Top 500 Supercomputer Ranking Gets Top Heavy.
With the Supercomputing 2003 trade show in Phoenix, Arizona, this week, Hans Meuer of the University of Mannheim, Erich Strohmaier and Horst Simon of Lawrence Berkeley National Laboratory, and Linpack benchmark creator Jack Dongarra from the University of Tennessee have released their semi-annual ranking of the Top 500 supercomputers.
The Top 500 is a rough indicator of vendor and technology preferences in the high performance computing (HPC) server market.
According to the ranking, the Earth Simulator parallel vector supercomputer, made by NEC Corp for the Japanese government and academic institutions to perform climate modeling, is still the most powerful machine in the world, with a performance rating of 35.7 teraflops of aggregate raw number-crunching power based on the Linpack Fortran benchmark.
That's maximum delivered performance, including overhead, not maximum theoretical performance, which is just the summation of all the flops in each processor. The Earth Simulator consists of 5,120 vector processors running at a mere 500 MHz.
The number two system is the "ASCI Q" machine based on Hewlett-Packard Co's AlphaServer SC45 servers. This cluster, located at Los Alamos National Laboratory, has 8,192 processors, running at 1.25 GHz and delivering 13.9 teraflops. This machine has been on the list for quite some time, but it looks like big Linux clusters are getting ready to bypass it.
Number three on the list is a new machine, dubbed "X," built by Virginia Tech from 1,100 Apple G5 workstations with 2 GHz processors and delivering 10.3 teraflops of power. This Apple cluster is linked together with Infiniband switched fabric made by Mellanox. This machine makes a very strong case for IBM's 64-bit PowerPC 970 processor, which is at the heart of the G5 workstations. This is only the third machine to every deliver more than 10 teraflops of computing power (real, not peak) on the Linpack test.
The fourth largest super on the list, nicknamed Tungsten, is installed at the National Center for Supercomputing Applications (NCSA), is comprised of two-way Dell Corp PowerEdge 1750 servers with a total of 2,500 3.06 GHz Pentium 4 Xeon DP processors from Intel and rated at 9.8 teraflops.
A variety of Linux clusters based on 32-bit Xeon and 64-bit Itanium processors and 64-bit Opteron processors from Advanced Micro Devices Inc are increasingly dominating the Top 500 list as these processors have reached parity with RISC/Unix architectures and clever third parties have been able to deliver interconnection fabrics that rival the best technology that IBM, HP, Cray, and SGI can put into the field using more established HPC architectures.
The fall Top 500 list has 208 clustered systems made from workstations or entry servers, up from 149 systems in the list that was published in June 2003. With this list, clustered machines are now the most popular type of supercomputer in the Top 500.
On this fall's list, the aggregate computing power of all 500 machines, across all architectures, was 528 teraflops, up from 375 teraflops only six months ago. The least-powerful machine on the Top 500 list is now rated at 403.4 gigaflops, while in June, that level was a much lower 245.1 gigaflops. The increasing use of Linux clusters by government agencies, research labs, and commercial institutions is driving out smaller machines.
Machines based on Intel Corp chips account for 189 supercomputers, with 32 of them based on Itanium and 157 of them based on Pentium 4 Xeon processors. There are a handful of AMD machines on the list, too. In the June 2002 Top 500 ranking, Intel had only three systems in the list, and then jumped to 59 systems in December 2002. In June 2003, Intel-based machines accounted for 119 systems, with 19 of them based on Itanium and 100 of them based on Pentium 4 Xeon.
The commodity pricing of Intel x86 servers has appealed to HPC customers, but it is the advent of commodity interconnection fabrics that is really driving the adoption of x86 machines in the supercomputer market. Linux is helping, too. The combination of x86 architecture, cheap interconnect and Linux is an unbeatable combination for a lot of HPC workloads, especially embarrassingly parallel applications like climate modeling and forecasting, genomics and proteomics, product test and design, and other kinds of simulation.
Still, HP's PA-RISC and IBM Corp's Power architecture systems are still popular on the list, and that is why these two vendors are fighting for top share of the Top 500. IBM, with 35.4 percent of the aggregate computing capacity of the Top 500 list, beats out HP's 22.7 percent share and NEC's 8.7 percent share. By system count, HP is the leader, with 165 systems compared to IBM's 159 systems. SGI is still in there, with 41 systems and its biggest machines based on the Altix 3000 shared memory Linux supers.
There are a large number of IBM PowerParallel, HP Superdome, Sun Starfire/Sun Fire, and SGI Origin machines on the list, as well as a ten Cray X1 vector supers. Cray and IBM will be vying for the top spots as they deliver their 100 teraflops Red Storm and ASCI Purple parallel clusters by early 2005, and IBM's Blue Gene/L super, which will be rated at about 360 teraflops, will very likely hold the top spot for as long as Earth Simulator seems to be.
|Printer friendly Cite/link Email Feedback|
|Date:||Nov 18, 2003|
|Previous Article:||Eclipse Swings on Java Answer.|
|Next Article:||European Services Spend to Grow by 65%.|