| Description |
In recent years, a number of trends have started to emerge, both in microprocessor and application characteristics. As per Moore's law, the number of cores on chip will keep doubling every 18-24 months. International Technology Roadmap for Semiconductors (ITRS) reports that wires will continue to scale poorly, exacerbating the cost of on-chip communication. Cores will have to navigate an on-chip network to access data that may be scattered across many cache banks. The number of pins on the package, and hence available off-chip bandwidth, will at best increase at sublinear rate and at worst, stagnate. A number of disruptive memory technologies, e.g., phase change memory (PCM) have begun to emerge and will be integrated into the memory hierarchy sooner than later, leading to non-uniform memory access (NUMA) hierarchies. This will make the cost of accessing main memory even higher. In previous years, most of the focus has been on deciding the memory hierarchy level where data must be placed (L1 or L2 caches, main memory, disk, etc.). However, in modern and future generations, each level is getting bigger and its design is being subjected to a number of constraints (wire delays, power budget, etc.). It is becoming very important to make an intelligent decision about where data must be placed within a level. For example, in a large non-uniform access cache (NUCA), we must figure out the optimal bank. Similarly, in a multi-dual inline memory module (DIMM) non uniform memory access (NUMA) main memory, we must figure out the DIMM that is the optimal home for every data page. Studies have indicated that heterogeneous main memory hierarchies that incorporate multiple memory technologies are on the horizon. We must develop solutions for data management that take heterogeneity into account. For these memory organizations, we must again identify the appropriate home for data. In this dissertation, we attempt to verify the following thesis statement: "Can low-complexity hardware and OS mechanisms manage data placement within each memory hierarchy level to optimize metrics such as performance and/or throughput?" In this dissertation we argue for a hardware-software codesign approach to tackle the above mentioned problems at different levels of the memory hierarchy. The proposed methods utilize techniques like page coloring and shadow addresses and are able to handle a large number of problems ranging from managing wire-delays in large, shared NUCA caches to distributing shared capacity among different cores. We then examine data-placement issues in NUMA main memory for a many-core processor with a moderate number of on-chip memory controllers. Using codesign approaches, we achieve efficient data placement by modifying the operating system's (OS) page allocation algorithm for a wide variety of main memory architectures. |
