AbstractMulti‐PSI is a distributed‐memory MIMD (multiple‐instruction, multiple‐data) computer that has 64 processors. Problems of knowledge processing, which are the object of the Multi‐PSI, must be divided into many partial problems in order to achieve a favorable load balance because a characteristic of the computation is the dynamic loss of uniformity. This, however, has the danger of increasing the communication overhead. Therefore, in order to constrain the communication overhead and achieve a favorable load balance, in this paper, we propose a programming method oriented to the above problem. In this methodology, the problem is formulated as multiple communicating objects and the degrees of freedom in the load allocation are maintained. On the other hand, when the load is allocated, the communication objects are estimated from the system performance and the objects' characteristics, and the processing granularity is adjusted so that it is within the tolerance. This program methodology was applied to the three program developments of the shortest path problem, the LSI routing problem, and logic simulation; each was implemented with high efficiency. For example, in the shortest path program, 260,000 small‐grained objects were distributed over 64 processors and reached about 75% efficiency. In addition, in the logic simulation composed of 12,000 objects (gates), a high absolute efficiency of 99 thousand events/second and a 48‐fold increase in speed with 64 processors were obtained. Along with confirming the effectiveness of this programming methodology, these results demonstrated. that high efficiency can be realized in parallel processing for some degree of granularity even in a distributed‐memory pa