Re: Long latencies with MD RAID 1 [was Re: libata/sata_nv latency on NVIDIA CK804 ]

[Lee Revell <rlrevell@joe-job.com>]

On Wed, 2006-03-15 at 19:21 -0500, Bill Rugolsky Jr. wrote:
> On Wed, Mar 15, 2006 at 06:44:02PM -0500, Lee Revell wrote:
> > On Wed, 2006-03-15 at 18:14 -0500, Bill Rugolsky Jr. wrote:
> > > [Meanwhile, I still have to switch contexts and look at the long
> > > softirq latencies that at first glance appear to be due to the use of
> > > mempool by the RAID1 bio code.] 
> > 
> > Can you post traces of them somewhere?  There are no long running
> > softirqs in the two you posted (the worst is only 200 usecs or so).
> 
> This is typical of what I'm seeing. It seems to be looping over lots
> of io request completions?
> 
> 	-Bill
> 
> preemption latency trace v1.1.5 on 2.6.16-rc6-git4-latency
> --------------------------------------------------------------------
>  latency: 1950 us, #8586/8586, CPU#0 | (M:desktop VP:0, KP:0, SP:0 HP:0 #P:1)

This looks very similar to what I see with my regular ATA drive (except
that the completions are handled in hardirq context).

You can cause less work to be done in each softirq by
lowering /sys/block/$DEV/queue/max_sectors_kb.

I would not consider ~2ms "long", there are some other softirqs that
induce 10-15ms latencies...

Lee