Workload Dependent Mitigation Approaches for Performance VariabilityКНИГИ » АППАРАТУРА
Название: Workload Dependent Mitigation Approaches for Performance Variability: Ensuring Timing Guarantees of Integrated Circuits Автор: Ji-Yung Lin, Michalis Noltsis, Dimitrios Soudris, Francky Catthoor Издательство: Springer Год: 2025 Страниц: 186 Язык: английский Формат: pdf (true), epub Размер: 25.9 MB
This book provides a holistic view of workload-dependent mitigation techniques for performance variability. The authors describe the use of design-time profiling information to reduce the uncertainties in future execution time calculation at run time, thereby offering the best option for minimizing system costs while reducing missed deadlines. Readers are introduced to an approach that combines dynamic voltage and frequency scaling (DVFS) with heterogeneous datapaths (HDP), enabling users to tackle performance variability of multiple timescales down to the sub-millisecond level.
Performance variability, the phenomenon that the execution time of the same application varies every time it is executed, is ubiquitous in modern computers. Performance variability is especially detrimental to time-sensitive and time-critical applications, which are the applications required to be completed before a certain deadline. With continuous development toward smaller devices and more dynamic architectures and software, performance variability increases and threatens the reliability of both time-sensitive and time-critical applications.
To reduce performance variability, real-time mitigation approaches are used in the online operation of various modern systems and platforms, including processors, memory organizations, and communication networks. Mitigation approaches use on-chip monitors to detect the amount of variability at run time. Based on the amount of variability and the prediction of future workload, a run-time controller changes the speed of the system or the platform by altering system knobs. Among all kinds of mitigation approaches, workload-dependent approaches operate on the principle of using the design-time profiling information to limit and reduce the uncertainties in future execution time calculation at run time. Therefore, only workload-dependent approaches can ensure full timing guarantees, so they are the most promising option for tackling performance variability while minimizing system costs.
This book elaborates on the concept and realization of a workload-dependent mitigation approach tackling various performance variability. We have improved the ways of tackling performance variability from the architecture and software levels based on two features, heterogeneous datapaths (HDP) and the concept of dynamic scenarios. Furthermore, we have developed a cross-layer reliability and variability mitigation framework to simultaneously ensure functional correctness and timing guarantees.
With the rapid development of the semiconductor transistor technology, electronic devices made of Very Large-Scale Integration (VLSI) chips become necessities in our daily lives, with countless functions such as computation, communication, entertainment, etc. Ever since the early age of VLSI chips in the 1970s, the number of transistors per area approximately doubled every year, following Moore’s Law. In the modern era, a typical chip can contain billions of transistors. The continuous transistor downscaling has fueled great improvement in the chip performance and cost. The huge number of transistors also enables complicated system designs. A highly integrated system-on-chip (SoC) design can contain hundreds of processor cores, multiple levels of memory, and sophisticated interconnection networks all inside a single chip. Timing guarantee is the main focus of this book among all reliability requirements. Among the diverse applications of VLSI chips, most applications care about the timeliness of delivering the calculation results.
In Chap. 2, we give an overview of variability in integrated circuits, in which every source of performance variability in each layer is addressed. Moreover, we discuss the general principles of the approaches to tackling variability, including guardbanding and run-time mitigation. In Chap. 3, we elaborate on the state-of-the-art mitigation approaches for performance variability, in which each component (monitors, system knobs, and adaptive control algorithms) is discussed in detail. In Chap. 4, we present the concept of system and adaptive scenarios, as well as deploy a PID mitigation scheme on a fine workload granularity, applying Dynamic Voltage and Frequency Scaling (DVFS) actuations. In Chaps. 5 and 6, we present the workload-dependent multi-timescale approach we have developed for tackling performance variability based on the concept of dynamic scenarios. Chapter 5 introduces the theoretical basis and the design of this approach, while Chap. 6 shows the simulation flow and the results. Thereafter, we propose the cross-layer reliability and variability mitigation framework in Chaps. 7 and 8. The principles and theoretical basis are described in Chap. 7, and the simulation results are discussed in Chap. 8. Finally, Chap. 9 summarizes the main messages and contributions of this book and discusses the future perspective of this work.
Скачать Workload Dependent Mitigation Approaches for Performance Variability
Навигация
Вход на сайт
Реклама
Workload Dependent Mitigation Approaches for Performance Variability 
