Automatic synthesis and optimization of chip multiprocessors

Abstract

The microprocessor technology has experienced an enormous growth during the last decades. Rapid downscale of the CMOS technology has led to higher operating frequencies and performance densities, facing the fundamental issue of power dissipation. Chip Multiprocessors (CMPs) have become the latest paradigm to improve the power-performance efficiency of computing systems by exploiting the parallelism inherent in applications. Industrial and prototype implementations have already demonstrated the benefits achieved by CMPs with hundreds of cores.CMP architects are challenged to take many complex design decisions. Only a few of them are:- What should be the ratio between the core and cache areas on a chip?- Which core architectures to select?- How many cache levels should the memory subsystem have?- Which interconnect topologies provide efficient on-chip communication?These and many other aspects create a complex multidimensional space for architectural exploration. Design Automation tools become essential to make the architectural exploration feasible under the hard time-to-market constraints. The exploration methods have to be efficient and scalable to handle future generation on-chip architectures with hundreds or thousands of cores.Furthermore, once a CMP has been fabricated, the need for efficient deployment of the many-core processor arises. Intelligent techniques for task mapping and scheduling onto CMPs are necessary to guarantee the full usage of the benefits brought by the many-core technology. These techniques have to consider the peculiarities of the modern architectures, such as availability of enhanced power saving techniques and presence of complex memory hierarchies.This thesis has several objectives. The first objective is to elaborate the methods for efficient analytical modeling and architectural design space exploration of CMPs. The efficiency is achieved by using analytical models instead of simulation, and replacing the exhaustive exploration with an intelligent search strategy. Additionally, these methods incorporate high-level models for physical planning. The related contributions are described in Chapters 3, 4 and 5 of the document.The second objective of this work is to propose a scalable task mapping algorithm onto general-purpose CMPs with power management techniques, for efficient deployment of many-core systems. This contribution is explained in Chapter 6 of this document.Finally, the third objective of this thesis is to address the issues of the on-chip interconnect design and exploration, by developing a model for simultaneous topology customization and deadlock-free routing in Networks-on-Chip. The developed methodology can be applied to various classes of the on-chip systems, ranging from general-purpose chip multiprocessors to application-specific solutions. Chapter 7 describes the proposed model.The presented methods have been thoroughly tested experimentally and the results are described in this dissertation. At the end of the document several possible directions for the future research are proposed.