nice python script. Seems to be in most distros these days.
Last edited by syg00; 08-19-2010 at 08:34 PM.