ASCI Red - System Structure

System Structure

The ASCI Red supercomputer was a distributed memory MIMD (Multiple Instruction, Multiple Data) message-passing computer. The design provided high degrees of scalability for I/O, memory, compute nodes, storage capacity, and communications; standard parallel interfaces also made it possible to port parallel applications to the machine. The machine was structured into four partitions: Compute, Service, I/O, and System. Parallel applications executed in the Compute Partition which contained nodes optimized for floating point performance. The compute nodes had only the features required for efficient computation – they were not purposed for general interactive services. The Service Partition provided an integrated, scalable host that supported interactive users (log-in sessions), application development, and system administration. The I/O Partition supported disk I/O, a scalable parallel file system and network services. The System Partition supported initial booting and system Reliability, Availability, and Serviceability (RAS) capabilities.

The Service partition helps integrate all of the different parts of ASCI Red together. It provides a scalable host for users, and it is used for general system administration. The I/O Partition provides a file system and network services, and the Service partition is made up of the log-in screens, tools for application development, and utilities for network connections. The Compute partition contains nodes that are designed for floating point performance. This is where the actual computing takes place. Every one of the compute nodes has two 333 MHz Pentium II Xeon processors, each with a 16 KB level-1 cache and a 256 KB level-2 cache. According to Intel, the ASCI Red Computer is also the first large scale supercomputer to be built entirely of common commercially available components.

All of ASCI Red's partitions are interconnected to form one supercomputer, however at the same time none of the nodes support global shared memory. Each of the nodes works in its own memory, and each shares data with the others through "explicit message-passing".