| Parallel processing accelerates computations by solving a single problem using multiple compute nodes interconnected by a network. The scalability of a parallel system is limited by its ability to communicate and coordinate processing. Circuit switching, packet switching and wormhole routing are dominant switching techniques. Our simulation results show that wormhole routing and circuit switching each excel under different types of traffic.; This dissertation presents a hybrid switching technique that combines wormhole routing with circuit switching in a single switch using virtual channels and time division multiplexing. The performance of this hybrid switch is significantly impacted by the efficiency of traffic scheduling and thus, this dissertation also explores the design and scalability of hardware scheduling for the hybrid switch. In particular, we introduce two schedulers for crossbar networks: a greedy scheduler and an optimal scheduler that improves upon the results provided by the greedy scheduler. For the time division multiplexing portion of the hybrid switch, this dissertation presents three allocation methods that combine wormhole switching with predictive circuit switching.; We further extend this research from crossbar networks to fat tree interconnected networks with virtual channels. The global "level-wise" scheduling algorithm is presented and improves network utilization by 30% when compared to a switch-level algorithm.; The performance of the hybrid switching is evaluated on a cycle-accurate simulation framework that is also part of this dissertation research. Our experimental results demonstrate that the hybrid switch is capable of transferring both predictable traffics and unpredictable traffics successfully. By dynamically selecting the proper switching technique based on the type of communication traffic, the hybrid switch improves communication for most types of traffic. |
