[Veritas-bu] NDMP backup delays - solved!
2002-09-10 09:25:52
Subject: |
[Veritas-bu] NDMP backup delays - solved! |
From: |
jlkennedy AT amcc DOT com (Jeff Kennedy) |
Date: |
Tue, 10 Sep 2002 06:25:52 -0700 |
Yep, this has been a serious problem for me in the past, sorry I didn't
catch it in your original post. One weekend I found my filer backups
were just hanging there for 30+ hours. I ran iostat on the master
(Solaris) and found that the internal drive was 100% busy. It took me a
bit to track it down but it was /usr/openv/netbackup/BPFSMAP_TMPDIR. In
there is where the inode maps are created and then checked. Once I
found the problem I linked that dir to a raid 1+0 array and they dropped
from 30+ hours to just under 16. With 4.5 I have seen it drop even
more.
~JK
Moshe Linzer wrote:
>
> You guys are right. I suspected the catalog processing, but couln't
> be sure. An email from Karen Shoenbauer of Veritas clued me in to
> checking the bptm logs. There I saw that after the backup completed
> for the first volume, NBU kicks off ndmp_bpfsmap_create, which runs
> for 17 hours! Only after this completes does the backup of the next
> volume begin.
> Apparantly 4.5 eliminates this catalog process (as Steve mentions), so
> it should speed up our total backup time significantly!
>
> As far as the dump filesystem walking goes, this has been improved
> with subsequent releases of Netapp software, and is now much quicker
> than it was.
>
> Moshe
>
> Steve Kappel wrote:
>
> > There are two things going on here. One is the catalog post-
> > processing that NetBackup NDMP must do to convert the inode-based
> > information provided by the NetApp into the native path-based
> > catalog. This overhead depends on the number of objects in
> > the backup. Starting with NetBackup 4.5 there is no longer any
> > catalog post-processing so this overhead is eliminated.
> > Secondly, dump walks the filesystem before it starts sending
> > backup data. A quick google search comes up with this
> > short overview of dump:
> > http://www.usenix.org/publications/library/proceedings/osdi99/full_papers/hutchi
> > nson/hutchinson_html/node6.html
> > Jeff Kennedy wrote:
> >
> >> I believe the problem is *before* the dump even begins. I have
> >> seen
> >> this problem on volumes that have a tremendous number of small
> >> files or
> >> are terribly fragmented. But I've never seen the problem it just
> >> start
> >> out of the blue, it's usually a gradual process that gets worse
> >> with
> >> time.
> >> ~JK
> >> John D Stephens wrote:
> >>
> >> > Moshe -
> >> > I have seen huge delays between volumes as well. What is going
> >> > on is this. After NBU starts the NDMP backup, it steps out
> >> > of the way and idle until the filer signals NBU that the dump
> >> > has completed. Then the filer sends all the meta data to NBU
> >> > and NBU then catalogs that data. Depending on the size of your
> >> > backup and the amount of files, it takes a little while for NBU
> >> > to chug through the catalogs. Hopefully, 4.5 is faster at this.
> >> > Check your iostat -x and look for the activity on your catalogs
> >> > after the filer finishes sending the NDMP job, but NBU still
> >> > shows it to be active. The only way around this is to use
> >> > multiple streams. Be carefull, multiple streaming does use the
> >> > filer's CPU more than normal.
> >> > Hope this helps.
> >> > John
> >> > Moshe Linzer wrote:
> >> >
> >> >> Lately we have started seeing delays between volumes when
> >> >> backing up our
> >> >> Netapp filer. I have a single AIT-2 tape attached to my F760,
> >> >> and I run
> >> >> a full backup of the filer, which contains 3 volumes. The
> >> >> performance
> >> >> on the first volume is as expected, but then there is a delay
> >> >> of
> >> >> sometimes up to 12 hours before the second volume begins to be
> >> >> backed
> >> >> up! The same thing happens before the third volume. volume
> >> >> sizes are
> >> >> between 375-480GB. I think this began after we applied some
> >> >> patches
> >> >> lately, but the patch descriptions don't seem to have any
> >> >> bearing on
> >> >> this problem. Had anyone seen this type of behaviour?
> >> >> Thanks,
> >> >> Moshe
> >> >> _______________________________________________
> >> >> Veritas-bu maillist - Veritas-bu AT mailman.eng.auburn DOT edu
> >> >> http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
> >> >>
> >> > --
> >> > +=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+
> >> > + John D Stephens ITS Design Systems +
> >> > + Texas Instruments 12500 TI BLVD, Dallas +
> >> > + jstephens AT ti DOT com 214-480-6229 +
> >> > +=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+
> >> > _______________________________________________
> >> > Veritas-bu maillist - Veritas-bu AT mailman.eng.auburn DOT edu
> >> > http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
> >> >
> >> --
> >> =====================
> >> Jeff Kennedy
> >> Unix Administrator
> >> AMCC
> >> jlkennedy AT amcc DOT com
> >> _______________________________________________
> >> Veritas-bu maillist - Veritas-bu AT mailman.eng.auburn DOT edu
> >> http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
> >>
> > __________________________________________________________________________
> > Steve Kappel steve.kappel AT veritas DOT com
> > VERITAS Software (Engineering)
> >
--
=====================
Jeff Kennedy
Unix Administrator
AMCC
jlkennedy AT amcc DOT com
|
|
|