Proc. of FGCS '94, ICOT, Tokyo, December 1994

The Evaluation of Parallel Inference Machines

Kouichi KUMON and Hiroyoshi HATAZAWA
Institute for New Generation Computer Technology
4-28, Mita 1-Chome, Minatoku, Tokyo 108, JAPAN
{kumon,f-hataza}@icot.or.jp

Abstract:

In this paper, we study the performance of various implementations of KL1 processing systems, including the parallel inference machine (PIM) system and the KLIC system. PIM systems are built on dedicated parallel hardware developed in the FGCS project for efficient KL1 execution, while the KLIC system is KL1 ported to run on UNIX workstations. This difference makes the design policies of both systems quite different necessitating a quantitative analysis. KL1 is a parallel logic language running on hardware ranging from a single processor to several hundred processors. Thus performance measurements must cover this range.

Firstly we use a set of small benchmark programs for single processor performance. It includes a naive reverse program to compare and characterize those systems. And we take a close look at the execution of instructions in both the PIM KL1 processing system and the KLIC system. The PIM/p system is a convenient platform because both the conventional KL1 system and the sequential core of the KLIC system are available on it, and thus we can compare both implementations on the same base. We also compare both system using the life program, which shows quite a different aspect to the append execution.

Then, we evaluate the cluster performance for automatic load balancing. Some PIM models have a cluster structure, in which processors are connected by a common bus and share memory. In parallel processing in a cluster, automatic load balancing is available to reduce the burden of writing load distribution code. It is thus meaningful to investigate load balancing mechanisms and measure the speed-up in parallel performance. To do this, we make execution models for a very small program, examine the parallel performance, and estimate the distribution overhead in the cluster.

Finally, we investigate the distributed performance of the systems by using simple communication dependent benchmark programs. The PIM models use a dedicated network for communication between clusters/nodes. However as the KLIC system uses PVM for message passing primitives, we found that the performance of KLIC wholly depends on the PVM performance.


All of this proceedings (Compressed PostScript) :

The Evaluation of Parallel Inference Machines (114KB)


www-admin@icot.or.jp