Sorting Large Data Files on POOMA

This paper reports on the results of the porting of a typical benchmark problem for distributed systems from its original platform [Baug89b] (the HC16 database machine 1) to POOMA, the Parallel Object Oriented MAchine 2. The benchmark, a sorting algorithm for a data set distributed over (sequential accessible) background devices of a number of processing nodes, was implemented on POOMA at two levels. The first level is the nucleus of the POOMA operating system and the implementation is written in C. The results show that the benchmark is able to exploit the full capabilities of the computing power, the storage devices and the communication bandwith. Porting the program to the first level has been done by B. A. W. Baugst¢ and J. F. Greipsland in September 1989. The second level is that of the programming language POOL, an object oriented higher level programming language for distributed symbolic processing (Parallel Object Oriented Language). Porting the program to the second level has been done by J. Kamerbeek in October 1989. The results show the suitability of that language for expressing the algorithm and again the applicability of the POOMA platform for this application, even where the POOL implementation was, at the time of the experiment, far from optimal. In both cases the speed up is, for large data sets, almost linear in the number of processors. This paper is organized in the following way. Section 2 introduces the POOMA architecture and some of its rationale. Section 3 defines the sorting problem and the specific algorithm. Section 4 deals with the C-version of the algorithm for POOMA and presents the results of a number of measurements. Section 5 covers the POOL implementation and its results. The paper ends with some conclusions.