Communication and scheduling in clusters: A user-level perspective

Cluster computing has become the preferred model for building high performance servers today. The rapid advances and commoditization of microprocessors, network hardware and system software has made it possible for clusters to be built from off-the-shelf components. However the performance of such clusters is hampered by the software overheads of communication and scheduling at each node. The improvement of cluster communication and scheduling performance are important areas of research for modern computer systems.; This thesis explores the use of user-level networking for improving cluster communication. By using intelligent network interface cards and optimized implementations of well-designed protocols, user-level networking can significantly reduce communication overheads. The thesis examines the design and implementation of user-level network interfaces from a software and hardware perspective. It studies the benefits of user-level networking in clusters and ways to improve the scalability of ULN implementations. It also investigates how quality of service can be efficiently provided in a ULN. User-level communication has been used to improve the scheduling of tasks across cluster nodes. The thesis examines communication-driven coscheduling mechanisms and their impact on parallel application performance. It is shown that communication and scheduling are closely related and overall cluster performance depends on optimizing the two subsystems together.