Photonic Interconnection Networks for Applications in Heterogeneous Utility Computing Systems

Growing demands in heterogeneous utility computing systems in future cloud and high performance computing systems are driving the development of processor-hardware accelerator interconnects with greater performance, flexibility, and dynamism. Recent innovations in the field of utility computing have led to an emergence in the use of heterogeneous compute elements. By leveraging the computing advantages of hardware accelerators alongside typical general purpose processors, performance efficiency can be maximized. The network linking these compute nodes is increasingly becoming the bottleneck in these architectures, limiting the hardware accelerators to be restricted to localized computing. A high-bandwidth, agile interconnect is an imperative enabler for hardware accelerator delocalization in heterogeneous utility computing. A redesign of these systems' interconnect and architecture will be essential to establishing high-bandwidth, low-latency, efficient, and dynamic heterogeneous systems that can meet the challenges of next-generation utility computing. By leveraging an optics-based approach, this dissertation presents the design and implementation of optically-connected hardware accelerators (OCHA) that exploit the distance-independent energy dissipation and bandwidth density of photonic transceivers, in combination with the flexibility, efficiency and data parallelization offered by optical networks. By replacing the electronic buses with an optical interconnection network, architectures that delocalize hardware accelerators can be created that are otherwise infeasible. With delocalized optically-connected hardware accelerator nodes accessible by processors at run time, the system can alleviate the network latency issues plague current heterogeneous systems. Accelerators that would otherwise sit idle, waiting for it's master CPU to feed it data, can instead operate at high utilization rates, leading to dramatic improvements in overall system performance. This work presents a prototype optically-connect hardware accelerator module and custom optical-network-aware, dynamic hardware accelerator allocator that communicate transparently and optically across an optical interconnection network. The hardware accelerators and processor are optimized to enable hardware acceleration across an optical network using fast packet-switching. The versatility of the optical network enables additional performance benefits including optical multicasting to exploit the data parallelism found in many accelerated data sets. The integration of hardware acceleration, heterogeneous computing, and optics constitutes a critical step for both computing and optics. The massive data parallelism, application dependent-location and function, as well as network latency, and bandwidth limitations facing networks today complement well with the strength of optical communications-based systems. Moreover, ongoing efforts focusing on development of low-cost optical components and subsystems that are suitable for computing environment may benefit from the high-volume heterogeneous computing market. This work, therefore, takes the first steps in merging the areas of hardware acceleration and optics by developing architectures, protocols, and systems to interface with the two technologies and demonstrating areas of potential benefits and areas for future work. Next-generation heterogeneous utility computing systems will indubitably benefit from the use of efficient, flexible and high-performance optically connect hardware acceleration