EVALUATION OF MASSIVELY PARALLEL COMPUTING FOR EXHAUSTIVE AND CLUSTERED MATCHED-FIELD PROCESSING

Abstract

Many computer algorithms contain an operation that accounts for a substantial portion of the total execution cost in a frequently executed loop. The use of a parallel computer to execute that operation may represent an alternative to a sheer increase in processor speed. The signal processing technique known as matched-field processing (MFP) involves performing identical and independent operations on a potentially huge set of vectors. To investigate a massively parallel approach to MFP and clustered nearest neighbors MFP, algorithms were implemented on a DECmpp 12000 massively parallel computer (from Digital Equipment and MasPar Corporation) with 8192 processors. The execution time for the MFP technique on the MasPar machine was compared with that of MFP on a serial VAX9000–210 equipped with a vector processor. The results showed that the MasPar achieved a speedup factor of at least 17 relative to the VAX9000. The speedup was 3.5 times higher than the ratio of the peak ratings of 600 MFLOPS for the MasPar versus 125 MFLOPS for the VAX9000 with vector processor. The execution speed on the parallel machine represented 64% of its peak rating. This is much better than what is commonly assumed for a parallel machine and was obtained with modest programming effort. An initial implementation of a massively parallel approach to clustered MFP on the MasPar showed a further order of magnitude increase in speed, for an overall speedup factor of 35.