Re: interrupt throttling stepping in too soon?

From: John Baldwin <jhb_at_freebsd.org>
Date: Thu, 13 Oct 2005 15:50:30 -0400
On Friday 07 October 2005 03:45 am, Stijn Hoop wrote:
> Followup to my own problem:
>
> On Thu, Oct 06, 2005 at 09:53:12AM +0200, Stijn Hoop wrote:
> > On the console there were multiple DMA_TIMEOUT messages for the disks of
> > the array, and just above those was a line about an 'interrupt storm for
> > atapci0, throttling'.
>
> After a rebuild which completed succesfully, I rsynced data to the disks.
> About 5 minutes later the same message appeared and the machine panic'd,
> this time destroying /var/log with it :-/
>
> Anyway, I decided to hammer the disks without waiting a long time for
> gvinum to build an array, and indeed I can reproduce this by mounting all 4
> drives, and for each drive dd'ing the whole drive to /dev/null plus
> rsync'ing data to it at the same time. It held up at 20 MB/s but I suspect
> that the box isn't keeping up with the combined total of interrupts because
> it goes down faster when I engage the network (rsync from a remote box).
>
> So, I guess my question becomes: is there a way to find out why this box
> can't handle enough interrupts? Could it be that the motherboard is flawed?
> Is this definitely a hardware error or is it still possible that the
> interrupt throttling is stepping in too soon and thus screwing the rest
> of the system?
>
> On a sidenote, I monitored systat -vmstat while doing the above and it
> appeared that IRQ 11 (ATA controller) was doing around 1000-1100
> irqs/s, and IRQ 5 (xl0) was doing around 800-1000 irqs/s. Is there a
> way to log this somehow?

Yes, when throttling is on, we limit the interrupts to about hz per second.  
You can try turning throttling off (by setting the limit to 0) or increasing 
the threshold via the sysctl hw.intr_storm_threshold.

-- 
John Baldwin <jhb_at_FreeBSD.org>  <><  http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve"  =  http://www.FreeBSD.org
Received on Thu Oct 13 2005 - 18:14:53 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:45 UTC