ADSM-L

Re: multiple instance recommendations

2006-05-19 22:06:03
Subject: Re: multiple instance recommendations
From: "Allen S. Rout" <asr AT UFL DOT EDU>
To: ADSM-L AT VM.MARIST DOT EDU
Date: Fri, 19 May 2006 22:05:38 -0400
>> On Fri, 19 May 2006 15:13:10 -0500, Dave Mussulman <mussulma AT UIUC DOT 
>> EDU> said:


> I have questions about server sizing and scaling.  I'm planning to
> transition from Networker to TSM a client pool of around 300 clients,
> with the last full backup being about 7TB and almost 200M files.  The
> TSM server platform will be RHEL Linux.

> I realize putting all of that into one TSM database is going to make it
> large and unwieldy.

You may be underestimating TSM; the size there is well in bounds for
one server.  The file count is a little high, but I'm not convinced
it's excessive.  The biggest server in my cluster has 70M files, in a
70% full DB of 67GB assigned capacity. So if that means 67GB ~= 100M
files, you might be talking 130-150GB of database.  I wouldn't do that
on my SSA, but there's lots of folks in that size range on decent
recent disk.

Do you feel your file size profile is anomolous?  My 70M file DB is
~23TB; that ratio is an order of magnitude off mine, and my big server
is a pretty pedestrian mix of file and mail servers.

> My understanding of a shared SCSI library indicates that the library
> is SCSI-attached to a server, but drive allocation is done via SAN
> connections or via SCSI drives that are directly attached to the
> different instances.  (Meaning the directly attached SCSI drives are
> not sharable.) Is that true, at least as far as shared libraries go?
> The data doesn't actually go through the library master to a
> directly connected drive, does it?

I have heard of some dual-attach SCSI setups, but never actually seen
one in the wild.  If I were going to point at one upgrade to improve
your life and future upgrade path, getting onto a shareable tape tech
would be it.  I have drunk the FC kool-ade.  It's tasty, have some. :)


> Other than the obvious hardware cost savings, I don't really see the
> advantage of multiple instances on the same hardware.  (I haven't
> decided yet if we would use one beefy server or two medium servers.)
> If you load up multiple instances on the same server, do you give
> them different IP interfaces to make distinguishing between them in
> client configs and administration tools easier?

They _must_ have different ports.  I use the same IPs, but different
CNAMES so I can move things around and not unnecessarily piss off
customers.  If you felt like using different IPs, I expect that would
work just fine too.


> Tape access-wise, is there a hardware advantage putting multiple
> instances on the same system?

Yes, it solves your drive sharing problem.  All the TSM instances
would be looking at /dev/rmtXX.  Your LM instance can do traffic
direction to figure out who's got the drive, and they are all using
the same bus, same attach.

> Any recommendations on any of this?  Your help is appreciated.


I like the 'beefy box' solution for all purposes except test instance.
Make sure it's got plenty of memory. 6G? 8?


- Allen S. Rout