Crunching Numbers, Two Ways
Kim S. NashOil-exploration firm CGG Americas replaced its supercomputers with a cluster of several thousand personal computers linked by specialized networking software. Here's one view of how the alternatives stack up. Silicon Graphics Origin 3000 Cluster of Dell Computer PCs Architecture Symmetric multiprocessing, where many processors share a single operating system and access each other's memory Parallel processing, where a computing job is split among many processors that each have their own operating system and memory Cost of system $2.9 million $2 million Number of processors 128 1,024 Cost per processor $22,656 $1,953 System memory 256 gigabytes Up to 2 terabytes Pluses Designed specifically for scientific applications that do highly complex modeling All processors can communicate quickly with all other processors, which means certain computing jobs can be done very quickly Relatively low cost of ownership Distributed architecture decreases risk that a single processor failure will stop entire job Power can be added incrementally Often uses Linux, a hot technology for which skilled workers are readily found Minuses Processor failure can halt an entire computing job Adding power generally adds expense: another $2.9 million machine Experienced technicians hard to find Setup and configuration often take several weeks Overall performance can be negatively affected by speed of the network connecting the processors Constant monitoring of individual parts is necessary; more parts means more potential points of failure Large physical space required as cluster grows; additional air conditioning and electric power are sometimes needed
Copyright © 2004 Ziff Davis Media Inc. All Rights Reserved. Originally appearing in Baseline.