ADSM-L

Re: Server Crash!!

1998-05-27 13:11:51
Subject: Re: Server Crash!!
From: Gene Mangum <gmangum AT UMICH DOT EDU>
Date: Wed, 27 May 1998 13:11:51 -0400
I believe they were referring to flushing disk buffers by issuing the
"fsync" system call.   It's not really bypassing the OS.   It tells the
OS that you need to be sure your data is committed to disk.

--
Gene Mangum
Gene Mangum
University of Michigan Medical Center


On Wed, 27 May 1998, Sanders, David wrote:

> " In fact, I once had a conversation with development
> who told me that where possible, ADSM will write through the operating
> system
> to circumvent any caching or buffering that the operating system may be
> performing.  "
>
> I don't understand that?  Everything has to use OS, right?  ADSM might
> use it's own channel programs but ultimately,,,,,,,,,,IOS does it??
>
> I'd be interested also in hearing the answer??
>
> Dave Sanders
> Sr. Technical Consultant
> DSanders AT massmutual DOT com
> 1295 State St, E060, Springfield, MA 01111
> 413-744-5095
> !@#$%
>
>
>
> > -----Original Message-----
> > From: Jason Meaden [SMTP:jasonm AT AU1.IBM DOT COM]
> > Sent: Wednesday, May 27, 1998 7:27 AM
> > To:   ADSM-L AT VM.MARIST DOT EDU
> > Subject:      Re: Server Crash!!
> >
> > G'day Dave,
> >
> > Right!
> >
> > However, when I say "my understanding is" - that is actually IBM speak
> > for - "I
> > am 99.99 percent certain".  In fact, I once had a conversation with
> > development
> > who told me that where possible, ADSM will write through the operating
> > system
> > to circumvent any caching or buffering that the operating system may
> > be
> > performing.  Of course, the disk hardware may have some level of
> > caching, and
> > there isn't much that can be done about that I guess.
> >
> > Someone in development who is monitoring this thread may have the
> > definitive
> > YES / NO for us.
> >
> > Regards,
> > --
> >   Mr Jason E Meaden                                  IBM Australia Ltd
> >   Software Service Specialist (Asia Pacific)         55 Coonara Avenue
> >   IBM Certified Specialist - ADSM             West Pennant Hills  2125
> >   Phone: 13 24 26 * Fax: 61 2 9354 7797 * Tie: 49427 * VM: RTP(MEADEN)
> >
> >
> >
> >
> >
> > ADSM-L AT VM.MARIST DOT EDU on 27/05/98 21:00:04
> > Please respond to ADSM-L AT VM.MARIST DOT EDU
> > To: ADSM-L AT VM.MARIST DOT EDU
> > cc:
> > Subject: Re: Server Crash!!
> >
> >
> > Hi Jason,
> > it seems then, that the issue is the buffering or not.  If it commits
> > after every activity, then mirroring is reasonable, if not it's not,
> > right?
> >
> > As far as the media, our server is MVS based and we have
> > fault-tolerant
> > devices so we can breathe a little easier, there.
> >
> > Dave Sanders
> > Sr. Technical Consultant
> > DSanders AT massmutual DOT com
> > 1295 State St, E060, Springfield, MA 01111
> > 413-744-5095
> > !@#$%
> >
> >
> >
> > > -----Original Message-----
> > > From: Jason Meaden [SMTP:jasonm AT AU1.IBM DOT COM]
> > > Sent: Wednesday, May 27, 1998 7:01 AM
> > > To:   ADSM-L AT VM.MARIST DOT EDU
> > > Subject:      Re: Server Crash!!
> > >
> > > G'day Dave,
> > >
> > > I can't say one way or the other if you should have mirroring
> > enabled
> > > or not.
> > > I can say that if you do, then you should let ADSM do it instead of
> > > the
> > > operating system (NT, AIX, or whatever).  You have to make a
> > judgement
> > > call on
> > > the cost of mirroring on performance and DASD usage compared with
> > the
> > > increase
> > > in availability.
> > >
> > > My understanding is that the log gets written too after every
> > activity
> > > and does
> > > not buffer the writes.  If you have a mirrored recovery log, and are
> > > using
> > > parallel writes, then any updates are written to both copies at the
> > > same time
> > > (for all intents and purposes).  In this instance, it is possible
> > that
> > > a system
> > > crash will cause a partial write to both copies thus negating the
