Colsa to TRIPLE Army cluster by year end

[macsrus]

I just heard that Colsa plans to triple the size of the Armys cluster by years end. It is suppose to grow to over 4672 xserves.

Can You imagine thats over 74 TFlops

I also heard that they will be adding Infiniband to the complete system later this year.....

Cool News Huh......

 

[PlaceofDis]

Link?

 

[Darwin]

Link first

But if it is true then they will be No.1 on the SuperComputer list, of course not that they where going to be that far down in the first place

 

[Falleron]

Sounds iffy to me. Would be interested in a link...

 

[mj_1903]

Coming from a person who says recently:

HeHe looks like Im buying.... because I just found out IT WILL HAVE A 64 BIT KERNEL and it NO LONGER will HAVE ANY OF THE JUNKY/CRAPPY (NEXT) STUFF LEFT IN IT...... YEA NETINFO IS GONE......

Ok.... sure, I believe you.

Tiger has a 32bit kernel, Tiger has the NeXT stuff still in it because it is Mac OS X and Apple has not dropped NetInfo.

Actually, the NeXT 'stuff' is quite possibly what makes Mac OS X so great.

 

[macsrus]

Tiger has a 32bit kernel, Tiger has the NeXT stuff still in it because it is Mac OS X and Apple has not dropped NetInfo.

Actually, the NeXT 'stuff' is quite possibly what makes Mac OS X so great.

Actually tiger has a 64 bit kernel..... That will be anounced next week at WWDC....
And as to your remark about Next stuff making OSX GREAT...... ARE YOU ON DRUGS ?...... BSD is what makes the OSX great.....

And Netinfo has been depricated in OSX since Jaguar...... each release removes more and more services from netinfo.....
As a matter of fact it can almost be completly turned off in Panther right now....
Apple has been, and is continuing to move everything out of netinfo and into ldap....

 

[legion]

A) The 25Tflops for the initial system is the theoretical max performance. Colsa has already said that they belive the most they could get as a Rmax is 15Tflops which would place it at 3rd on the current list. Think of it this way, the System X Terrascale cluster had 2200 G5 chips and attained around 10Tflops. How would 3,132 G5 chips in Mach 5 attain 25Tflps? That would mean with a 42.36% gain in chips they are getting a 150% gain in performance. (And, BTW, they were already using Infiniband.)

B) Even if they tripled the cluster, clusters don't scale linearly. They actually scale closer to a logarithmic curve.

C) Does anyone honestly believe that Apple could deliver that many Xserves??!!

D) It will never be at the top of the list because BlueGene /L (final) will be done by then and that will be at a measurable (not theoretical) Rmax of 350Tflops.

E) If you're going to make stuff up, at least make it seem somewhat interesting like a move to Quadrics fabric switches.

 

[macsrus]

A) The 25Tflops for the initial system is the theoretical max performance. Colsa has already said that they belive the most they could get as a Rmax is 15Tflops which would place it at 3rd on the current list. Think of it this way, the System X Terrascale cluster had 2200 G5 chips and attained around 10Tflops. How would 3,132 G5 chips in Mach 5 attain 25Tflps? That would mean with a 42.36% gain in chips they are getting a 150% gain in performance. (And, BTW, they were already using Infiniband.)

Virginia Tech never achieved better than 60 percent effcientcy

For 2 reasons

1. Their system was never stable because of the lack of ECC memory...
They had trouble even keeping the system up at all.... much less the time it takes to complete Linpack runs

2. The problem size they ran under linpack was Small. It calculates out to barely 1 gig of memory per processor..... If you check all other top 10 systems you will find that they all used approx 2 gig memory or better for their Linpack problem sizes. Heck the current number 2 (thunder) uses almost 4 gig per processor...
Using a large problem size and using xserves instead of the desktops should increase TECHS linpack results to 11 to 12 Tflops....

