[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: kjournald tuning

On Oct 03, 2002  16:34 -0700, Jay Thorne wrote:
> ...what I'm writing to you about is kjournald.
> When this apparently looping state is reached, kjournald shows D 
> state for several seconds. During that same time, block-out "bo" in vmstat 
> consistently shows 0. kjournald is choking. Or something. Also, this appears 
> to get a lot of processes looping on write somewhere in the webserver. 
> Looping on the lockfile, maybe? The runqueue gets huge > 40 often , and 
> apache spirals up to all 75 processes in use. These pauses seem to occur at 
> pretty exact intervals. Something like every 240 seconds. With the disk in 
> question mounted as ext2 I don't get the issue. At all.
> I'd like to find a way to tell kjournald to wait less time with smaller 
> buffers so that the big pause while kjournald flushes and subsequent long 
> runqueue and its associated context switch overhead does not occur. 

In Linux's tradition of pre-emptive bug fixing, the parameter you seek
was just added to the official 2.4.20-pre9 kernel announced yesterday.
It allows you to specify the flush interval on a per-filesystem basis.
Since sct submitted it, I assume it is also in the RH kernel, but
probably only in a RawHide kernel and not an official errata kernel yet.

Summary of changes from v2.4.20-pre8 to v2.4.20-pre9
<sct redhat com>:
  o 2.4.20-pre4/ext3: Bump ext3 version number
  o 2.4.20-pre4/ext3: Fix LVM snapshot deadlock
  o 2.4.20-pre4/ext3: jbd commit interval tuning
  o Sanity check for Intermezzo/ext3
  o ext3 commit notification for Intermezzo
  o Fix the order of inodes being marked dirty in a couple of corner cases

It appears that this patch is actually _not_ in the RH 12.5 kernel, but
maybe there is a more recent RawHide kernel?  If not, the patch is at:


Cheers, Andreas
Andreas Dilger

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]