Re: 2.6.21-mm1: many processes end up in D state

On Mon, 30 Apr 2007 20:14:05 +0200
"Jiri Slaby" <jirislaby@xxxxxxxxx> wrote:

I have a problem with higher disk loads (e.g. running git-log or yum update).
Many processes end up in D state and system is unusable -- I'm not able to run
anything but smooth mouse moving when this happens.

If I wait for a 20-30sec it becomes usable. This happens in 2.6.21-rc7-mm2 and
also in 2007-04-28-05-06 broken-out snapshot. I think 2.6.21-rc6-mm1 worked
fine, but I'm uncertain. If it is important, let me know to re-test.


It is important, but I doubt if retesting 2.6.21-rc6-mm1 will clarify
things a lot.

Could you try switching to a different IO scheduler please? Anticipatory
would suit.

As I wrote below the sysrq-t, switch to noop didn't help, but it seems
that it's harder to reproduce with that:

<cite it's_bad_to_write_anything_below_logs="true">
Note that yum works on lvm on raid0 and git too, but on the another md volume.
Both ext3s. Drivers are sata_promise and ata_piix (sata disk); CFQ scheduler.
Using noop is no change (but seems to be harder to reproduce with it). I figured
out that it probably happens when 2+ processes are on both "processors" (HT on
P4) and are IO wait (multiload-applet shows red above the half).

Swap usage is 0 all the time.
</cite>

My comprehension skills on Monday morning are even less than usual ;)

I would check the anticipatory scheduler as well, please. I don't know
what no-op would do with a workload like that, but it probably isn't very
good.

You appear to believe that it's related to the CPU scheduler? That's a bit
unexpected - it sounds more like a VFS/IO thing? But stranger things have
happened.

I guess it's time to end the staircase experiment in -mm.
http://userweb.kernel.org/~akpm/js.bz2 is my current rollup (against
2.6.21) minus staircase and related things. Pretty please.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Relevant Pages

  • Re: [PATCH] Consolidate the request merging
    ... moves the backmerging out of the io schedulers and into the elevator ... We save 1.6kb of text and as a bonus get backmerging for noop as ... the deadline scheduler. ... IMO it is a good idea for the noop scheduler to do as much merging as ...
    (Linux-Kernel)
  • Re: [PATCH] Consolidate the request merging
    ... moves the backmerging out of the io schedulers and into the elevator ... We save 1.6kb of text and as a bonus get backmerging for noop as ... the deadline scheduler. ... IMO it is a good idea for the noop scheduler to do as much merging as ...
    (Linux-Kernel)
  • Re: io scheduler silly question perhaps..
    ... Try benchmarking Anticipatory or Deadline against Noop, ... flash, but perhaps we need a simple "merge only" scheduler for this ...
    (Linux-Kernel)
  • Re: Fall back io scheduler for 2.6.15?
    ... >> I actually thought that elevator= got removed, ... Not an accident at all, actually, because the original patch i ... Since noop is guaranteed to be ... default scheduler is not loaded. ...
    (Linux-Kernel)
  • Re: wrong in_flight diskstat in 2.6.16.1
    ... I see weird output from /sys/block/sd/stat on our AMD64-X2 smp ... scheduler was used, change to noop didn't change anything. ... It is vanilla 2.6.16.1 and here are absolute values in hex and one second differences below: ...
    (Linux-Kernel)