Advances in multi-paradigm computing are bringing supercomputers with petascale speeds closer to reality.
Speaking at the Solid Earth and Environment Grid (SEE Grid) conference in Canberra yesterday, SGI weather and climate business development manager, Ilene Carpenter, said that interest in petascale computers, which can perform quadrillions of calculations per second, is increasing.
"There's a lot of interest around the world in deploying petascale systems in the next two to five years," Carpenter said.
SGI is developing several methods to overcome the barriers to efficient processing presented by computing on such a large scale.
In order to attain petascale speeds using current supercomputer designs, it requires tens or even hundreds of thousands of processors.
But because of inherent system inefficiencies large-scale process clusters can operate with as little as 5 per cent efficiency.
SGI says it has developed a method to increase efficiency in large-scale process clusters. This method, multi-paradigm computing, combines a range of processing techniques into a single dual-chip system.
The multi-paradigm project, dubbed Ultraviolet, will combine a scalar-system foundation chip with a chip providing vector, Processor in Memory (PIM) and application-specific processing capacity.
SGI is also developing large-scale shared memory systems for petascale computers.
"We're continuing to design big shared memory systems...with large cache coherency domains," Carpenter said.
Cache coherency measures the ability of multi-processing systems to share cached memory without errors caused by conflicts.
"We think [large cache coherency is] an important element in improving productivity on systems. It will also allow some of the next generation languages that depend on globally accessible memory to work well," she said.
Developers of petascale systems will have to tweak existing multi-processor management techniques to cope with the increased load.
For example, SGI is adding features which will increase the scalability of synchronization techniques.
"If we're designing a system that's going to have 50,000 processors," Carpenter said, "and you need to do a barrier and synchronize all of them, we know we can't use the same methods that we use today with 2,000 or 3,000 processors".
Vendors also working on petascale computing include supercomputer giants Cray Inc and Fujitsu.
In contrast to the methods Western developers currently adopt using thousands of retail-strength CPUs, Japanese developers are attempting to design petascale computers which use a much lower number of high-performance processors.
In the system being designed by Fujitsu, for example, each processor will operate at around 100 gigaflops per second.
In order to develop efficient petascale systems, experts predict that advances in cooling and power-distribution systems will be needed to accompany advances in memory and processor speeds and connectivity efficiency.