ADSM-L

Re: DB storage/performance question ADSM or Legato?

1998-11-02 18:58:30
Subject: Re: DB storage/performance question ADSM or Legato?
From: Ben Kokenge <ben AT EDMS DOT NET>
Date: Mon, 2 Nov 1998 17:58:30 -0600
Thanks for the input, and I hope this thread continues.  I have developed
a couple solutions on my own and I would like to see what y'all think.

As I have mentioned, we have 4 servers (HP-UX running oracle and sybase)
with about 200GB each that need to be fully backed up nightly  Here is what
I propose:

1. Run an ADSM server on each DB server.  That way there is no network
issues and I know I can write 5 or 6MB/sec per 3590 drive that hangs off
the server via SCSI.  Stream that 3 or 4 times and you get ~20MB/sec, which
is 72GB/hr.  I site:  http://www.storage.ibm.com/software/adsm/adperfpr.htm
where they backed up 736GB in 1.5 hours.  Much overhead and drives with
ADSM server DB on each HP box.  I could then do server to server backup
with all the HP boxes which is beneficial.

2. This is a stretch, but the IBM VSS is out in force now.  I could have
the actual oracle DB stored on the VSS with all it's mirroring properties.  Then
I can have an ADSM tape library on one of the HP servers to do the backups
from the VSS which is local to the system, b/c of the functionality of the VSS.
In other words have all DB servers store data on the VSS (Versatile Storage 
Server)
with direct sacs connect.  Then the one ADSM server can backup all the data
quickly.  Run five 3590's concurrently and you are looking at 100GB/hr or 
better.
That is acceptable backup time.

I think David has a valid point where one should look at the restore time via
hardware crashes and the like, but over 50% of data loss is due to user
error, so DB restores from corrupt file are more likely.

These are my thoughts.

Thanks,  Ben



DAVID HENDRIX wrote:

> Matthew poses the same logic and questions we usually see: let's just
> mirror the thing real well and get real fault tolerant.  Then we wont need
> any daily/weekly/whatever backups - right?
>
> Unfortunately, this scheme stops working when an unforseen programmer
> error, user error, etc. gets propogated to the primary and all copies of
> the database in question.  I usually win the argument here on this issue by
> citing two examples our revenue systems had with this kind of problem.  One
> was very serious and had only been discovered 3 MONTHS after the problem
> actually occured.  These examples I use could have cost our company
> millions of dollars if we had not had a "backup" - the mirror's were
> corrupt, the primaries were corrupt and there was no way to rebuild.  But,
> hey, the hardware was just fine.
>
> If you don't mind taking the unforseen hit, then just mirror - it's a
> business decision, not a technical one.
>
> We struggle with providing appropriate hardware and resource to backup all
> of our growing and very large databases (most these days come in the 100's
> of GB to +1TB).  SAN, FC/AL, etc. will help.  The software needs to
> catch-up and we need to stop funneling data "through" a server and start
> sending data from device to device.  ADSM, Networker, Netbackup,
> Alexandria...doesn't matter.  VLDB toasts each one because the paradigm
> requires the client/server communication.  Yes, even with backup clients
> (Networker), and slave servers (Netbackup).  You still need sufficient
> backplane, CPU and I/O bandwidth on the "server" to send data to local
> devices at appropriate speeds without impacting the application you are
> trying to protect.
>
> ADSM doesn't win the performance war, but it does well enough for us (last
> example I gave to this list server was ~6MB/sec/stream for an aggregate
> ~300GB/hour backup speed on a E10K with ADSM).
>
> We are looking forward to see what the vendors do with SAN.  They can't get
> there fast enough for us...
>
> Thanks,
>
> David
<Prev in Thread] Current Thread [Next in Thread>