Xserve & Xserve RAID Updated

[Mechcozmo]

Of course I have a maxed out Xserve setup. And I need the RAID array for my free iTunes collection. My wife has her own Xserve. What's annoying is that the kids are hounding me to get them one each too. I guess I'll have to give in or I wouldn't be a good father.

Lookin' to adopt? I like Japanese food, too.

 

[~loserman~]

Yes, but if you have embarrassingly parallel codes that will run on it, it's a pretty good answer.


I might ask, "where is the dual-core PPC970" ?? I don't see them on the Xserve or BladeCenter pages....

If you invented the numbers for that one (hmmm... (9.2*2)*(2.5/2.3) = 20) then I'll say that the dual-core Xeon is 16.2 ((7.2*2)*(3.6/3.2)).

What is the application, O/S, etc? Any real-world relevance? (Even IBM openly says that the MADD instruction on PPC gives it some great benchmark numbers that don't translate into real application speedups....)

IBM will release JS40 blade servers with 2 dual-core 2.5Ghz PPC970s as product late this month. We have ordered 2016 processors worth for delivery Oct 22.

The numbers I quoted are not made up they are the theoretical Maximums of the CPUs in question.

After close to 10000 hours of extensive application runs at our research facility, I can attest to the fact that our PPC clusters easily out perform our Xeons, Athlons, and Opterons. The only exceptions would be we have apps that have higher memory bandwidth needs and these perform best on our Opterons.

In scientific computing floating point and memory bandwidth/latency are the most important factors a platform provides.

I disagree with your assessment with Fused Multiply Add not adding much to real world applications. Sure it doesn't help out a word processor but almost every scientific application benefits from it greatly.
Scientific apps are what clusters are designed for in the first place.

I realize from a long history of reading your posts you tend to present the view that Intel hangs the moon but you also seem to have no real idea what scientific computing is about.

Are the largest portion of Clusters running on Intel and AMD you bet.
It's isn't because they are better though. It is because of the principle of commodity computing. The whole Beowolf Cluster Idea came out of using commodity parts to build scientific computers at a fraction of the cost by people(In the first case NASA Goddard Giga Ops Project) who needed some processing power but couldn't get any time on a Supercomputer.
PowerPC's have only lately become interesting as a commodity computing platform with their introduction in Xserves and JS20 and new JS40 blade servers. While they never really fit the bill as commodity their price performance made them very compelling.
Personally I think their main hold back probably was OS X and Linux. OS X is pure crap as an OS for scientific computing. It has terrible memory management and an absolutely horrible TCP/IP stack(latency out the wazzoo). Not to mention how inefficient it's scheduler is.( terrible stalls and blocking of processes on the CPU).
As to Linux it held it back also, because the 2.6 kernel didn't really support it well until 2.6.9+
Now the PPC970 will probably go into obscurity except for IBM.
In the end I think it can easily be proven as a technically superior chip to any Intel has ever offered with a possible exception to the ill fated Itannic.

It's O.K. for you to like Intel.(Hell for some things I like Intel too)but don't disseminate FUD.

 

[AidenShaw]

IBM will release JS40 blade servers with 2 dual-core 2.5Ghz PPC970s as product late this month.

OK, so I assume that you are testing on prototype or pre-production units?

The numbers I quoted are not made up they are the theoretical Maximums of the CPUs in question.

LOL - you're not even benchmarking real systems? Just peak theory?

When you said "We have done extensive testing on our Opteron and Xeon and PPC clusters" I assumed that you were running programs and timing them....

I disagree with your assessment with Fused Multiply Add not adding much to real world applications.

Not my assessment, IBM's:

http://www-03.ibm.com/chips/power/powerpc/newsletter/jun2004/ppc_process_at_work.html

In practice, only a small portion of peak capacity is achieved because a processor is rarely scheduled to do simultaneous “multiply and adds” in double precision.

However, the LINPACK benchmark, which is often used to rank supercomputers (the Top500 Supercomuter list), makes extensive use of simultaneous multiply and add.

 

[~loserman~]

Now that's what I call a wind tunnel.
That's over 12,000 fans in that one picture!

LOL.

You should hear what an Xserve was like with Linux on it prior to 2.6.9(before fan control). The fans were maxed out.
If those systems ran that the room would be deafening.

 

[Chaszmyr]

The new XServe RAID still doesn't use SATA, does that seem weird to anyone else?