Re: EVA question: How many vdisks should I create?
From: Scott Vieth (svieth_at_wi.rr.com)
Date: 10/07/03
- Next message: Fred Kleinsorge: "Re: DS15 systems have arrived"
- Previous message: Scott Vieth: "Re: EVA question: How many vdisks should I create?"
- Maybe in reply to: David J. Dachtera: "Re: EVA question: How many vdisks should I create?"
- Next in thread: Rob Young: "Re: EVA question: How many vdisks should I create?"
- Reply: Rob Young: "Re: EVA question: How many vdisks should I create?"
- Reply: jlsue: "Re: EVA question: How many vdisks should I create?"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: 7 Oct 2003 08:08:49 -0700
> This is all news to me. Our normal config would be as few LUNs as
> possible, with as many disk spindles in the group as possible. Regardless
> of the OS.
This is not the "best" way to do things. Other OSes have built-in
throttling mechanisms which regulate the number of pending IO requests
per LUN. For example, Windows or Tru64 have a "queue depth" of
something like 32. Don't quote me on that number.
If you have one LUN (on an EVA) served to a Windows box, and that
Windows box has 32 pending IO request to that LUN, it will wait for
one of the IOs to complete before it issues the next IO to that LUN.
So if you have one LUN on an EVA (still taking Windows here), you can
have (1 x 32) pending IO requests. If you create ten LUNs on the same
EVA and present them to the same Windows box, now you can have (10 x
32) pending IO requests on that system.
*This* is why I originally asked the question about having one LUN vs.
many LUNs.
On a Windows system or Tru64 system, you want *more* LUNs so you can
throw more IOs at an EVA.
On VMS, it doesn't matter. The OS doesn't throttle IO requests like
that (which is why I can bury my HSG80s with VMS 7.3-1 and make the
HSG80s go tits-up).
I don't have a Windows box connected to my EVA (except for the SANMA)
to test this out. I've got a Solaris box (which probably has the same
"you can only issue 'n' IOs per LUN before I'll stop sending IO
requests" throttle that Tru64 and Windows do) connected to the EVA but
I don't have time to tear down the single LUN and divide it into a
whole bunch of LUNs in the name of performance. The IO performance on
that system is not as critical as the performance of our VMS systems.
> There are other more pressing considerations in the area of managing very
> large LUNs (backup, restore, dept. ownership of xxGB, whatever). Most of
> these can be alleviated by changes in operating procedures (though, not
> always the political ones can be solved). If that's not an issue in your
> shop, then larger LUNs should work fine, regardless of the OS.
Since VMS does not have the "throttling" mechanism I mentioned above,
we are going to focus our design on matching the number of LUNs to the
number of backup streams we want to run in parallel.
Since we have four DLT8000 tape drives directly attached to the ES40
where our production app (IDX) runs, we are going to present four
LUNS. Yes, we could get away with one HUGE lun, but then we could
only run one backup stream.
If anyone from VMS Engineering or Storage Engineering reads what I
posted above and disagrees with the way I explained the throttling in
various OSes, please speak up. This topic was addressed in one of the
EVA-related seminars in Atlanta. I don't have the name of the seminar
at-hand but I could dig it up if someone wants to research this issue
further. The Powerpoint presentation from that seminar is probably
available online.
Thanks,
-Scott Vieth :^)
- Next message: Fred Kleinsorge: "Re: DS15 systems have arrived"
- Previous message: Scott Vieth: "Re: EVA question: How many vdisks should I create?"
- Maybe in reply to: David J. Dachtera: "Re: EVA question: How many vdisks should I create?"
- Next in thread: Rob Young: "Re: EVA question: How many vdisks should I create?"
- Reply: Rob Young: "Re: EVA question: How many vdisks should I create?"
- Reply: jlsue: "Re: EVA question: How many vdisks should I create?"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|