Re: Sun may have Solaris for Itanium already ready
From: Yousuf Khan (removethisspam.bjsk90.removethispam_at_hotmail.com)
Date: 10/29/03
- Next message: Tim: "Re: SunPCI 1.3 just returning a blank window"
- Previous message: CJT: "Re: Sun may have Solaris for Itanium already ready"
- In reply to: Peter Perlsų: "Re: Sun may have Solaris for Itanium already ready"
- Next in thread: Sander Vesik: "Re: Sun may have Solaris for Itanium already ready"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: Wed, 29 Oct 2003 22:53:46 GMT
"Peter Perlsų" <nospam@nospam.com> wrote in message
news:3f9fc52b$0$27375$edfadb0f@dread16.news.tele.dk...
> > Anything where the words "high performance" might be uttered. If you
> > are running a process that has to share memory/data then you better not
> > use a cluster or you lose 90% of the performance.
>
> Nonsensical generalization. It depends on the task.
>
> For scientific uses, I can think of the DTU here in Denmark which, one
> year, installed a cluster for computing protein folding, and leasing of
> processing time for international customers. That cluster is comprised
> of 32 PowerMac G4 duals, connected by a 100 Mbit ethernet. After some
> months of testing and tweaking, the buyers were very impressed with the
> performance.
There's only a small subset of problems that can be handled well by loosely
coupled clusters of computers, such as what you're describing. Here's a
recent Cray marketing blurb about a new line of Opteron-based supercomputers
that they are now marketing with a proprietary interconnect technology which
Cray calls Black Widow. Their rationale for going with a proprietary
interconnect was precisely because not all supercomputing tasks can be
handled well through a loosely-coupled supercluster. The speed of the
interconnects make a big difference to what kind of problems it can handle.
--------------------------------------------
http://www.cray.com/news/0310/rsproduct.html
"Red Storm is architected from the ground up for superior reliability and
ease-of-administration," he said. "It provides redundancy features and
powerful capabilities for system-wide management, including resiliency and
repair in the event of disk or processor failures. This creates the basis
for better scientific productivity and progress." Clusters loosely link
together multiple servers or PCs with relatively low-bandwidth connections
that are inadequate to make efficient use of the processors. "Clustered SMP
systems with commercial interconnects are fine for handling small problems
or big problems that are simple in nature, but their efficiency can drop to
less than five percent on really challenging problems and workloads, versus
five to ten times better than that for a well-designed MPP system like Red
Storm," Ungaro said. "For challenging work, MPP systems are far more
cost-effective."
--------------------------------------------
It should be noted that there are other supercomputer being built with
Opterons also, but they cost much less than Cray's Red Storm or smaller
versions of it. That's because they use off-the-shelf interconnects such as
networking gear, or Infiniband. But as they say, you get what you pay for.
> > Thousands. All of the Fortune 500 for a start. Automakers, the
> > petrolium industry, each vile branch of the government, banks and
> > investment firms, aerospace, internet search and profiling (google,
> > yahoo, etc), and the list goes on. On a much smaller scale, one of our
> > simulators is written in C and gets compiled using standard compilers.
> > Within the last two years we tried to move it from HP-UX on PA-RISC to
> > Linux on Xeon and couldn't even compile. The compiler demanded more
> > memory than the 4GB address space of Xeon allowed, just to optimize and
> > link a single executable.
>
> Thousands... in a computing world that will count its unit number in
> billions before too long.
What exactly has billions of little PCs got to do with any of this? The
requirements for databases with 100's of Gigs of data are beyond question --
they exist. They do exist and they exist quite commonly. Quite a few are
already running in the Terabytes.
There are also billions of little houses in this world, but businesses seem
to want to operated out of big office buildings and factories, and there
are only 100's of thousands of those in this world. So what do you suggest
that businesses should operate out of houses from now on exclusively, no
matter what they do?
> >> And again, who says that such a gargantuan hypothetical database cant
> >> be split up across multiple systems, and be served in a distributed
> >> manner ?
Databases require a common store for performance reasons. It's as simple as
that. They do split databases over multiple systems in parallel right now,
but they always use a common store, such as a large shared disk array. But
each of the parallel instances of the database still need to keep as much of
the database in memory locally as they can too.
> At least I know that the Xeon has an physical address space of 64
> gigabytes, and not 4 as you claim.
The Xeon's 64GB address space is a hack which is only available to the
operating system and not the applications. This hack allows an OS to keep
more applications in memory at the same time, but each application is still
limited to less than 4GB. So a database (being an application) would still
be limited to 4GB.
Yousuf Khan
- Next message: Tim: "Re: SunPCI 1.3 just returning a blank window"
- Previous message: CJT: "Re: Sun may have Solaris for Itanium already ready"
- In reply to: Peter Perlsų: "Re: Sun may have Solaris for Itanium already ready"
- Next in thread: Sander Vesik: "Re: Sun may have Solaris for Itanium already ready"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
