Slashdot Mirror
New Top500 List Released at Supercomputing '06
Guybrush_T writes "Today the 27th Edition of the Top 500 List of World's Fastest Supercomputers was released at ISC 2006. IBM BlueGene/L remains the world fastest computer with 280.6 TFlop/s. No new US system in the top10 this year, since they all come from Europe and Japan. The French Cluster at CEA (French NNSA equivalent) is number 5 with 42.9 TFlop/s. The Earth simulator (no 10) is no longer the largest system in Japan since the GSIC Center built a 38.2 TFlop/s Cluster, reaching the 7th place. The German cluster at Juelich is number 8 with 37.3 TFlop/s. The full list, and the previous 26 lists, are available on the Top500.org site."
= FLoating Operations Per per second
Should be just Flops with no /
I would find that suprising too, except that there are 3 Microsoft systems on the list (5 OSX systems too). I actually read the page though.
Actually, as I posted above, there is a new system in the US. It's currently the #3 system.
http://www.top500.org/system/8128
Just to follow up, you can get OS information here: http://www.top500.org/stats/27/osfam/ (by family)
... I can only assume that the systems using "Linux" as the particular operating system are using a custom-made distro, instead of one of the commercial ones (which are listed separately on the detailed chart). Unless they just failed to report one in particular.
OS (# systems) (Percent)
Linux 367 73.40%
Windows 2 0.40%
Unix 98 19.60%
BSD 4 0.80%
Mixed 24 4.80%
Mac OS 5 1.00%
Totals 500 100%
Alternately there's a more refined breakdown listing them by Operating System type and version. Oddly, "Linux" is listed both as an operating system family and as a distinct flavor/distro
As for the Windows-based systems, there were one each for Windows 2003 Server and Windows Compute Cluster Server 2003.
"Ladies and gentlemen, my killbot features Lotus Notes and a machine gun. It is the finest available."
Correction, MMX added SIMD-type functionality to the original pentium line, but MMX only worked on integers, and reused floating point registers, making the proc incapable of doing FP math and SIMD at the same time..
SSE was when intel "got it right".
And it's still not that commonly used a feature.
I don't need no instructions to know how to rock!!!!
With over 900,000 computers in the system, SETI@home has the ability to compute over 250 TFLOPS (as of April 17, 2006).
http://en.wikipedia.org/wiki/Seti@home
IBM's Blue Gene is faster (and more flexible). Big networks of computers like SETI's are good at crunching static radio telescope data or brute force RC5 cracking. When it comes to most real world problems, the nodes must communicate and share data, which over the internet makes it far too slow. Real supercomputers do not use any type of networking between nodes, they have a shared memory bus.
Cthulhu Saves.
Because quite a few of them do use liquid cooling, but they don't order hobbyist solutions. Many Crays still use 3M Fluorinert, which doesn't damage sensitive electronics when you spray it right on a processor.
-Those who would give up essential liberty to purchase temporary safety deserve neither. -Ben Franklin
Rmax and Nmax are two measures of the capability of a supercomputer.
The first, Rmax, is the LINPACK benchmark. The LINPACK benchmark is a measure of floating point operations per second for the cluster. They usually include theoritical ((Max FLOPs for one CPU) * (number of CPUs) == Rpeak) along with the actual. Obviously, theoretical values will always be larger then actual due to wasted CPU cycles.
Nmax is the size of the problem (i.e. the dimension of the solved linear equation)
So, Rmax is the maximum performance obtained by the LINPACK benchmark during the test, Nmax is the size of the problem the best performance was achieved at.
~nate
Rmax represents the maximum acheived measured FLOPs as a result of an xhpl run.
Nmax represents the problem size. Nmax generally is aimed to be a problem that consumes as much memory as possible without swapping.
Rpeak is the theoretical max FLOPs possible according to the processor used. For example, a PPC chip is theoretically capable of 4 Flops per clock, so multiply the clock by the number of cores in the cluster. x86_64 is theoretically capable of 2 flops per clock, so multiply cores by two. Note that AMD clock for clock doesn't do any better than intel in *this* particular benchmark, so Intel clusters inherently can climb this list better, despite poor memory performance and other factors that make them less useful in a general supercomputing sense. Itanium can acheive better floating point (I believe 8 flops per clock).
