A two-phase scheduling algorithm for efficient collective communications of MPICH-G2

Abstract

In this paper, we propose a packet-level parallel data transfer and a Two-Phase Scheduling(TPS) algorithm for collective communication primitives in MPICH-G2. The algorithms are characterized by two unique features: 1) a concurrent data transfer of packets from a source node to multiple destination nodes and 2) a scheduling of enhancing the performance of collective communications by early identification of bottleneck incurring nodes. The proposed technique is implemented and the performance improvement is measured. According to the performance evaluation, the proposed method has achieved about 20% performance improvement against conventional block data transfer methods when a binomial tree is used for the communication in LAN. In TPS algorithm, the distribution of messages to bottleneck incurring nodes is delayed to minimize the affection of the node to the total performance. Using TPS algorithm on WAN, significant performance improvement has also been achieved for various data sizes and number of nodes.