Please create an account to participate in the Slashdot moderation system


Forgot your password?
Technology (Apple) Businesses Apple Technology

Big Mac achieves around 14 TFlops with 128 Nodes 307

mzs writes "The Virginia Tech G5 cluster has achieved around 80% of its peak performance in preliminary Linpack testing with 128 nodes according to Jack Dongarra at the Top 500. "They're getting about 80 percent of the theoretical peak," Dongarra said. "If it holds, and it's unclear if it will, it has the potential to be the world's second most powerful machine." Typically getting 60% of peak in the Top 500 lists is quite good. If the Big Mac cluster achieves 60% of peak it would displace the 2,300 2.4 GHz Xeon cluster at LLNL for the number three spot on the current list."
This discussion has been archived. No new comments can be posted.

Big Mac achieves around 14 TFlops with 128 Nodes

Comments Filter:
  • Mmm Big mac (Score:2, Funny)

    by pheared ( 446683 )
    Anyone remember Happy Meal Ethernet and Big Mac Ethernet?
  • nope... (Score:3, Informative)

    by Durinia ( 72612 ) * on Thursday October 16, 2003 @02:26PM (#7231639)

    Title is wrong - they get 80% efficiency on 128 nodes. The 14 TFlops number is if that efficiency is held through the full size of the machine (2000+ processors).
    • It is amusing that the submitter apparently had not bothered to read the article.

    • Re:nope... (Score:3, Informative)

      by nate1138 ( 325593 )
      And that isn't very likely. The efficiency of large clusters drops with every node. Expect somewhere in the 60% range as a final efficiency. I think.

      • I was going to tack on something like "...which it won't." to the end of my post, but I thought it would get interpreted as an opinion as opposed to pointed out that the submitter was on crack. ;)

    • Re:nope... (Score:2, Insightful)

      I was going to say that something seemed terribly wrong... because when I read that I interpreted it as 128 nodes were pumping out 80% of it's peak at 14Tflops...

      if that were the case then the last 972 nodes would almost be a complete waste if they only gave a 20% preformance increase.

      good call on that Durinia!

  • by BWJones ( 18351 ) on Thursday October 16, 2003 @02:27PM (#7231648) Homepage Journal
    So, looking at this, I am wondering if the federal constraints on computer exports are still in place? This Apple supercluster shows that just about anyone now could afford to build a supercomputer giving smaller countries access to compute cycles never before dreamed of for relatively few $$'s

    • It's a good question actually. The answer is of course 'national security' and 'nuclear weapons' and stuff like that, but the real kicker is this:

      If a nation could not export the computers needed for a cluster, they could always build the cluster here in the US and utalize the thing over the Internet. In fact, they could probably contract a domestic company to do it for[THE REMAINDER OF THIS POST HAS BEEN CENSORED BY ECHELON]
  • G4, G5 (Score:3, Funny)

    by ( 156602 ) * on Thursday October 16, 2003 @02:27PM (#7231650) Homepage Journal
    When will Slashdot add or change the G4 icon to G5?
  • by grub ( 11606 )

    There are weenies that will say "Psstt.. you know that #2 computer in the Big 500? It only has one button on the mouse!"
  • What is it using for disk space? MCR @ LLNL is using the Lustre [] file system with DataDirect Networks [] storage.
    • Btw the main developer of Lustre is a really cool dude. I had some questions about snapshotting filesystems on linux, he had done some preliminary work on one similar to Netapp's for linux. I asked him about his work and if he knew about any similar work. He said his work was on hold indefinitly while he worked his paying gig on Lustre but that if he ever got Lustre complete enough that he didn't have to work 80+ hour weeks he would probably take it up again, he also pointed me to some solutions that he tho
  • With only 1100 macs to achieve the second rank of the TOP 500 computers for a measly 5.2 million... it makes one wonder why in the world the opteron or itanium's arent used in this type of cluster to achieve similar if not greater results?

