Hi Thorsten et al
Well I finally found the _cause_ of the slow I/O problem. I rewrote the
deadline component for staircase 3 times and then I even wrote a completely
new scheduler for my own experimentation! Then I threw all those away.
Anyway this patch (only applies to 2.6.16-ck2) does a number of cleanups for
staircase. Specifically there are two major improvements. The first is that
the sleep type in disk reads was not being accounted for at all meaning disk
reads would have unfairly low cpu usage under load. The problem with fixing
this previously was that it would then allow other sneaky tasks to use kernel
activity on their behalf to get unfair cpu usage. The second major change is
accounting properly for kernel activity on behalf of a task. This fixes a
number of "exploit" cases that have been suggested for some time, and
improves the fairness of tests like the starve.c that Mike Galbraith
published. I have to thank Mike for bringing these to my attention that I've
tried hard to ignore for some time.
Thorsten could you please test to see if this fixes the problem for you? I
_know_ Mike isn't going to test it for me (sorry again...)
These changes should all end up in staircase v15 all going well.
Cheers,
Con
As an attachment since most of my mailing list users prefer that.