Predictable multi-processor system on chip design for multimedia applications

TLDR: This dissertation presents predictable architectural components for MPSoCs, a Predictable MP soC design strategy, automatic platform synthesis tool, a run-time system and an MPSoC simulation technique to design and manage these multi-processor based systems efficiently.

Abstract: The design of multimedia systems has become increasingly complex due to consumer requirements. Consumers demand the functionalities offered by a huge desktop from these systems. Many of these systems are mobile. Therefore, power consumption and size of these devices should be small. These systems are increasingly becoming multi-processor based (MPSoCs) for the reasons of power and performance. Applications execute on these systems in different combinations also known as use-cases. Applications may have different performance requirements in each use-case. Currently, verification of all these use-cases takes bulk of the design effort. There is a need for analysis based techniques so that the platforms have a predictable behaviour and in turn provide guarantees on performance without expending precious man hours on verification. In this dissertation, techniques and architectures have been developed to design and manage these multi-processor based systems efficiently. The dissertation presents predictable architectural components for MPSoCs, a Predictable MPSoC design strategy, automatic platform synthesis tool, a run-time system and an MPSoC simulation technique. The introduction of predictability helps in rapid design of MPSoC platforms. Chapter 1 of the thesis studies the trends in modern multimedia applications and processor architectures. The chapter further highlights the problems in the design of MPSoC platforms and emphasizes the need of predictable design techniques. Predictable design techniques require predictable application and architectural components. The chapter further elaborates on Synchronous Data Flow Graphs which are used to model the applications throughout this thesis. The chapter presents the architecture template used in this thesis and enlists the contributions of the thesis. One of the contributions of this thesis is the design of a predictable component called communication assist. Chapter 2 of the thesis describes the architecture of this communication assist. The communication assist presented in this thesis not only decouples the communication from computation but also provides timing guarantees. Based on this communication assist, an MPSoC platform generation technique has been presented that can design MPSoC platforms capable of satisfying the throughput constraints of multiple applications in all use-cases. The technique is presented in Chapter 3. The design strategy uses three simple steps for platform design. In the first step it finds the required number of processors. The second step minimizes the communication interconnect between the processors and the third step minimizes the communication memory requirement of the platform. Further in Chapter 4, a tool has been developed to generate CA-based platforms for FPGAs. The output of this tool can be used to synthesize platforms on real hardware with the help of FPGA synthesis tools. The applications executing on these platforms often exhibit dynamism e.g. variation in task execution times and change in application throughput requirements. Further, new applications may often be added by consumers at run-time. Resource managers have been presented in literature to handle such dynamic situations. However, the scalability of these resource managers becomes an issue with the increase in number of processors and applications. Chapter 5 presents distributed run-time resource management techniques. Two versions of distributed resource managers have been presented which are scalable with the number of applications and processors. MPSoC platforms for real-time applications are designed assuming worst-case task execution times. It is known that the difference between average-case and worst-case behaviour can be quite large. Therefore, knowing the average case performance is also important for the system designer, and software simulation is often employed to estimate this. However, simulation in software is slow and does not scale with the number of applications and processing elements. In Chapter 6, a fast and scalable simulation methodology is introduced that can simulate the execution of multiple applications on an MPSoC platform. It is based on parallel execution of SDF (Synchronous Data Flow) models of applications. The simulation methodology uses Parallel Discrete Event Simulation (PDES) primitives and it is termed as "Smart Conservative PDES". The methodology generates a parallel simulator which is synthesizable on FPGAs. The framework can also be used to model dynamic arbitration policies which are difficult to analyse using models. The generated platform is also useful in carrying out Design Space Exploration as shown in the thesis. Finally, Chapter 7 summarizes the main findings and (practical) implications of the studies described in previous chapters of this dissertation. Using the contributions mentioned in the thesis, a designer can design and implement predictable multiprocessor based systems capable of satisfying throughput constraints of multiple applications in given set of use-cases, and employ resource management strategies to deal with dynamism in the applications. The chapter also describes the main limitations of this dissertation and makes suggestions for future research.