> This phenomenon appears to be caused by some interaction between the
> delivery of I/O interrupts and timer interrupts when running under
> VM (or LPAR with shared CPUs, for that matter).  We haven't quite
> understood this whole issue yet, unfortunately.  (Maybe Martin knows
> more details ...)

Unfortunatly no. I havent't found a way to counter this problem yet.
Seems like the virtualization (VM/LPAR) likes to "cluster" the timer
interrupts with I/O interrupts. The order of the interrupts gets lost,
the I/O interrupt is always delivered first. The timer interrupt comes
in as soon as the interrupts get enabled again. This has a bad effect
because most I/O interrupts trigger a bottom half which enables the
interrupts. But ticks that arrive while the cpu is in a bottom half
are system ticks and not user/idle ticks. So due to the change in the
order of the interrupts the stochatic tick accounting gets wrong.

blue skies,
   Martin

Reply via email to