And for anyone seeking to compare Rpeak/Rmax numbers with published Cell figures, keep in mind that game consoles (and by extension cell) brag about their single precision (32-bit) floating point performance, whereas this list only deals with double precision numbers (64 bit). Cell actually is nothing special at get top500 relevant benchmark results.
Many people feel this very specific benchmark is a poor indicator of the overall effectiveness of a cluster, and consider hpcc (which includes hpl as a subset) to be a better holistic method to evaluate the value of a cluster.
XML is like violence. If it doesn't solve the problem, use more.
The speed doubling time is still about 18 months (== 10x in five years). Two more doublings from the 2005 or 2006 280 TFlops is around 2008-2009. Its a version of Moore's law for supercomputing. Though processor speed hasnt been gaining as fast in recent years, improve clustering technology and software seems to be compensating.
"Exaflops in 2020!"
United Airline's intranet is called SkyNet.
"No one likes working in a hamster wheel, and your shop smells of cedar shavings from here." - TaleSpinner
shared memory bus is a pretty subjective measure. Sine the mid-90's no supercomputers have used true flat-memory shared buses. Instead they are connected by some sort of switched point-to-point network. In many of the mpp machines, these networks are built into the memory controllers on each node. (blue gene, crays, earth simulator, etc) The bandwidth and latency of these networks is orders of magnitude better than gigE, but it's still a network of sorts.
Given that clusters these days are made from commodity components (Xserve G5s, for instance) and how large clusters are these days, you end up with a pretty astounding failure rate. We lose roughly two peices of hardware (in order of most to least common: memory, cpu, motherboard, disk, power supply) a week, and our cluster (Turing, 640 nodes) is fairly small. We aren't even into the late-in-life crazy-disk-failure mode that most machines get at 3-5 years old. Think about the logistical nightmare if we had to try to "drain" a system of coolant before pulling it out to service it.
Plus then you'd have to have all that (very custom) cooling equipment, pumps, etc. You'd have to watch for leaks closely, which is also a problem with air cooling and the refrigerant lines, but those have a lot less surface area of pipe/connectors to go wrong: a loop per rack for rack-mounted cooling, not a loop per machine.
Plus, as other posters have said: we'd like accurate numbers.
Slashdot Patriotism: We Support our Dupes!
Real supercomputers do not use any type of networking between nodes, they have a shared memory bus.
Don't say that, all it does is show how little you know.
There are 2 common types of interfaces between nodes on a supercomputer: shared memory and message passing.
Shared memory is where all the nodes can access memory over some sort of network. In order for communication to happen all 2 nodes need to do is read and write to the same location in memory. There is little talk about the network protocol used at this level because for the most part it is an emulation of layer 2 of the OSI (as if all you are doing is ordering the hardware around).
Message passing would best be described in terms of layer 7. Communication occurs between 2 nodes via messages that are sent back and forth (hence the name). The most common message passing scheme is MPI. In MPI, there is a concept of a sender and a reciever. The reciever calls MPI_Recv and the sender calls MPI_Send and a message is sent from send to recv. You could almost think of this as an HTTP communication; the server is listening, the client sends information, the server sends back, except in MPI the reciever must be calling MPI_Recv and waiting for a send from a specific sender and the sender must call MPI_Send to send the information to the reciever (there really [well, sorta] isn't a concept of a timeout). In my experience, this makes MPI (I use MPICH2) difficult to debug, if A calls send to B and B calls send to A at the same time, your program blows up (often with very little useful information).
On the cluster I do my work on, the implementation of MPI sends TCP/IP packets over ethernet (much like 256 of this top500 list). The libraries could be written to do the work over Myrinet or any other network.
For future reference please learn some factual information before you go spouting bull. If you follow this link, and choose interconnect family, you would find that most of the supercomputers in the top500 list are using some standard network interconnect.
Yeah, but is the NSA sitting on a supercomputer as a "last resort" for encryption problems?
Not supercomputers in the sense they are being discussed here. The top500 list is computers that excel at floating-point operations. I have never seen an encryption method which uses floating point at all. They all use integer operations. DES, RSA, AES, MD5, SHA-1, etc. All 100% integer. In most cases cracking encryption algorithms really boils down to some sort of a search algorithm, so it wouldn't suprise me if there is overlap between systems that excel at data-mining and ones that excel at cracking encryption.