Architecting Energy Efficient Servers.

Abstract

This dissertation investigates how energy efficient servers can be architected using current and future technology. We leverage recent trends in packaging and device technology to deliver low power and high throughput. Specifically at the package level, this dissertation looks at 3D stacking technology that has emerged as a promising solution in achieving energy efficiency by delivering high throughput at a low cost. It shows how one would leverage this new technology into a datacenter. 3D stacking technology can be used to implement a simple, low-power, high-performance chip multiprocessor suitable for throughput processing. Our proposed architecture leveraging this technology, PicoServer, employs 3D technology to bond one die containing several simple slow processing cores to multiple memory dies sufficient for a primary memory. The multiple memory dies are composed of DRAM. 3D stacking technology also enables wide low-latency buses between processors and memory. These remove the need for an L2 cache allowing its area to be re-allocated to additional simple cores. The additional cores allow the clock frequency to be lowered without impairing throughput. Lower clock frequency along with the integration of non-volatile memory in turn reduces power and means that thermal constraints, a concern with 3D stacking, are easily satisfied. The PicoServer architecture targets server applications,which exhibit a high degree of thread level parallelism. An architecture targeted to efficient throughput is ideal for this application domain. At the memory device level, this dissertation investigates how the system memory could be re-architected to reduce the rising power consumption of system memory and disk drives. Flash memory has emerged as a strong candidate to reduce system memory power while remaining cost effective than conventional system memory. This dissertation discusses how Flash could be integrated at the system level and provides insights on the architectural support for Flash in servers. Our architecture uses a two level disk cache composed of a relatively small DRAM, which includes a primary disk cache, and a Flash based secondary disk cache. Further, based on our observations, we found that the Flash based disk caches should be split into a read optimized disk cache and write optimized disk cache.