    This is tremendous advertising for apple, but what about clusters of Power4's and 5's? why wouldnt they out-perform this cluster? at 14tflops... would 2200 macs be fairly equal to the earth simulator?

    And lastly with IBM seeing their G5's at 3Ghz in 8 months or so.... d

  • by GillBates0 ( 664202 ) on Thursday October 16, 2003 @02:30PM (#7231691) Homepage Journal
    Big Mac also achieved around 14 KTons with 128 kids.
  • ...If the Big Mac cluster achieves 60% of peak ...

    Big Mac cluster? What's that do, display in real time the flow of special sauce molecules over the surface of a flame broiled patty?
  • IBM has 49 of the top 100 from the last list. Who cares whose on top if you make almost half of the 100 fastest computers in the world?

    Yes, this is a troll post. I am pissed the submission after the Napster 2 one wasn't the itunes4Windows announcement.

  • Topic Icon (Score:3, Informative)

    by 47Ronin ( 39566 ) <> on Thursday October 16, 2003 @02:37PM (#7231777) Homepage
    Time to replace the G4 icon with a G5 pic dont you think??

    Like this for example > []
    • You're talking to the same people who have had an incorrect US flag since as long as I remember.

      Even *after* I sent a new one. Three times.

      I'm thinking that the G4 icon will be there when the G7 Virgin Tech cluster comes out.

      • You're talking to the same people who have had an incorrect US flag since as long as I remember.

        Hey - You're right! It looked a little odd - so I zoomed it up a bunch and sure enough: 12 stripes! I wonder which colony they dropped?

        (Reminds me of the time I was driving home from work and some guy had painted a US flag on the side windows of his Chevy Blazer. As a former teacher, it saddened me that this "patriot" figured we only had 36 states - six rows of six stars. Also, in his world, there were 15

      • Virgin Tech
        That's a pretty accurate description of the school, based on the people I know that go there :)
    • Well it's also time to replace the Caldera C logo with a SCO/SCOX tree on a yellow background, but they don't do that either. Caldera has nothing to do with the SCO case having been bought and changed in nature before the case happened.
  • Break into the labs and use this baby to run some RC5 cracks! My team will be unstoppable!
  • Great feat - IBM!

    Why is everyone so obsesses with Apple when it comes to G5? What does Apple have to do with the G5?! This is the kiddie version of IBM's success CPU named Power4, if I'm not totally incorrect.

    Apple, to me, is a group of cosmetologist hangarounds. ;)
  • The submitter stated 128 nodes. This is wrong. The article states 128 processors which would be 64 nodes.
  • Big Mac achieves around 14 TFlops...

    All this time the answer to my supercomputing needs was at McDonalds. What a burger!

  • It should be noted that the final result doesn't need to be anywhere near 80% efficiency to assure it second place on the list. The third place machine is only around 7 TFlops, meaning Big Mac could operate at as low as 40 to 45 percent efficiency and still take second.
  • I wonder if they are using GCC3.3 or IBM XL C. Apparently the performance from code generated by the IBM compiler far exceeds that generated by the latest PPC GCC.
    • And what OS are they using? I haven't read about that anywhere but I would venture to guess it's some commodity os (linux) they don't need to mention. And yes at 80% efficiency they have to use altivec optimalization .. if it's gcc hopefully we'll see better altivec support sooner rather than later.
  • Premature (Score:3, Insightful)

    by jmichaelg ( 148257 ) on Thursday October 16, 2003 @02:51PM (#7231924) Journal
    The article says
    "We're just making up numbers here," Dongarra cautioned. "We don't have real numbers yet. If they get 80 percent, it will be slightly faster than (ASCI Q, the current No. 2 on the Top 500 list)."
    Then a little later, the article says:
    Lockhart cautioned that even if Big Mac beat most of the machines in the current Top 10, the list, which is compiled twice a year, is a moving target. Lockhart said there are four or five new supercomputers coming online that also may qualify for places in the Top 10.
    So to summarize, the data aren't in and nobody will really know where the machine ranks until all the data are in. About the only outcome one would expect is that the machine would outperform older technology.
  • Of course, it might get out of hand....everyone knoes it takes two hands to handle the whopper!
  • Anyone know the theoretical peak of the 2300-node xeon cluster at the Lawrence Livermore lab? I've read elsewhere [] that it was (to be) a 1920-node cluster with a theoretical peak of 9.2 teraflops.

