Optically-Connected Memory

Growing demands on future data centers and high-performance computing systems are driving the development of processor-memory interconnects with greater performance and flexibility than can be provided by existing electronic interconnects. A redesign of the systems' memory devices and architectures will be essential to enabling high-bandwidth, low-latency, resilient, energy-efficient memory systems that can meet the challenges of exascale systems and beyond. By leveraging an optics-based approach, this thesis presents the design and implementation of an optically-connected memory system that exploits both the bandwidth density and distance-independent energy dissipation of photonic transceivers, in combination with the flexibility and scalability offered by optical networks. By replacing the electronic memory bus with an optical interconnection network, novel memory architectures can be created that are otherwise infeasible. With remote optically-connected memory nodes accessible to processors as if they are local, programming models can be designed to utilize and efficiently share greater amounts of data. Processors that would otherwise be idle, being starved for data while waiting for scarce memory resources, can instead operate at high utilizations, leading to drastic improvements in the overall system performance. This work presents a prototype optically-connected memory module and a custom processor-based optical-network-aware memory controller that communicate transparently and all-optically across an optical interconnection network. The memory modules and controller are optimized to facilitate memory accesses across the optical network using a packet-switched, circuit-switched, or hybrid packet-and-circuit-switched approach. The novel memory controller is experimentally demonstrated to be compatible with existing processor-memory access protocols, with the memory controller acting as the optics-computing interface to render the optical network transparent. Additionally, the flexibility of the optical network enables additional performance benefits including increased memory bandwidth through optical multicasting. This optically-connected architecture can further enable more resilient memory system realizations by expanding on current error dectection and correction memory protocols. The integration of optics with memory technology constitutes a critical step for both optics and computing. The scalability challenges facing main memory systems today, especially concerning bandwidth and power consumption, complement well with the strengths of optical communications-based systems. Additionally, ongoing efforts focused on developing low-cost optical components and subsystems that are suitable for computing environments may benefit from the high-volume memory market. This work therefore takes the first step in merging the areas of optics and memory, developing the necessary architectures and protocols to interface the two technologies, and demonstrating potential benefits while identifying areas for future work. Future computing systems will undoubtedly benefit from this work through the deployment of high-performance, flexible, energy-efficient optically-connected memory architectures.