Virginia Tech to Build Top 5 Supercomputer?

hype7 writes "ThinkSecret is running a story which might explain exactly why the Dual 2GHz G5 machines have been delayed to the customers that ordered them minutes after the keynote was delivered. Apparently, Virginia Tech has plans to build a G5 cluster of 1100 units. If it manages to complete the cluster before the cut-off date, it will score a Top 5 rank in the Linpack Top 500 Supercomputer List. Both Apple and the University are playing mum on the issue, but there's talk of it all over the campus."

Do they have a need for it?

[John Seminal]

I might be wrong, but I think VT is a public school. Why do they need the 5th most powerful cluster in the world? How about lowering tuition? Does the school really not need anything else more than a cluster? In 5 years, will it still be worth the money they are paying? Is there an immediate need to the cluster?

Parent is stupid

[Anonymous Coward]

You obviously know very little about distributed computing.

Latency isn't the most paramount issue, otherwise render farms and clusters wouldn't be as popular as they were today.

and lets not forget about projects like distributed.net and Seti@Home. Latency is not at all the concern for them.

Re:Macs ?

[damiam]

For certain types of processing (rc5 cracking is one example), Macs completely smoke PCs. For example, distributed.net stats show that a 667Mhz G4 can process more keys/second than a 2.8Ghz P4. Considering how much faster a 2Ghz G5 would be, a 1100-node cluster would be damn powerful if you were doing work that mapped well onto Altivec.

Getting a bit ahead of ourselves

[Richard Mills]

"If it manages to complete the cluster before the cut-off date, it will score a Top 5 rank in the Linpack Top 500 Supercomputer List."

Err... I think somebody's getting a bit ahead of themselves here. =) Building parallel computing systems is complicated, and it may end up being quite a bit harder to realize the predicted performance than thought (not an uncommon occurrence). I'll believe it when they have the actual Linpack numbers.

Re:What?

[Anonymous Coward]

"I don't recall the last time they were wrong about anything they've posted"

iWorks.

Hoax

[Anonymous Coward]

The cluster has nothing to do with Apple. They were contemplating using PPC970 CPUs (basically becasue they're cheaper than the Power4's), but AFAIK there were never plans to use complete G5 computers, or anything else made by Apple. And I doubt IBM would take CPUs from Apple to give them to V.Tech. The delay is mot likely due to performance issues; they're probably fine-tuning the OS so it's competitive with x86 / Windows systems.

Re:What about latency?

[Glonoinha]

-You win the moron of the article award. Congrats.

Now you are one optimistic AC. The day is still young. I am giving 30:1 odds that there are going to be way better morons than Thinkit3 before this thread is archived.

Re:What?

[Anonymous Coward]

Frankly, I don't recall the last time they were wrong about anything they've posted.

How about this report? [thinksecret.com] I'm still waiting, nearly a fortnight later, after they said "one Apple retail source said the new 15-inch and 17-inch models are expected to arrive later this week".

Re:What about latency?

[Aapje]

It all depends on what you need the cluster for. Some computations need constant communication, others can go on for hours, days or even weeks without feedback. If you're smart, you use supercomputers for the first kind of tasks and clusters for the second kind.

Universities (and big business) often work together and exchange resources. Virginia Tech gets a large amount of bargaining power by having control over a large amount of processing power. They can easily trade CPU time on their cluster for CPU time on a low-latency supercomputer.

Re:AltiVec

[Anonymous Coward]

SSE2 and Altivec are not identical, so there's still areas where a G4/G5 would (and does) thrash an intel-anything to pieces - the same can happen with other areas where the Pentium would kill the PPC.

One example is in RC5-72 crunching. My plain old consumer eMac can crunch 10.47 million keys per second using altivec. It's not even an optimal architecture for getting all the data through to the altivec unit (133mhz bus etc). The fastest x86 based box doesn't come -near- that. (see http://n0cgi.distributed.net/speed/ for stats)

That's one extreme case, and possible the most extreme one. Only knowing the final purpose for the VT cluster would reveal just how sensible picking PPC machines is.

Re:Macs ?

[Llywelyn]

True, but it is entirely possible that they are building this for more than just to say "we have a computer on the top list and you all sux0r5". They are depending on the dual-floating point units on the G5 for LINPACK and AltiVec for whatever else they put it to.

Without more details its hard to tell

Re:is that so?

[carpe_noctem]

Wait a sec...you mean the thinksecret story, or the macrumors one? I think "real" is a bit of an overstatement here...

Re:AltiVec

[discstickers]

The SSE2 unit on the P4 or the Opteron would have nearly the same performance and cost a whole heck of a lot less.

Uh, no. 2 years ago, my roommate and I were both running the distributed.net client. I have a 500 Mhz Powerbook G4 (100Mhz bus). He had a 1.4GHz P4 with rambus RAM. I got 4Million keys/sec. He got 2MKeys/sec.

So clock for clock, my machine was nearly 4 times faster.

Re:MOD PARENT DOWN

[class_A]

ooops, yuo = own3d!!!!
http://n0cgi.distributed.net/speed/query.php?cputy pe=99&cpumhz=1000&recordid=1&contest=rc572&multi=0 [distributed.net]
Power PC 7450/7455 G4 1000 MacOS X 10.2 2.9005 RC5-72 10,594,666.00

Re:1100 G5s still can't...