> > > whole idea of
> > > mirroring.  This method will protect you from a media crash that
> > > removes one of
> > > the mirrored copies.  ADSM would mark the lost copy offline, and
> > > continue to
> > > use the other one.  You would then shut down everything smoothly,
> > > replace the
> > > faulty media, and restart without too much effort.
> > >
> > > If you are using sequential writes to your mirrored copies, then
> > ADSM
> > > will set
> > > a dirty flag on the primary copy, perform the I/O, ensure the write
> > is
> > > successful, then flag the copy as 'clean' then it will set the dirty
> > > flag on
> > > the secondary copy and perform the I/O on that copy, ensure it was
> > > successful,
> > > then mark that copy as clean, and repeat.
> > >
> > > In the event of a server crash, ADSM will automatically detect which
> > > copy is
> > > clean at restart and use that version.  Yes, this causes a
> > performance
> > > hit.  In
> > > the grand scheme of things, not a huge penalty I think, but
> > something
> > > you
> > > should test on your own system.
> > >
> > > On the other hand, I have customers who do not care about the
> > recovery
> > > log.
> > > They backup their database twice a day, and accept that if they lose
> > > the
> > > recovery log, they may lose up to 12 hours of updates.
> > >
> > > Regards,
> > > --
> > >   Mr Jason E Meaden                                  IBM Australia
> > Ltd
> > >   Software Service Specialist (Asia Pacific)         55 Coonara
> > Avenue
> > >   IBM Certified Specialist - ADSM             West Pennant Hills
> > 2125
> > >   Phone: 13 24 26 * Fax: 61 2 9354 7797 * Tie: 49427 * VM:
> > RTP(MEADEN)
> > >
> > >
> > >
> > >
> > >
> > > ADSM-L AT VM.MARIST DOT EDU on 27/05/98 20:32:33
> > > Please respond to ADSM-L AT VM.MARIST DOT EDU
> > > To: ADSM-L AT VM.MARIST DOT EDU
> > > cc:
> > > Subject: Re: Server Crash!!
> > >
> > >
> > > Since I tend to be against mirroring, this could change my mind if
> > > what
> > > Jason says is true!
> > >
> > > There are 2 questions that I have:
> > > 1) Does the recovery log get written to after every activity, or
> > does
> > > it
> > > buffer activity for awhile before writing?
> > >         *if every activity immediately writes to the primary log and
> > > then to the mirrored-log, then I guess I'd have to rethink the
> > > mirroring thing
> > >         *if every activity DOESN't get written to the primary log
> > > immediately, then mirroring accomplishes nothing since the content
> > > of the buffer is broken and either neither log gets written to, or
> > > both
> > > get written to badly.
> > > 2) If you have a busy system and a -write to a mirrored log- has to
> > > wait
> > > for a write to finish on the primary log, does this negatively
> > affect
> > > performance?
> > >         *I guess this would have to be accepted as a cost for having
> > > the
> > > redundancy
> > >         *Or, it's not detectable
> > >
> > >
> > >
> > > Dave Sanders
> > > Sr. Technical Consultant
> > > DSanders AT massmutual DOT com
> > > 1295 State St, E060, Springfield, MA 01111
> > > 413-744-5095
> > > !@#$%
> > >
> > >
> > >
> > > > -----Original Message-----
> > > > From: Jason Meaden [SMTP:jasonm AT AU1.IBM DOT COM]
> > > > Sent: Wednesday, May 27, 1998 5:47 AM
> > > > To:   ADSM-L AT VM.MARIST DOT EDU
> > > > Subject:      Re: Server Crash!!
> > > >
> > > > G'day Scott,
> > > >
> > > > You have a corrupt recovery log due to a partial write that was in
> > > > progress
> > > > when the system died.
> > > >
> > > > You should probably restore the DB from your last backup.  Don't
> > > > bother with a
> > > > rollforward, even if you had that enabled.  It will probably roll
> > in
> > > > the same
> > > > error, and the server would still not start.
> > > >
> > > > You could also do a 'dump load audit' but this is very time
> > > consuming
> >
> >
> >
> >
>
<Prev in Thread] Current Thread [Next in Thread>