Kira (kirsc.delete@this.aeterna.ru) on August 6, 2012 1:25 am wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on August 6, 2012 1:02 am
> wrote:
> > David Kanter (dkanter.delete@this.realworldtech.com) on August 5,
> 2012 3:49 pm
> > wrote:
> > > > > Oracle keep supporting the (far
> slower) M-series
> > SPARC?
> > >
> > > Honestly,
> > > the
> 'Itanium has bad performance' excuse
> > makes no sense. It's definitely
> faster
> > > than Niagara and potentially the
> > Fujitsu
> M-series.
> >
> > Potentially!? What it supposed to mean? Wait for
> >
> Poulson?
> >
> > >
> > > > I'd expect
> > > >
> >
> > fully loaded M9000
> > (64-socket SPARC64 VII+) to be approximately
> twice faster
> > >
> > > >
> > than fully loaded Superdome2
> (32-socket Tukwila) in majority of
> > >
> > commercial
> > >
> > applications.
> > >
> > > > Do you have a hard data
> >
> that could
> > > contradict my
> > > > expectations?
> > >
>
> > > I think
> > the relevant comparison point is
> > > Poulson,
> as it should be out any day
> > now.
> > >
> > > I'm a little
> surprised that they
> > > only have 32S systems,
> > since they used to
> scale up to 64S...and now they have
> > > better
> >
> interconnects.
> >
> > That's the same as Fujitsu. Back in dual-core days
> they used
> > to scale up to 128S, but now the biggest M9000 is 64S.
> >
>
> > >
> > > But I
> > don't perceive Itanium as being
> dramatically
> > > slower than Fujitsu.
> > >
> >
> > >
> DK
> >
> > Well, perception is one thing and the facts is something
> else.
> > We
> > certainly don't have enough facts to know whether top HP
> gear is "dramatically
> > slower" that than top Fujitsu M-series or just a
> little slower. But, IMO, we
> > have enough info to be sure that Kira's
> original claim (M-series is far slower
> > than Integrity) is false.
> >
>
> >
> > Anyway, the comparison between supporting Itanium
> > and
> supporting M-series was invalid regardless of relative performance, for
> more
> > than one reason:
> > 1. Oracle resells M-series. So, it's partly
> their own.
> > 2.
> > Incremental effort required to support M-series on
> top of T-series (the same OS,
> > the same ISA, mostly the same toolchain) is
> far lower than required to support
> > HP-UX/Itanium on top of ... well ...
> nothing.
> >
>
> Where's this proof that Tuk is slower than S64 VII? I posted
> SPEC numbers. Are you seriously claiming that a chip that has significantly
> lower SPECint numbers, a smaller cache, a shared bus, an off-chip memory
> controller, and supports an older generation of memory... is faster? Evidence,
> please? Or at least a rationale?

Well, see TCP-H below.
But even your own SpecInt_rate scores show single Tukwila socket slower than *two* Sparc64 VII+ sockets.

http://www.spec.org/cpu2006/results/res2010q4/cpu2006-20101201-13867.html
http://www.spec.org/cpu2006/results/res2010q2/cpu2006-20100426-10756.html

882 vs 531
Is it not enough?