[Anonymous Coward]

No computer available can really play Doom3. It's not fucking out yet you stupid fucking cunt. Maybe I missed the announcement, that the game was cancelled for the mac. But considering that it was first demoed on the platform, that every single id game has been ported and that you are a fucking moron, I'd say that 1100 G5s can play 1100 copies of Doom perfectly well. Provided they are all equipped with in- and output devices, which is rather unlikely. So I guess you're right, Mr. (or Mrs. or maybe even Ms.) shitface. 1100 can't, but a single one will be able to. In closing I would invite you to remove your head from you stinking rectum and replace it with your dick. If you are female I apologise for calling you cunt, call you wanker instead and invite you to stick a fist up your privates. Thank you very much.

Re:Poor choice on Apple's part

[runenfool]

You got a link for dual Opterons for 1500 bucks with all the goodies in the G5? If yes, Id love to see it.

(seriously - I can take out a loan :) )

1% of G5 orders

[Viadd]

Just for perspective, there have been over 100,000 G5s ordered, so this cluster is about one percent of the backlog. In other words, assuming that Apple ships all pending orders in about a month, the G5 I ordered will be delayed by about 8 hours.

And 8 hours@12.4GFlops...damn you Virginia Tech, you owe me a third of a quadrilion floating point multiplies!

Re:Poor choice on Apple's part

[benh57]

And that opteron system will use 3x the power and heat. Universities don't spend millions on a 'non efficient choice'. Smart people did the math, and they had facts - which you don't.

Re:What about latency?

[stingerman101]

Doesn't it also depend on the interconnect technology they decide on? Since the G5 architecture uses hypertransport for the PCI-X connection and switches the communications to the processor, we are not talking about the same level of latency that older generation PC's had, are we? Isn't it premature to draw conclusions until we better understand how Apple and Virginia Tech plan to architect this new super type computer cluster?

Re:AltiVec

[sevenofnine]

Yeah right, so what your saying is that they should build their own cluster by putting together 1100 units of grey boxes?
In your "calculations" i think you forgot to add the added manhours for doing this...
and thats not even thinking about the order / failure rate...
buy mem for 1100 machines is bound to have 1, 2 or more (properbly the ladder) mem blocks not working...
And im sure there is alot more reasons to this that you ( and i) missed...

Re:Another brilliant idea, inspired by Apple

[dbirchall]

With 1,100 machines in the cluster, you'll probably be running into something's mean time between failures pretty darn often, whether it's memory getting a bit wrong, or one of the 9,900(!) fans needing replacing. :)

But... a cluster should be redundant enough to withstand that sort of minor inconvenience and go on functioning without the errant node while it gets fixed, reboots or whatever.

I'll admit that building something smart enough to say "Node 206, you have a memory error. Bad G5, no donut!" is beyond the scope of my understanding.

Re:Think Secret's Record

[Nick dePlume]

I'd argue that it's solid in that regard as well.

I believe we're honest and up-front about the reliability of what we report. If something has not been confirmed by multiple reliable sources, we say so. If something is mere rumor, we also say so, but most of the time, we would rather not report it if we can't confirm it.

As for mistakes, after Macworld Expo, for example, we typically take a look at the announcements and compare them to our pre-expo reporting to see what, if anything, we got wrong. We don't hide anything.

Nick dePlume
Publisher and Editor in Chief
Think Secret

But, what will they *do* with it?

[chiph]

Sure, having a top-5 supercomputer is cool, and the bragging rights will garner the school some additional funding and scholarships, but at the end of the day after the benchmarks are run and the empty Jolt(tm) cans are recycled, what will they *do* with it?

Chip H.

Re:Talk about a ton of desktops in a server room

[valdis]

Exposure to the facilities? All I'll say there is that I know of at least 4 people who have posted to this story that could throw a paper airplane from their cubicle and have a reasonable chance of landing it on the blueprints. ;)

I said there wasn't capacity on that generator. I didn't say anything about the existence of other generators - and there's a distinction between pulling copper to get power grid capacity to the cluster and having emergency power backup for same. That diesel is for emergency backup, not local power generation - we've got a nice coal-fired beast on campus for that.

In addition, you have to remember that the cluster most likely has different backup power requirements than our production systems. Our production systems (the central mail server, the large database machines, etc) really need to stay up - we're talking about things that if the power substation near the airport goes out, the whole campus is screwed till it comes back online, if we don't have backup power.

Look at how many services were out when we did the cut-over - everything was affected, from E-mail to the library catalog. That's what emergency power is for - so you don't have an event like that without the multi-week notice that the install had.

On the other hand, does a compute cluster really need that level of power backup, or can we significantly trim the budget by merely having a really good power conditioner to make sure there's nice clean non-ripply power, and enough battery backup to allow the cluster to do a suspend-to-disk and poweroff cleanly? Yes, if we take a power hit in that scenario, some researchers have to wait for their results - but the business of the university as a whole isn't disrupted.

Or possibly the plan is to have the conditioner and batteries for now, and have someplace to cable in a backup diesel generator in the future - remember that 1,100 G5s, a bunch of Infiniband switches are a chunk of change - and then you get the expense of the power/cooling work. Suddenly you start thinking about how to put the non-essential costs into the next fiscal year. ;)

(The above is not intended as an actual statement of the actual design or plans, just an illustrative discussion of the fact that there are a lot of tradeoffs that need to be made when planning and designing large-scale installations of any sort).