- Niall Litchfield <niall.litchfield_at_dial.pipex.com>
wrote:
> Hi
>
> The bad news is that I don't believe that
> calculating IO/Sec *can* be done
> for a *new* system. At least I'd like to see how it
> is done. I'm willing to
> bet that any formula for doing it will include (x%)
> for 'overhead', which
> actually means 'stuff I don't know about'. Of course
> if the *new* system is
> a replacement for an old system with known IO
> requirements and the workload
> is similar (or predictably different) then obviously
> a calculation/lower
> bound could be set.
>
> (of course if one has the exact data set that you
> will use, and the IO
> required by each and every sql statement in use, and
> the exact number of
> clients and the exact machine and software
> configuration that will be used
> for always then one can measure your IO requirement.
> I have never seen such
> a situation.)
>
> Actually however I think that this bad news is
> rather mitigated by the fact
> that I don't believe that capacity can be calculated
> ahead of time for a
> *new* system either. It will entirely depend on the
> take up of the
> application and any changes to the design/usage post
> go-live.
>
> I think that that leaves us in a relatively good
> position, namely that we
> can estimate values for B,P etc based on our skill,
> judgement (and budget :(
> ), and that because none of the figures are *hard*
> figures it ought to be
> possible to negotiate *sensible* disk purchases.
> They key is to take into
> account all the demands on the system (as Cary
> says). I'm afraid that for
> *new* systems though getting into formulae for
> *calculating* requirements is
> likely to give false assurances. Time to brush up on
> negotiating skills (and
> to find how how to effectivey bribe your sys admin
> and/or budget holders).
>
> Yours unscientifically
>
> Niall
this is where the z-space kicks in (discreteness).
if you're using direct attached storage, the cabinets
typically hold 14 drives spread across a split
backplane per enclosure.
10 => PCI-X slots open > 0
4 Server PCI-X buses
n dual channel external SCSI RAID controllers
n dual channel external SCSI RAID enclosures
n*14 => drives > 0
so we will attempt to spread the IO across the PCI-X
bus channels, across RAID controller channels, across
drives.
so pick the following:
1 cabinet, 1 dual channel SCSI RAID controller, 14
drives
2 cabinets, 2 dual channel SCSI RAID controllers, 28
drives
3 cabinets, 3 dual channel SCSI RAID controllers, 42
drives (we have a winner :) ).
4 cabinets, 4 dual channel SCSI RAID controllers, 56
drives
Received on Tue Jan 20 2004 - 13:29:26 CST