Performance of multithreaded computations on high-speed networks

Abstract

Inter-host communication has always been a performance bottleneck for Distributed Computing Systems exchanging large amounts of data between hosts. This is due to the the low speed, and low shared bandwidth provided by the networking technologies that are currently used, such as Ethernet and Token Ring. With the emergence of several high-speed and high-bandwidth networks like ATM and Gigabit Ethernet, fast inter-host communication for solving real time problems, has become possible. Almost all software systems that support distributed concurrent computing use a process based computing model and a message-passing based communication model. However, recent research developments suggest that a thread based model for computing and scheduling has specific advantages over the traditional model. In this thesis, we have implemented a thread based computing model and a thread based intra-host communication model and have tested two networking technologies namely ATM and Ethernet for supporting these models. We implement a recursive distributed matrix multiplication algorithm and the results are very encouraging, indicating the feasibility of developing portable and better performing thread based systems for distributed concurrent computing.