| Clusters built from commodity PCs/Workstations are becoming a predominant and cost-effective style for designing scalable servers. Wide spread use of cluster systems in a diverse set of applications has spurred significant interest in designing such servers considering performance, scalability, and quality-of-service (QoS) as the design objectives. In addition to these parameters, optimizing energy consumption in these architectures has recently emerged as a major concern since server power usage is becoming a significant fraction of the total ownership cost. Understanding the limitation of the current cluster interconnects, the recently proposed InfiniBand(TM) Architecture (IBA) specification focuses these issues for designing future System Area Networks (SANs) or clusters. However, IBA design is currently in its infancy since the released specification outlines only higher level functionalities, leaving it open for exploring various design alternatives. Design and analysis of high performance and energy efficient IBA-style SANs is the focus of this thesis.; In this thesis, we address five issues for designing high performance and energy efficient cluster interconnects. First, a pipelined switch/router and a network interface card (NIC) are proposed. In this study, we investigate how to accommodate various applications in a unified framework. The second part of the thesis deals with the integration of admission and congestion control mechanisms in a SAN environment. A simple, credit-based congestion control algorithm, which relies on the hardware already available in the NIC to regulate traffic injection, is proposed. Third, we investigate four co-related techniques (routing, multipathing, packet dropping, and multicasting) for providing high and predictable performance in IBA. A comprehensive simulation testbed for IBA has been developed to address these design issues. Next, using a cycle-accurate simulator of an IBA compliant interconnect fabric and actual designs of its components, we investigated the energy behavior on regular and irregular interconnects. To minimize power in the links, we propose a novel dynamic link shutdown (DLS) technique. The DLS technique makes use of an appropriate adaptive routing algorithm to shutdown the links intelligently. Finally, analytical models for QoS capable cluster networks are developed. The models capture various design intricacies and can be used as an effective design tool. |
