Acceleration of the hardware-software interface of a communication device for parallel systems

During the last decades the ever growing need for computational power fostered the development of parallel computer architectures. Applications need to be parallelized and optimized to be able to exploit modern system architectures. Today, scalability of applications is more and more limited both by development resources, as programming of complex parallel applications becomes increasingly demanding, and by the fundamental scalability issues introduced by the cost of communication in distributed memory systems. Lowering the latency of communication is mandatory to increase scalability and serves as an enabling technology for programming of distributed memory systems at a higher abstraction layer using higher degrees of compiler driven automation. At the same time it can increase performance of such systems in general. In this work, the software/hardware interface and the network interface controller functions of the EXTOLL network architecture, which is specifically designed to satisfy the needs of low-latency networking for high-performance computing, is presented. Several new architectural contributions are made in this thesis, namely a new efficient method for virtual-tophysical address-translation named ATU and a novel method to issue operations to a virtual device in an optimal way which has been termed Transactional I/O. This new method needs changes in the architecture of the host CPU the device is connected to. Two additional methods that emulate most of the characteristics of Transactional I/O are developed and employed in the development of the EXTOLL hardware to facilitate usage together with contemporary CPUs. These new methods heavily leverage properties of the HyperTransport interface used to connect the device to the CPU. Finally, this thesis also introduces an optimized remote-memory-access architecture for efficient split-phase transactions and atomic operations. The complete architecture has been prototyped using FPGA technology enabling a more precise analysis and verification than is possible using simulation alone. The resulting design utilizes 95 % of a 90 nm FPGA device and reaches speeds of 200 MHz and 156 MHz in the different clock domains of the design. The EXTOLL software stack is developed and a performance evaluation of the software using the EXTOLL hardware is performed. The performance evaluation shows an excellent start-up latency value of 1.3 μs, which competes with the most advanced networks available, in spite of the technological performance handicap encountered by FPGA technology. The resulting network is, to the best of the knowledge of the author, the fastest FPGA-based interconnection network for commodity processors ever built