Two-Tree Collective Communication in Distributed Heterogeneous Systems

This paper studies the collective communication in the grid computing environment, which is characterized by the combination of distributed heterogeneous networks as well as uneven, long communication delay. Efficient collective communication requires communication schedule, which in turn requires network information. When the network information is not accurate or network faults occur, the performance of collective communication can be significantly degraded. This paper proposes TTCC (Two-Tree Collective Communication) for scheduling collective communication in the grid. It provides an efficient and reliable schedule even in this unfavorable network condition by maintaining two disjoint communication trees. Benefits of the proposed method are manifested via simulation, where the performance degradation with TTCC is much slower than those using conventional scheduling algorithms.