PC clusters for signal processing: an early prototype

A cluster of commodity personal computers (PCs) is an attractive platform for signal processing, especially for applications like sensor arrays where there is significant parallelism in the computation. This paper reports the results of experiments with several small cluster configurations to explore the opportunities and limitations of this approach. Experiments included uniprocessor and multiprocessor PCs, 100 Mbit/sec and 1 Gbit/sec Ethernet. All experiments were receive-only, starting from raw spectrum samples provided by a wideband A/D card, and all signal processing code ran as application processes on top of standard Linux. The applications used to drive the experiments were delay and sum beamforming and the simultaneous reception of multiple AMPS cellular channels. The current standard PCI I/O bus (32 bits at 33 MHz) was the most significant bottleneck in the experiments. This suggests that the arrival of the next generation of commodity PCs, with faster I/O busses (such as 64-bit PCI at 66 MHz), may open the door to their use as a low-cost platform for a variety of low-end real-time signal processing applications.