Design and implementation of message passing library on Maestro network

Most interconnections and communication software of recent clusters are LAN or WAN-based which include large communication overheads, and are not optimal for parallel computing. This paper presents a design and an implementation of message passing library called MAP. MMP is designed so that the potential performance of Maestro cluster can be extracted efficiently. Maestro has a dedicated network that consists of high performance network interfaces(NI) and a switch. In this paper, we propose two optimization techniques: optimizing pipelined transfer and autonomous memory management by NI. Evaluation has been done using round-trip communication and application programs. Through the evaluation, we have confirmed that MMP utilized dedicated hardware of Maestro efficiently and achieved high performance.