Now for what you have pointed out about Colsa's info.... from their press release.... I agree. They said they would only reach about 15 Tflops.
Notice they also said they were using Gig E.
15 Tflops is an optomistic number for Gig E.... but is not unreasonable

I have little doubt that a cluster of 4500 xserves with infiniband could reach 50+ Tflops Rmax

As far as clusters not scaling.... that has not been proven yet....
As a matter of fact current results of large systems have been showing the opposite to be true.

Lastly... while Quadrics is the reigning king of high speed low latency Cluster Interconnects.... it will not remain so for long....
Infiniband will replace all cluster interconnects due to ....
1. it being a standard non proprietary solution
2. it has a very aggressive upgrade roadmap.... and has been meeting that roadmap so far....
3. Since it is offered by several companies it is already better priced than competing solutions

I expect by the end of 2005 to early 2006 noone will even be talking about the myrinet or quadrics solutions at all

 

[Abstract]

A) The 25Tflops for the initial system is the theoretical max performance. Colsa has already said that they belive the most they could get as a Rmax is 15Tflops which would place it at 3rd on the current list. Think of it this way, the System X Terrascale cluster had 2200 G5 chips and attained around 10Tflops. How would 3,132 G5 chips in Mach 5 attain 25Tflps? That would mean with a 42.36% gain in chips they are getting a 150% gain in performance. (And, BTW, they were already using Infiniband.)

They used slower processors than the ones available today. They also didn't use ECC.

 

[mymemory]

If you go and help you get free pizza!

 

[King Cobra]

If you go and help you get free pizza!

Gee, only if Folding did the same... every week...

Yeah, I'm contesting the 75Tflop comments above (well, at least the ones that support it). Triple the number of processors does not mean triple the flops. Even if the old max. was 15Tflops, the upgraded supercluster prob. won't reach 45Tflops. I'm not sure what the final number will come out to be, but it'll be interesting to see how close the upgraded supercluster compares with the Earth Simulator (prob. my favorite computer cluster ever) in terms of Tflops.

 

[mj_1903]

Actually tiger has a 64 bit kernel..... That will be anounced next week at WWDC....
And as to your remark about Next stuff making OSX GREAT...... ARE YOU ON DRUGS ?...... BSD is what makes the OSX great.....

And Netinfo has been depricated in OSX since Jaguar...... each release removes more and more services from netinfo.....
As a matter of fact it can almost be completly turned off in Panther right now....
Apple has been, and is continuing to move everything out of netinfo and into ldap....

Actually, no I am not on drugs. I am a Mac OS X programmer, and know that the great stuff in Mac OS X is not the FreeBSD foundation (well, its not actually the foundation) but instead its the next derived components and engineering team.

If, and this is a big if, Mac OS X was so good because of BSD, then why do they use the Mach kernel? Why not the BSD kernel? If it was so good, why don't they use X11 instead of the NeXT engineered user-land that we are using right now? Why if it was so good don't we use C instead of the NeXT designed Obj-C/Cocoa? Why does Apple write 80% of their major apps in Cocoa if the BSD frameworks are so good?

Arguably, BSD could be removed from Mac OS X and conceivably we would not notice the difference. Yes we would lose the ability to do a lot of things, but they would and could be re-engineered. If we removed the NeXT derived components though, you would be back on a BSD command line without our nice Mach kernel powering you along.

NetInfo is still an integral part of Mac OS X and I have seen nor heard nothing about it being removed.

If Mac OS X is moving to a 64bit kernel, how are the iApp, Mac OS X, Quicktime, and other core teams testing and debugging Mac OS X right now on 32bit machines? Most of them have G4 towers, not G5's. If you mean that Mac OS X will have two kernels shipping with machines, one being 64bit optimized and the being 32bit optimized, then say that, but there is a 32bit kernel in Mac OS X and as far as I know it is remaining that way.

In fact, I would bet my house on the fact that we will have a 32bit kernel because you don't fork a kernel and still have the same number of employees working on it.

Not to mention of course that Tiger is basically 25% complete at this time. Anything can and will change in the time frame it takes them to release it.