On a multiprocessor, load balancing is an essential technique to improve the performance of parallel processing by efficiently utilizing the processing power of the system. In this paper, we propose a dynamic and distributed algorithm for load balancing on TRAN-SCUBE, a multiprocessor system without shared memory. The algorithm adopts receiver-initiated asychronous strategy by which an idle processor first initiates a "hadshake" procedure to locate a heavily loaded processor, then absorbs som work from it.This algorithm is further demonstrated in the context of parallel soving of a maze problem.Simulated results show the it is effective and is more applicable to large size problems.