Topology Agnostic Methods for Routing, Reconfiguration and Virtualization of Interconnection Networks

Abstract

Modern computing systems, such as supercomputers, data centers and multicore chips, generally require efficient communication between their different system units; tolerance towards component faults; flexibility to expand or merge; and a high utilization of their resources. Interconnection networks are used in a variety of such computing systems in order to enable communication between their diverse system units.

Investigation and proposal of new or improved solutions to topology agnostic routing and reconfiguration of interconnection networks are main objectives of this thesis. In addition, topology agnostic routing and reconfiguration algorithms are utilized in the development of new and flexible approaches to processor allocation. The thesis aims to present versatile solutions that can be used for the interconnection networks of a number of different computing systems.

No particular routing algorithm was specified for an interconnection network technology which is now incorporated in Dolphin Express. The thesis states a set of criteria for a suitable routing algorithm, evaluates a number of existing routing algorithms, and recommend that one of the algorithms – which fulfils all of the criteria – is used. Further investigations demonstrate how this routing algorithm inherently supports fault-tolerance, and how it can be optimized for some network topologies. These considerations are also relevant for the InfiniBand interconnection network technology.

Reconfiguration of interconnection networks (change of routing function) is a deadlock prone process. Some existing reconfiguration strategies include deadlock avoidance mechanisms that significantly reduce the network service offered to running applications. The thesis expands the area of application for one of the most versatile and efficient reconfiguration algorithms available in the literature, and proposes an optimization of this algorithm that improves the network service offered to running applications. Moreover, a new reconfiguration algorithm is presented that supports a replacement of the routing function without causing performance penalties.

Processor allocation strategies that guarantee traffic-containment commonly pose strict requirements on the shape of partitions, and thus achieve only a limited utilization of a system’s computing resources. The thesis introduces two new approaches that are more flexible. Both approaches utilize the properties of a topology agnostic routing algorithm in order to enforce traffic-containment within arbitrarily shaped partitions. Consequently, a high resource utilization as well as isolation of traffic between different partitions is achieved.