Improving the Scalability of High Performance Computer Systems

Improving the performance of future computing systems will be based upon the ability of increasing the scalability of current technology. New paths need to be explored, as operating principles that were applied up to now are becoming irrelevant for upcoming computer architectures. It appears that scaling the number of cores, processors and nodes within an system represents the only feasible alternative to achieve Exascale performance. To accomplish this goal, we propose three novel techniques addressing different layers of computer systems. The Tightly Coupled Cluster technique significantly improves the communication for inter node communication within compute clusters. By improving the latency by an order of magnitude over existing solutions the cost of communication is considerably reduced. This enables to exploit fine grain parallelism within applications, thereby, extending the scalability considerably. The mechanism virtually moves the network interconnect into the processor, bypassing the latency of the I/O interface and rendering protocol conversions unnecessary. The technique is implemented entirely through firmware and kernel layer software utilizing off-the-shelf AMD processors. We present a proof-of-concept implementation and real world benchmarks to demonstrate the superior performance of our technique. In particular, our approach achieves a software-to-software communication latency of 240 ns between two remote compute nodes. The second part of the dissertation introduces a new framework for scalable Networks-on-Chip. A novel rapid prototyping methodology is proposed, that accelerates the design and implementation substantially. Due to its flexibility and modularity a large application space is covered ranging from Systems-on-chip, to high performance many-core processors. The Network-on-Chip compiler enables to generate complex networks in the form of synthesizable register transfer level code from an abstract design description. Our engine supports different target technologies including Field Programmable Gate Arrays and Application Specific Integrated Circuits. The framework enables to build large designs while minimizing development and verification efforts. Many topologies and routing algorithms are supported by partitioning the tasks into several layers and by the introduction of a protocol agnostic architecture. We provide a thorough evaluation of the design that shows excellent results regarding performance and scalability. The third part of the dissertation addresses the Processor-Memory Interface within computer architectures. The increasing compute power of many-core processors, leads to an equally growing demand for more memory bandwidth and capacity. Current processor designs exhibit physical limitations that restrict the scalability of main memory. To address this issue we propose a memory extension technique that attaches large amounts of DRAM memory to the processor via a low pin count interface using high speed serial transceivers. Our technique transparently integrates the extension memory into the system architecture by providing full cache coherency. Therefore, applications can utilize the memory extension by applying regular shared memory programming techniques. By supporting daisy chained memory extension devices and by introducing the asymmetric probing approach, the proposed mechanism ensures high scalability. We furthermore propose a DMA offloading technique to improve the performance of the processor memory interface. The design has been implemented in a Field Programmable Gate Array based prototype. Driver software and firmware modifications have been developed to bring up the prototype in a Linux based system. We show microbenchmarks that prove the feasibility of our design