Doctor of Philosophy

dissertationThe computing landscape is undergoing a major change, primarily enabled by ubiquitous wireless networks and the rapid increase in the use of mobile devices which access a web-based information infrastructure. It is expected that most intensive computing may either happen in servers housed in large datacenters (warehouse- scale computers), e.g., cloud computing and other web services, or in many-core high-performance computing (HPC) platforms in scientific labs. It is clear that the primary challenge to scaling such computing systems into the exascale realm is the efficient supply of large amounts of data to hundreds or thousands of compute cores, i.e., building an efficient memory system. Main memory systems are at an inflection point, due to the convergence of several major application and technology trends. Examples include the increasing importance of energy consumption, reduced access stream locality, increasing failure rates, limited pin counts, increasing heterogeneity and complexity, and the diminished importance of cost-per-bit. In light of these trends, the memory system requires a major overhaul. The key to architecting the next generation of memory systems is a combination of the prudent incorporation of novel technologies, and a fundamental rethinking of certain conventional design decisions. In this dissertation, we study every major element of the memory system - the memory chip, the processor-memory channel, the memory access mechanism, and memory reliability, and identify the key bottlenecks to efficiency. Based on this, we propose a novel main memory system with the following innovative features: (i) overfetch-aware re-organized chips, (ii) low-cost silicon photonic memory channels, (iii) largely autonomous memory modules with a packet-based interface to the proces- sor, and (iv) a RAID-based reliability mechanism. Such a system is energy-efficient, high-performance, low-complexity, reliable, and cost-effective, making it ideally suited to meet the requirements of future large-scale computing systems