Supercomputers are notoriously unreliable. They contain so many more processors, memory chips, disks and cables than an ordinary computer, that the "law of averages" alone dictates failures will occur much more frequently. In addition, some supercomputing software and hardware components are products of cutting-edge research, new and not yet "debugged." Finally, supercomputers are usually placed under heavy computing and communications workloads 24 hours a day.

Supercomputer failures and "downtime" are extremely costly, due to the importance of their "computational mission" and the scarcity of available systems. To reduce downtime, modern supercomputers are built with a focus on three key aspects of operation:

• Reliability - likelihood of a failure occurring in the running system
• Availability - "uptime" of system and/or system resources in the presence of a failure
• Serviceability - ability to quickly detect, isolate, and recover from failures that occur

Some supercomputer designs incorporate additional hardware and software to support system RAS. Redundant power supplies and cooling systems are common, for example. A single failure in a redundant component typically does not cause downtime (impact the availability of the system). Hot-swappable components like disks and power supplies also are commonly employed to improve system serviceability.

The ASCI Red system features a Ethernet network, separate from the primary interconnect, that is used to detect and recover from failures to improve serviceability and manageability.

SMASH RAS

The IBM Blue Gene system adopts an architectural approach called SMASH (Simple, Many and Self-Healing). SMASH attempts to build into the system features that will minimize failures and downtime.

Conclusion

Supercomputing is currently the most advanced of all high-performance computing (HPC) approaches. HPC alternatives like clustering and grid computing hold great promise for achieving high performance computing and communications at a low cost, but for now, supercomputers remain the world's fastest computing systems best suited for solving many critical scientific problems.

Supercomputers remain expensive and highly specialized. They run a limited set of applications and require non-stop "care and feeding" to keep running smoothly.

The IBM Blue Gene system represents the next generation of advanced supercomputing technology. When complete, it will operate at speeds up to 100 times greater than today's systems. In addition to solving computational problems, Blue Gene should influence the way in which mainstream computers of the future are built.