Share |

What is high performance computing?

A high-performance computing system, or supercomputer, consists of thousands of computer cores in a single room, connected by extremely fast interconnects in order to minimize the time it takes for cores in separate computers to communicate (that time is the latency, and HPC systems are low-latency). HPC systems are normally ranked by the number of floating operations they can execute each second (flops).

For example, as of 10 January 2011, top-ranked HPC system in the world is the Tianhe supercomputer in Tianjin, China. Tianhe consists of over 186,000 computing cores, and can do between 2500 and 4700 teraflops.

Supercomputers are required to solve high-performance problems in which distinct computing nodes need to communicate frequently, resulting in a communications bottleneck; these sorts of calculations are examples of fine-grained parallelism. They can also be used to do coarse-grained parallel and task-parallel calculations, although many of those can be accomplished on systems that are less expensive to operate.

In contrast, grid computing (often called “high throughput computing”) is usually used for tasks that can be broken down into many small, repetitive tasks in which there is a huge volume — such as looking at the results from a supercollider, in which there is enough data generated in one year (15 Petabytes) that it would be the equivalent of filling up 500 years’ worth of DVDs if the data were music ripped to an MP3 at 192kbps. A grid of computers also allows for many smaller computers to be combined, thus helping scientists to tackle problems that cannot be solved well on a single system, and to solve problems much more quickly than can be done on a single machine.

Just to make things confusing, you can hook up one supercomputer to another, creating an HPC grid. Some examples of HPC grids include DEISA in Europe and TeraGrid in the U.S. In Europe, DEISA works alongside PRACE to create a pan-European HPC service and infrastructure. In the US, the TeraGrid integrates high-performance computers, data resources and tools, and high-end experimental facilities around the country.