When threads are migrated from heavily loaded nodes to lightly loaded nodes for load balance in software distributed shared memory systems, the communication cost of maintaining data consistency is increased if migration threads are carelessly selected. Program performance is degraded when loss from increased communication exceeds the benefit from load balancing. This study addresses the problem with a novel selection policy called reduction of inter-node sharing costs. The main characteristic of this policy is simultaneously considering thread memory access types and global sharing. The experimental results show that this policy can reduce the communication of benchmark applications by 50% during load balancing.