From: pSeries AIX Geek (aixgeek@YAHOO.COM)
Date: Mon Nov 11 2002 - 22:02:00 EST
I think the rationale is this:
- Typical performance would be seriously degraded by
requiring a MWC write on every jfslog write (consider
that the MWC resides on the outer edge of the disk).
- A normal file system write operation that would
change metadata consists of (very high-level) a write
to the jfslog, the write to the file system, and a
commit back to the jfslog. If the jfslog is mirrored,
then the "write" operation to the jfslog (either the
initial or the commit) won't complete until the all
copies are written. Now, suppose that for some reason
this didn't complete (system crash before the commit
is mirror-written to the other copy). Then, the MWC
would say, "ah, I need to flush the primary copy to
the secondary" (I believe jfslog devices have
scheduling set to sequential), so your commit is thus
written.
Remember, MWC ONLY helps in cases where a crash occurs
after mirror has been written but BEFORE the PPs are
marked sync'd.
I can't see a case where the jfslog really needs it.
In a worst case scenario -- the commit is never
applied and then an fsck would read the jfslog and
correct whatever discrepancies are there.
- pAG
--- Holger.VanKoll@SWISSCOM.COM wrote:
> Hello,
>
> I noticed that mwc is (by default) turned off for
> paging-spaces and jfslogs. I understand this
> concerning paging-space, but why shouldn´t I care
> about consistent mirrors of jfslogs ?
>
> This thread
>
http://marc.theaimsgroup.com/?t=98198083800007&r=1&w=2
> also ends with this question...
>
> Regards,
>
> Holger
>
>
__________________________________________________
Do you Yahoo!?
U2 on LAUNCH - Exclusive greatest hits videos
http://launch.yahoo.com/u2
This archive was generated by hypermail 2.1.7 : Wed Apr 09 2008 - 22:16:19 EDT