An optimised and generalised node for fat tree classes

  • Adamantini Peratikou

Student thesis: Doctoral Thesis


Fat tree topologies have been extensively used as interconnection networks for high performance parallel systems, with their most recent variants able to easily extend and scale to accommodate different system sizes and requirements. While each progressive and evolved fat-tree topology includes some extra advancements compared to the original ones, these topologies do not fully address or resolve the issues that large scales systems face.

We propose an extended Zoned node, architecture as an alternative to conventional fat trees, and other variants. The extension relates to the provision of extra links to balance the number of routing switches, and hence increases the bisection bandwidth, and also to the extra layers that provide an inherent fault tolerance. In this work we emphasize on controlled power consumption, managed network complexity, faster message transmission, lower latency and higher throughput. These features are desirable for high performance parallel systems. We will show through semantics that our zoned node specifies most variants of the fat trees topologies such as k-ary n-tree and XGFT. We will also show that a replication of several zoned nodes into super nodes takes a broader view of complex interconnections such as dragonfly and PERCS.

We propose a generic source routing algorithm that we call in this thesis “the single-sliced-addressing” that works for all the classes of the zoned nodes, and we will prove through analysis and simulation that it outperforms the previous routing scheme deployed in some variant fat tree topologies. Most variants of fat-tree topologies do not address optimisation issues. In this work, we develop an optimising system that identifies parameters and components of the zoned node that lead to an optimized architecture. The optimization process is achieved based on minimising the overall cost of the network that is directly related to its complexity and therefore proportional to the relative power, which serves as the objective function that is minimised based on the traffic constraints to maintain a lower delay and a higher throughput. The simulation results show that the extracted optimised zone node performs well under various load conditions and traffic patterns compared to non-optimised variants of fat tree topologies.
Date of AwardApr 2014
Original languageEnglish
SupervisorMo Adda (Supervisor) & Bryan Carpenter (Supervisor)

Cite this