Hi Tom,
Its possible that your backup throughput of
5-10 MB/sec via NDMP to a single DLT drive
is limited by the DLT tape device. Anyone
with a filer can run basic benchmarking on
their filesystem(s) by running a test "dump
to null" by running on the console or via
rsh:
dump 0f null /vol/vol_name
Our rule of thumb is to add another 5%
overhead in time for the actual writing to
tape.
Our supported tape devices has increased
recently in 6.1 to include:
AIT
AIT-2
DLT7K
DLT8K
Mammoth-1
Mammoth-2
9840
3590E
You may get faster throughput either by
running concurrent qtree backups from each
volume or by employing a faster tape device.
I'll preempt the "when are you going to support
LTO" question by saying we have this drive in
our dev labs, but please speak to your sales
team about availability of this and other tape
devices.
I would be interested in receiving either private
or public (in this forum) what you (our customers)
are planning to do regarding tape devices in the
short & long term though.
Cheers,
Grant
-----Original Message-----
From: Tom "Mad Dog" Yergeau [mailto:MadDog@fool.com]
Sent: Tuesday, May 01, 2001 3:21 PM
To: 'Eric Hester'; ''Toasters@Mathworks. Com' '
Subject: RE: Backup solutions for NDMP with F740's
Eric,
When we went from over-the-wire backups to direct-attach NDMP backups, I saw
a big increase in backup performance. Especially in incremental backups.
NetApp is very good at finding changed files and dumping them to tape
quickly, since on an NDMP backup WAFL figures out what has changed. If
you're backing up over the network, your backup server has to parse your
entire directory structure looking for changed files. WAFL, on the other
hand, can just do an inode scan which is blazingly fast compared to walking
a large directory tree over the wire.
We generally see 5-10 MB/sec throughput via NDMP to a single DLT drive,
depending on how well the data compresses. Oddly enough, database files
compress really well and gzip'd files compress really poorly. :-)
The only downside is that we use NetBackup 3.4, which does not support the
NDMP "DAR" feature. Which means if you have a backup that spans three
tapes, and you need to restore a 5K file that's on the third tape, you have
to read the first two tapes and and to get to the file you need. So a
restore can be painful. But, that is VERITAS' fault for not supporting DAR,
and if you use Legato or Syncsort, I think you won't have this problem.
(If anyone from VERITAS is on this list... DAR support would be really
really nice!)
MD
-----Original Message-----
From: Eric Hester [mailto:erich@nuvox.net]
Sent: Tuesday, May 01, 2001 4:47 PM
To: ''Toasters@Mathworks. Com' '
Subject: Backup solutions for NDMP with F740's
We have been using an ADIC Scalar 218 for some time with our F740 cluster
with
fairly good results until recently. Our Netapps provide the storage for our
Web
and Mail clusters. We have experienced pretty serious growth and are seeing
some
significant slow downs in backup turnaround time. We have up until this
point
also just been backing up the data over NFS mounts to our NetBackup server.
Being that most of the data comes from the mail server side and we use
maildir,
it is not a huge amount of data but is a huge number of smaller files. I'm
not
very confident that NDMP would completely correct the problem, but it seems
that
removing the network/nfs overhead couldn't hurt. I think I have made the
decision to go with the ADIC Scalar 1000 with 2 drives dedicated to the
Netapps
(1 to each in the cluster) with NDMP and 4 more drives dedicated to the
backup
server for other systems.
2 questions:
1) What are some impressions on NDMP as a performance gain for backups?
Right
now the network does not seem to bottleneck the process as we have dedicated
100m links from the servers to a separate NFS/backup network, and 4 channel
etherchannel from each netapp to that same network. From MRTG graphs this
network is mostly idle (10-20m max) even during backups. I don't appear to
be
able to keep the DLT's saturated (in other words the activity light doesn't
stay
on the whole time, so it is not streaming non stop) during backups, and I
don't
know if this is NFS overhead, or there is another problem. Is NDMP more
capable
of keeping a DLT fully saturated and therefore more efficient, or do I still
need to look at this as a Netbackup issue?
2) A little OT but - Does anyone have any good experience with this
particular
library (ADIC Scalar 1000)? I loved our smaller Scalar.. but have more long
term
experience with StorageTek. My salesperson has done a good job convincing me
of
the advantages of this box, and I am pretty much sold. Just curious if
anyone
else has any horror stories or glowing recommendations.
Eric Hester
Director of Systems Engineering
NuVox Communications
eric(a)nuvox.net