[rescue] IBM on ebay
Francisco Javier Mesa-Martinez
lefa at cats.ucsc.edu
Fri May 17 20:58:06 CDT 2002
On Fri, 17 May 2002, Sridhar the POWERful wrote:
> On Fri, 17 May 2002, Francisco Javier Mesa-Martinez wrote:
>
> > > The only RS/6000's that have anything approaching "mainframe class"
> > > performance are SP's, which are essentially smaller RS/6000's with
> > > ancillary I/O processing and shared-memory clustering. I want one. 8-)
> >
> > As a poor fool that has had to spend many sleepless nights programming
> > SP system(s), I can tell you that the SP/2 is not a shared memory
> > machine. There is no single adress space, it is just a glorified
> > cluster with each node being an actual RS6K workstation, and some
> > fancy switching topology for internode communication. The only ime
> > shared memory operation can be achieved is inside an SMP node, all
> > internode communication is performed via message passing (usually
> > using some sort o'MPI).
>
> I played with PE on regular RS/6000's and seen the memory exchange/locking
> stuff. What does this stuff do? I haven't really done anything with it,
> but it looked like explicit transfer memory sharing.
PE (POE as I think is also called) is pretty much a development
environment. But you still need MPI, since the programming model for the
SP2 is still message passing based.
> > Why would anyone want one of these machines is beyond me... but it
> > seems that IBM was able to market them pretty well.... A large SGI
> > system with ccNUMA, now THAT I do want!
>
> There *are* IBM NUMA machines. Just not RS/6000's.
Yeah, the numa-q I guess, but it seems that IBM is not marketing the
sequent stuff too well. And I do not know how much of the numa-q is being
applied to future RS6K, a nice cluster of power4 with NUMA would kick some
serious ass...
> Peace... Sridhar
>
> --
> "How do you fight such a savage?"
>
> "With heart, faith, and steel. There can be only one."
> -MacLeod and Ramirez, "Highlander"
More information about the rescue
mailing list