The road from Gigaflop/s to Petaflop/s

Mannheim, 21-6-97 Paul Messina from CalTech presented approaches, investigations and possibilities to reach Petaflop/s performance at the Mannheim Supercomputer Seminar. There are applications that need this enormous amount of computing power: 1,000 times faster than the current fastest machines. Messina talked about studies that propose and examine innovative approaches to achieve this performance in the year 2007 time frame. The latest machine at the CalTech computing centre, a 256 processor Exemplar, has been installed this month. Messina gave details on this machine.

CACR Computing Resources

The CalTech supercomputing centre CACR has a collection of high-performance systems. The newest machine is a Hewlett-Packard/Convex Exemplar X-class, 256 PA-8000 processors, 184 GFlop/s aggregate peak speed, 64 GBytes global shared memory (installed April - June 1997).

As this is the largest Exemplar system, Messina reported first experiences gained on the 64 processor machine. The performance on message-passing codes well tuned to the Cray T3D (ocean modelling, electron-molecule collisions) is 4-5 times better.

An astrophysics plasma code (a sequential code) parallelised well up to 16 processors.

The hardware and software were stable.

The other computing resources at CARCR are: Intel Paragon, 512 computational nodes (38 GFlop/s, 20 GBytes), Intel Delta, 512 computational nodes (30 GFlop/s, 8 GBytes), Cray T3D, 256 processors (38 GFlop/s, 16 GBytes), IBM SP2, 19 nodes in total, different types, one high-node with 8 CPUs.

Messina also runs a Beowulf Class PC Cluster, 62 Pentium Pros with an exceptional price-performance $12/peak Mflop/s, $30/sustained Mflop/s, this means 1 Gflop/s delivered costs less than $30K.

Petaflop/s studies

CACR is involved in Petaflop/s studies since 1994. Messina reminded the audience of the scales involved, it might look like 10,000 100 Gigaflop/s processors or 1,000,000 1 Gigaflop/s and told that such a machine will do in 5 minutes, what takes a Gigaflop/s computer 10 years to do.

With current technology such a computer consumes 1 to 2 Gigawatts of power. He projected technology in DRAMs, logic capacity and microprocessor path to 2007 and believes that with 2004-technology a Petaflop/s machine can be built in 2007. The cost is still a prohibiting factor. A brute force approach would cost more than $40 billion, with major advances in the next 15 to 20 years Messina expects prices to go down to $100 million to $1000 million.

Messina believes that 4 types of systems and architectures have to be studied:

extrapolation of current off the shelf technology machines such as IBM SP2 or SGI Origin
exotic technology with additional levels of memory hierarchy
processors in memory
special purpose computers

Uwe Harms