    This article reports that the 2300-noder operates at 7.6 teraflops, but i was wondering what percentage of its theoretical peak that is.
    • According to the June Top500 List [], LLNL's third-ranked MCR Xeon cluster has an Rmax of 7634.00, and an Rpeak of 11060.00.

      Rmax is the best it's attained; Rpeak is the theoretical peak in a perfect world with 100% multiprocessing efficiency.

      One of the really amazing things about the NEC Earth Simulator (aside from its sheer power) is that its Rmax is over 85% of its Rpeak. A lot of other systems only get Rmax of 60-70% of Rpeak.

      I've read that PowerPC chips are pretty efficient in SMP scenarios, and if

  • SETI array (Score:2, Insightful)

    by darc ( 532156 )
    Interestingly, a fun number to compare against is SETI@Home's array power, which is approximately 15 teraflops. [See the SETI@Home FAQ]

    Although they don't run Linpack, and therefore can't be considered on Top500 the same way, it's still cool to know that SETI would still place second on the supercomputing list. Back in 2001, they were averaging a very large number of teraflops as well, (>10TF) the figure is on the internet somewhere. In 2001, that was greater than the top three supercomputing sites comb
  • Pretty cheap (Score:3, Insightful)

    by randombit ( 87792 ) on Thursday October 16, 2003 @02:58PM (#7231992) Homepage
    From article: Dongarra said the cost is so low he questioned whether the college got a special discount.

    At $5.2 mil for 1100 machines, I think they paid full market price; that's over $4,500 per machine, and currently Apple is selling dual 2 Ghz G5's for ~$3000. And that's with lots of extras that they wouldn't want in a cluster (ATI 9600, CDRW, etc), which hopefully they convinced Apple they didn't need... (else they've got a whole lot of Mac keyboards sitting around!)

    I wonder how much of the cost was the actual machines, and how much was infrastructure and networking stuff (I can just see 1,100 Macs all powered off one extension cord and a bunch of surge protectors).
    • Well, 4GB mem per machine and NETWORK cost. Infiniband cards cost more than 100$, and switching architecture with 1100 ports isnt cheap, either...
    • And that's with lots of extras that they wouldn't want in a cluster (ATI 9600, CDRW, etc), which hopefully they convinced Apple they didn't need...

      They plan to use the GPU on the 9600 in the future for extra compute cycles. It's a cheap coprocessor.
  • What will the effect of the 'magical' software they have to detect memory errors be?
  • the article suggests that maybe va tech got a big discount from apple because of the huge cost difference from other sooperputers so I went to the apple online store.

    I pulled the modem out of the default dual 2ghz g5 mac and upped the ram to 4gigs, as I vaguely remember the big mac node config being from some other article somewhere. Cost 5,320. Tried to up the quatity to 1100 in my cart but the web form would only allow 3 digits, so I did 110. Got 563,200. That looks like a volume discout because my
    • I did the store thing twice to get my numbers. Second time I for got to replace the superdrive with the combo. That accounts for the automatic discount. Unit price is 5,120 on that config * 110 is 563200 which is spit out in the cart.

      Sorry folks. I'll sheepishly retreat into the corner now
  • by crovira ( 10242 ) on Thursday October 16, 2003 @03:09PM (#7232161) Homepage
    I'm picking nits here but...
  • VA Tech must've gotten about 1.6 TFlop/s on 128 nodes. 14 TFlop/s is waaaaaay in excess of the peak for 128 dual G5 nodes (by almost an order of magnitude in fact).

