Addressing the heterogeneity challenges in many-core systems

In recent years, the microprocessor industry has evolved towards the many-core system era, and power has now become a first-order design constraint. Moreover, the continuing scaling of semiconductor devices has led to technology heterogeneity effects that make the power problem even more severe. Traditional design methods that do not consider variation effects can lead to gross inaccuracies. Accurate power models and tools that consider the technology heterogeneity effects early in the design stage are highly desirable. Furthermore, it is expected that more and more applications with diverse requirements will run simultaneously on such many-core systems. However, the application heterogeneity can cause severe resource contention problems and affect performance in unpredictable ways.;This thesis addresses the above two heterogeneity challenges faced by many-core system design: technology heterogeneity and application heterogeneity. First, it develops power modeling tools that take technology heterogeneity effects into account; Second, it introduces a unified quality-of-service (QoS) framework to tackle application heterogeneity challenges.;In the area of technology heterogeneity, this thesis proposes ORION 2.0, an architecture-level power modeling framework for networks-on-chip (NoCs). Early stage power estimation for NoCs has become extremely important as they now consume a significant amount of total chip power. ORION 2.0 was built on top of the original ORION 1.0 power modeling framework and extensively enhances the ORION 1.0 models. ORION 2.0 can be easily used by system-level designers and enables a wide range of power and thermal optimization opportunities in many-core system design.;The next part of this work presents PolarisPT, a system-level early stage design space exploration tool for NoCs. As on-chip resources keep increasing, the design choices for NoC architectures will continue expanding. Technology heterogeneity further increases the complexity of future designs. Because of this, tools that help designers prune the expanding design space of NoCs and aid designers in making design decisions in such highly heterogeneous environment become crucially important. PolarisPT helps the designer rapidly explore NoC power-performance trade-offs with respect to technology variation effects.;In the area of application heterogeneity, this thesis proposes a QoS architecture framework that enables the simultaneous management of three critical shared resources: cache, NoC and memory. The proposed framework is very effective in providing QoS-aware resource management with low overhead. We show that the proposed static QoS with joint resource managements achieves greater performance than an architecture that only consider individual QoS support.;Finally, this thesis proposes a dynamic QoS management framework to adaptively improve system-level performance. The primary motivation for this work is that applications exhibit runtime variations in resource demands and that static resource allocation may not utilize the allocated resources efficiently at all points in time. Therefore, we propose a framework that dynamically responds to runtime variation and reallocates shared resources among applications whenever suitable. Our proposed framework can significantly improve performance for lower-priority applications, while maintaining performance for high-priority applications.;As heterogeneity problems become more prominent in many-core system design, a complete methodology that can guide designers in designing power/performance efficient systems at design time and optimize the resource utilization at runtime is highly desirable. The models, tools, designs, and mechanisms presented in this thesis take a first step in addressing the heterogeneity challenges faced by future many-core system designers and pave the way for further optimization.