In a previous article, we presented a per cluster logical cache strategy for a software distributed shared memory (SDSM) system for interconnects of loosely-coupled machine clusters. We assume that inter-cluster links are slower than intra-cluster ones. We propose an adaptive strategy that aims at increasing the efficiency of our first cache solution. By exploiting sharing access pattern of applications related to locality within the same cluster, the adaptive cache reduces the number of messages generated by our previous cache solution. Performance evaluation results show that the former outperforms the latter. We also present a new barrier algorithm, which is integrated with the adaptive cache and more adapted to the communication latency hierarchy of the platform.