Back in 1993, Prof. Hans Werner Meuer and Erich Strohmaier launched the TOP500 project as a way for the industry to track the most powerful computers in the world. The resulting list, which is released twice a year, ranks the top computers based on the well-known Linpack benchmark. To most people, those 500 systems represent the definitive list for the absolute fastest computers in the world. Because of that, getting a spot on the TOP500 is eagerly sought by system vendors, component makers, and science/research organizations.
But the TOP500 project encompasses something more valuable than just bragging rights. The 30 lists that have been compiled since June 1993 represent a history of the supercomputing industry from the perspective of the machines themselves. Recently Prof. Meuer composed a retrospective of the 15-year record of the TOP500, in which he reviewed the project’s history (and prehistory), the rationale for the rankings, the limitations of Linpack and the list, the evolution of supercomputing technologies, and how the list can be used to forecast the capabilities of future machines.
One of the most useful aspects of the list is how accurately it does predict the advance of computing power. Over the last 15 years, the top and bottom systems on the list as well as the aggregate FLOPS of all 500 machines have followed a fairly constant exponential progression. Writes Meuer:
“Based on the extrapolation of these fits, I announced at ISC’97 that, eight years from then, i.e. in 2005, there would be only Teraflop/s systems on the TOP500 list, even though the 1997 list had only one: Intel’s ASCI Red. Probably many of the ISC’97 attendees thought I was crazy to make a forecast eight years out for such a dynamic market. And I do have to admit that I did not feel too comfortable myself — going out on a limb like that in front of many of the world’s top HPC experts. You certainly can imagine that I was extremely anxious to see the next TOP500 lists.”
But true to his prediction, by June 2005, there were only teraflop systems on the TOP500 list. Future projections include the first petaflop (Linpack) system on the list in November 2008 (certainly conceivable), a one petaflop entry point to the list in June 2016, and the first Linpack exaflop system by November 2018.
One of the other trends Meuer traces is the how international representation has changed over time. In the very first list, the dominant supercomputer players were the U.S. with 225 systems (45 percent), Japan with 111 systems (22.2 percent) and Germany with 59 systems (11.8 percent). Every other country was in the single digits percentage-wise.
Today, although the top five systems are spread through four countries — the U.S. Germany, India and Sweden — the dominance of the U.S. in the overall list has actually grown to 283 systems (56.6 percent). No other country even tops 10 percent. Overall though, the trend points toward democratization of supercomputing, with more countries represented in the list, and rapidly developing economies, like China (10 systems) and India (9 systems), already making their presence felt. The continued globalization of computing seems assured, but the open question is whether U.S. dominance of the list will continue.
There is no such democratization trend for HPC system manufacturers. In the early days of supercomputing, Cray Research, Fujitsu, Thinking Machines, Intel, Convex and NEC all owned at least 5 percent of the TOP500 market, based on number of systems. Fifteen years ago, Cray Research dominated the pack with 205 systems (41 percent). Those were the days when customized systems were the norm.
With the rise of HPC commoditization (x86 processors, Ethernet, Linux) over the last decade, the market has completely transformed itself; supercomputing has become much more general purpose, which has ushered in a new set of vendors. In November 2007, IBM led the TOP500 with 232 systems (46.4 percent), while HP came in second with 166 systems (33.2 percent). Dell came in a distant third with 24 systems (4.8 percent). Only in the top 50 systems do you see companies like Cray, SGI, Dell and Bull having a more pronounced presence. But if past volatility is a guide, new system manufacturers will emerge, while established ones may disappear. Today, Sun Microsystems and SGI are two companies that are aiming for a larger presence in high end computing.
One of the more compelling aspects of the TOP500 is how it can be used to trace the evolution of supercomputer technology. Meuer attempts to use Bell’s Law of Computer Classes to explain the evolution of supercomputing. Proposed by Gordon Bell in 1972, Bell’s Law describes how new computer architectures form, become established and then eventually phase out. A new class of computer appears to be introduced every ten years or so, although the phase out period may extend to multiple decades. For mainstream IT, the classes have been mainframes, minicomputers, workstations/personal computers, and distributed web computers. For supercomputing, the classes are quite different. Meuer categorizes them as follows:
- Data parallel systems
Vector (Cray Y-MP and X1, NEC SX etc.)
SIMD (CM-2 etc.)
- Custom scalar systems
MPP (Cray T3E and XT3, IBM SP etc.)
Scalar SMPs and Constellations (Cluster of big SMPs)
- Commodity clusters
NOW, PC cluster, Blades etc.
- Power-efficient systems
BG/L or BG/P as first examples of low-power systems. They might be able to form a new class, but we do not know this yet.
Perhaps one other computing class is emerging — that of the heterogenous machines. Currently the only heterogeneous machine on the TOP500 list is Sun’s TSUBAME supercomputer (at #16), with ClearSpeed floating point acceleration hardware hooked into Opteron-based Sun File servers. But Cray, IBM and others are constructing hybrid systems with exotic accelerators integrated into CPU-based machines. IBM is building the one petaflop Cell BE-Opteron-based Roadrunner machine for Los Alamos National Laboratory, while Cray is planning to deliver an XT5h hybrid system with Opteron and vector blades for the UK’s High End Computing Terascale Resource (HECToR). Both systems are scheduled to be completed later this year.
It’s possible that GPU-accelerated high-end systems are in the works, using NVIDIA or ATI gear, but nothing has been made public. In an EEtimes Asia article this week, NVIDIA chief scientist David Kirk predicts that by 2012 three of the top five supercomputers in the world will be equipped with graphics processors.
Over the past week, two of the most powerful supercomputers ever built were inaugurated at their respective facilities. Last Friday, Sun’s 504 teraflop “Ranger” system was officially dedicated by the NSF and TACC in Austin, Texas. Halfway around the world in Germany, Research Centre Jülich cut the ribbon on its new 223 teraflop “JUGENE” Blue Gene/P supercomputer. JUGENE currently sits at the number two spot on the TOP500 list with a Linpack mark of 167 teraflops. Although Ranger has yet to be officially benchmarked, it should easily eclipse the Linpack results for JUGENE should TACC decide to make a submission for the next TOP500 list in June. Considering that later this year a 556 teraflop Blue Gene/P is scheduled to be installed at Argonne National Laboratory and a one petaflop “Baker” Cray (presumably an XT5 or XT5h) system is set to be delivered at ORNL, the top of the TOP500 could be in for a big shakeup in 2008.
The 31st TOP500 list will be published June 18th in the Opening Session of the Interational Supercomputing Conference (ISC’08) in Dresden, Germany.