Specializing General-Purpose Computing: A New Approach to Designing Clusters for High-Performance Technical Computing

High-performance technical computing stresses computer systems in many ways, from CPU performance to memory systems to inter-system communication. Over the past twelve years, clusters of commodity hardware running Linux have become the most common tool for high-performance computing. However, the dynamics of such applications are often very different from those of applications that drive the design of commodity computer systems, which means that commodity systems may be cheap for computing, but they are not efficient for many technical applications.

At the same time, it has become increasingly clear that the costs of software dominate the overall investment. Providing more efficient computing for the same software can therefore be extremely valuable in improving the delivered performance for technical applications. With Linux on clusters as the primary environment for such applications, we can consider how to design a system optimized for executing high-performance technical code on Linux. SiCortex has recently introduced one example of such a system. The SC5832 and SC648 are computer systems designed from the silicon up to run Linux cluster applications. The systems feature a large number of processor cores, integrated high-speed interconnect fabric, a small footprint, and low power usage. In essence, the hardware design has taken the important components for a cluster node and shrunk them onto a single chip (plus DRAM); the software provides a coherent, turnkey Linux cluster computing environment.

We will examine the technology trends that converge to make this design approach both feasible and attractive, including the role of lowering power consumption as a means of gaining performance. In addition, we will describe the design of individual processor nodes, the interconnect fabric, and the challenges of assembling a Linux cluster distribution.