System-Level Modeling and Analysis of 3D Multi-Processors on Chip for Future Cloud Computing

点击翻译按钮获取中文摘要

As our economy and society is mainly based on digital information, during the last decade an emergence of a digital universe occurred. As a consequence of this evolution, cloud computing and data centers have become ubiquitous - they are essential in nearly every sector of the economy and have become the backbone for communications, business, academic, and governmental systems. However, several challenges will be faced in the near future. With the new applications in the cloud and the massive data to be processed, the computing performance requirements are increasing dramatically. Today’s supercomputers’ power is expressed in several PFLOPS (1015 Floating Point Operations per Second) and 1MW is required for one PFLOP. The requirements for 2016 are estimated to exaFLOPS (1018 Floating Point Operations per Second) while, for economical and ecological reasons, the energy consumption needs to be limited to 20MW. The expected growth in computing power is two orders of magnitude larger than the acceptable increase in energy. A key resource to overcome this energy challenge is given by the new Integrated Circuits technologies: three-dimensional silicon integration will allow more than 1,000 times higher bandwidth communications at low power per channel using local interconnects between die layers and between die stacks. Integrated silicon nanophotonics will provide low-power and high-bandwidth optical interconnections between different parts of the system on a chip, board, and rack levels. A combination of these technologies will likely be required to build exascale systems. In this context, the advanced integration technologies enabled the emerging of microservers: a server may be replaced by a set of microservers, consuming less energy and providing a custom required computation power. The design of future microservers is still an open problem. One of the main open questions is related to the thermal characterization in microservers, since they are based on high-density packed chips where the heat dissipation is poor. Therefore, advanced analysis and modeling tools are required in order to assess the impact of this new technology in future cloud infrastructures. In this context, our proposal concentrates on the definition of an approach for system-level modeling and analysis of systems integrating advanced technologies with direct application to architectural and thermal optimizations in microservers.