Slashdot Mirror
Big Mac Officially Ranks 3rd
An anonymous reader noted that
according to Wired, it will be announced officially on Monday the Big Mac supercomputer is the third-fastest super-computer. The article also talks about some of the amazing supercomputers in the planning stages. The sort of stuff that will make Big Mac look like that old TI-85 collecting dust in your drawer.
Clicky for the official November list
Virginia Tech has not (and will not) call the computer "Big Mac". BBC used the name when it first started appearing in the news, and everyone else picked it up, IIRC.
The people in charge of the cluster don't want to call it "Big Mac" because (1) they don't want a lawsuit from McDonalds, and (2) who wants to be associated with nasty, greasy fast food?
They've worked out a solid candidate for a name (it's not official yet) that isn't quite as catchy as "Big Mac", but it also doesn't have any of the downsides.
Hmmmm:
PowerPC 970 G5 1800 rc572: 13,400,000.00
AMD Athlon64 1600 rc572: 5,771,251.00
Intel Pentium 4 3500 rc572 4,960,583.00
Hmm, guess this means my submission a couple hours ago won't go through (dangit, Wired!)...
Here is the official press release and the list.
There is a lot of good points to note all around. The first is the G5 Terascale cluster at Virginia Tech at #3 (10.28 Tflops/s, 2200 CPU, Infiniband) is the first academic computer to break 10 teraflops/s. This extra performance was promised at Mac OS X Developer's conference last month. Not to sure if the price is a testament to Infiniband ($1.5 million cabling, cards, and routers) or the Macs ($4.2 million list).
Good thing too because in a surprise move the NCSA cluster made the list at #4 (9.82Tflops/s, 2500 CPU, Myrinet). This cluster is built using Dell's running Pentium 4 XEONs and Red Hat Linux! One subtle point to note is that they didn't get all the systems online in time (there should be 2900 CPUs, not 2500). I bet some programmer at PSC and an ex-Chief Scientist of SDSC is appreciating having a hand in edging out NCSA for #3--not to mention Apple beating Dell for #3.
The fastest Itanium cluster is at #5 (8.63 TFlops/s, 1936 CPU, Quadrics) which is looking like the odd man out boxed in by a PC based systems using Myrinet, the P4 Xeon above, and the most powerful Opteron system at #6 (8.05 Tflops/s, 2816 CPU, Myrinet). Another point of similarity:did I mention it's also using Linux?
And finally, It's easy to overlook #73, a single compute node of BlueGene/L (1.44 Tflops/s, 1024 CPU). Imagine 128 of these connected together and you have something that will easily take #1 when it's completed even if we handicap it 20-40%. As noted on SlashDot earlier, this will be running Linux.
If you had read anything about this project, you would know that they have implemented their own low-overhead software error correction to compensate for the non-ECC memory. Presumably the benchmarks were done with this enabled.
It's hard to be religious when certain people are never incinerated by bolts of lightning.
The G5 is a cool processor, but it isn't the reason the VT cluster is so fast, the Infiniband interconnect is. The LINPACK benchmark that is used to determine position on the Top 500 list depends very strongly on the latency of the network connection.
Infiniband has ~ 8-12 us latency (probably even less by now), while ethernet is an order of magnitude slower. In real-life applications it's actually worse than this suggests.
We have tested a real-life application (socorro) using both gigabit ethernet and Myrinet (slightly slower than Infiniband), and gigE took 600 seconds to finish a run, while Myrinet took 4.
VT's cluster is using the largest Infiniband network yet built (or at least announced). The previous largest Infiniband network was O(100) machines. VT could have built the cluster using Xeons, Itaniums, or Opterons and arrived at roughly the same level of performance.
"...Then, as you have read it, you can tell the world how you do this "low-overhead software error correction to compensate for the non-ECC memory"..."
Taken directly from interview with Srinidhi Varadarajan:
Q: How do you deal with Error Correction in Memory?
A: There's a lot of traffic on Ars Technica and other places. We do failure recovery, memory doesn't report. One of the things we've noticed is that failures aren't an issue yet. The reason they can be competent is the LINPACK test, which is showing 16 digits of accuracy. We are planning on moving to ECC systems in the future. They may have to run things twice for a bit.
(tig)
Ignorance and prejudice and fear
Walk hand in hand
Actually, you are horribly horribly wrong about two things.
You definitely could not do that with Opteron or Xeon systems. VT was in negotiations about price and delivery time with Dell and Apple. Apple beat out Dell's prices (shocking!!!).
Also, the G5 makes a great cluster computer. It comes standard with gigabit ethernet and has very easy access to parts (no screws required to install anything).
Finally, the Apples make a good cluster because in 5 years or so when they disassemble it they have 1,100 really nice desktop machines. PC's need to be upgraded more often to serve as a desktop computer (that's why Macs have awesome resale value compared with PCs).
Help I'm a rock.
The 1.5 GHz Itanium 2 costs over $3000 per chip, and even the 32-bit Xeon 3.06 GHz is about $1000, while the 2 GHz PPC 970 is about $300 or $400.  In addition, VT wants 64-bit chips, so Xeon is a nonstarter.
Excluding the Earth Simulator, the 2 GHz G5 has the highest Flops per CPU, even 5% higher than the 1.5 GHz Itanium 2 and 10 times cheaper:
#2 Alpha 13880 / 8192 = 1.69
#3 G5 10280 / 2200 = 4.67
#4 Xeon 9819 / 2500 = 3.92
#5 Itanium 8633 / 1936 = 4.45
#6 Opetron 8051 / 2816 = 2.85
I would guess that your guess is wrong.
Have you seen a G5? And do you know why your typical fans fail? Could it be perhaps because they are undersized, underpowered, cheap fans being pushed to their limits to try and brute force enough air churning around to keep your system cool? The G5s have a properly engineered case with a specific path for airflow, and abundant, high-quality fans mounted on rubber shock absorbers to dampen vibration. I suspect that these fans will have few problems. And if they do, won't it be a painful process to replace them: unlatch side panel, remove plastic airflow enclosure panel, put hand on fan assembly, slide out, slide in new fan assembly. Click, done. I wish my servers were that easy to replace things on.
Cheap hard drives? Yeah, sure, whatever. They use Seagate Serial-ATA drives. I don't know about you, but Seagate stopped fitting into my definition of 'cheap and crappy' about 8 years ago.
I have never dealt with ECC memory, so I grant you that point, however.
Random and weird software I've written.