    Here's how I figure that:

    A G5 proc can do 4 64-bit FP ops per clock cycle (2 FPUs with each capable of doing a multiply/add op), so that's 8 GFlops/s per 2GHz proc, or 16 GFlop/s per dual-proc node. For 128 nodes, that's a little over 2.048 TFlop/s peak. Dongerra said they were getting 80% of peak, which would be 1.638 TFlop/s.

    The thing

    • I also want to know how much floor space this thing takes up. with 12 nodes per rack, you're looking at 92 racks to cover that many machines.. the picture they posted shows 15+ racks in a row, say they have 22 racks per row, which would give us room for IB switches, gige switches, and disk servers. that's 256 nodes per row.. you're looking at maybe 4.5 rows of 22 24" racks.. the G5 being an 8" wide box. with some extra room for cable managment, and rack structure, you're probably talking 28" wide for ea
  • Ok you frothing Macaholics. *Someone* explain to me *how* such a largely clustered machine was put together with 80% efficiency *WITHOUT* ECC. This issue has been posted several times on this story, without any satisfactory answer.

    If there are too many nodes the error rate goes up super-linearly (since interconnect errors have to be factored in.) There's no getting around this. If they are just running *without* ECC and hoping the results just come out correctly, then this is worthless -- its amounts t
    • AFAIK ECC isn't really that important. when a machine with ECC detects an error it shuts down, when a machine without ECC has an error it usually crashes. My guess is that the machines get power-cycled every now-and-then, and somewhere in the init the machine checks the RAM. RAM errors are pretty far-between on Macs when using Apple-branded RAM.

      ECC is overrated.
      • ECC detects an error it shuts down,.

        Not necessarily. Know what ECC stands for? One of the C's is for Correction. Given enough extra bits, you can arrange the coding so that valid memory words all differ from one another by two or more bit flips. Given a single bit flip, and assuming it was only a single bit flip, you can uniquely determine the corresponding valid memory word. If the hardware is functioning properly (e.g., it was just a cosmic ray hitting a RAM cell), then you can rewrite the valid combinat
    • 1. Take a deep breath.

      2. Step away from you computer and walk out side for a bit. Try not to be scared by all the blue sky and that really bright yellow ball up there.

      3. Come back in after sudden exposure to intense light causes immediate sunburn.

      4. Read about Deja Vu here [].
  • Wow (Score:3, Funny)

    by Kelz ( 611260 ) on Thursday October 16, 2003 @03:56PM (#7232828)
    Had no idea hamburgers could run that fast.
    Must be the "special" bun in the middle.
  • My double whopper is going to beat that!
  • Does anyone know a site that lists the floating point values of common processors like the Pentium 4 and Athlon series? I would like to calculate the power of therotical clusters but I can't find the Gigaflop/Teraflop information anywhere. Anyone?
  • Big Mac achieves around 14 TFlops with 128 Nodes [...] If the Big Mac cluster achieves 60% of peak it would displace the 2,300 2.4 GHz Xeon cluster at LLNL for the number three spot on the current list.

    Uh, huh. So, if those statements were true, the G5 would be 12 times faster than a Xeon. Quite a feat of engineering that would be. But I don't think so. The folks at VT are extrapolating their 64 node performance to a 2300 node cluster.

    This tells us two things:
    • Mac advocates have stooped to new lows
  • AIST is getting a similar scale Linux cluster using Opteron processors, a bit larger (are all the G5 dual proc boxes? If so things are still in the same ball park). about 1100 dual opteron systems, and a few other Intel boxes tossed in. One good article is here [].

    I have submitted this as story, but it evidently never news like the G5 cluster... Linux and AMD no longer our favorite 'underdogs' anymore?
  • iTunes has problems. If you read the forums, people are reporting that iTunes REARRANGES your existing MP3 files into it's own directory scheme. Is that rude or what? Granted there is an option to turn it off, but what a bad default setting if you don't catch it!

    For me, .99 a song is still too much money. But maybe I could live with that.

    But it also won't download or let you create MP3 files from music you download from them. I've got a car player with a 40GB hard drive that only supports MP3

Adding features does not necessarily increase functionality -- it just makes the manuals thicker.