Energy-efficient architectures for chip-scale networks and memory systems using silicon-photonics technology

Today's supercomputers and cloud systems run many data-centric applications such as machine learning, graph algorithms, and cognitive processing, which have large data footprints and complex data access patterns. With computational capacity of large-scale systems projected to rise up to 50GFLOPS/W, the target energy-per-bit budget for data movement is expected to reach as low as 0.1pJ/bit, assuming 200bits/FLOP for data transfers. This tight energy budget impacts the design of both chip-scale networks and main memory systems. Conventional electrical links used in chip-scale networks (0.5-3pJ/bit) and DRAM systems used in main memory (>30pJ/bit) fail to provide sustained performance at low energy budgets. This thesis builds on the promising research on silicon-photonic technology to design system architectures and system management policies for chip-scale networks and main memory systems. The adoption of silicon-photonic links as chip-scale networks, however, is hampered by the high sensitivity of optical devices towards thermal and process variations. These device sensitivities result in high power overheads at high-speed communications. Moreover, applications differ in their resource utilization, resulting in application-specific thermal profiles and bandwidth needs. Similarly, optically-controlled memory systems designed using conventional electrical-based architectures require additional circuitry for electrical-to-optical and optical-to-electrical conversions within memory. These conversions increase the energy and latency per memory access. Due to these issues, chip-scale networks and memory systems designed using silicon-photonics technology leave much of their benefits underutilized. This thesis argues for the need to rearchitect memory systems and redesign network management policies such that they are aware of the application variability and the underlying device characteristics of silicon-photonic technology. We claim that such a cross-layer design enables a high-throughput and energy-efficient unified silicon-photonic link and main memory system. This thesis undertakes the cross-layer design with silicon-photonic technology in two fronts. First, we study the varying network bandwidth requirements across different applications and also within a given application. To address this variability, we develop bandwidth allocation policies that account for application needs and device sensitivities to ensure power-efficient operation of silicon-photonic links. Second, we design a novel architecture of an optically-controlled main memory system that is directly interfaced with silicon-photonic links using a novel read and write access protocol. Such a system ensures low-energy and high-throughput access from the processor to a high-density memory. To further address the diversity in application memory characteristics, we explore heterogeneous memory systems with multiple memory modules that provide varied power-performance benefits. We design a memory management policy for such systems that allocates pages at the granularity of memory objects within an application