Automated Compilation Framework for Scratchpad-based Real-Time Systems

Abstract

ScratchPad Memory (SPM) is highly adopted in real-time systems as it exhibits a predictable behaviour. SPM is software-managed by explicitly inserting instructions to move code and data transfers between the SPM and the main memory. However, it is a tedious job to decide how to manage the SPM and to manually modify the code to insert memory transfers. Hence, an automated compilation tool is essential to efficiently utilize the SPM. Another key problem with SPM is the latency suffered by the system due to memory transfers. Hiding this latency is important for high-performance systems. In this thesis, we address the problems of managing SPM and reducing the impact of memory latency. To realize the automation of our work, we develop a compilation framework based on the LLVM compiler to analyze and transform the program code. We exploit our framework to improve the performance of the execution of single and multi-tasks in real-time systems. For the single task execution, Worst-Case Execution Time (WCET) is of great importance to assure correct and safe behaviour of the system. So, we propose a WCET-driven allocation technique for data SPM that employs software prefetching to efficiently manage the SPM and to overlap the memory transfer and the task execution in a predictable way. On the other hand, multi-tasking requires the system to be schedulable such that all the tasks can meet their timing requirements. However, executing multiple tasks on a multi-processor platform suffers from the contention of the accesses to the shared main memory. To avoid the contention, several scheduling techniques adopted the 3-phase execution model which executes the task as a sequence of memory and computation phases. This provides the means to avoid the contention as well as to hide the memory latency by using a Direct Memory Access (DMA) engine. Executing memory transfers using the DMA allows overlapping the memory transfers with the computations on the processor. Using the 3-phase model in systems with limited sizes of local SPM may necessitate a segmentation of the task. Automating the segmentation process is necessary especially for systems with large task sets. Hence, we propose a set of efficient segmentation algorithms that follow the 3-phase execution model. The application of these algorithms shows a significant improvement in the system schedulability. For our segmentation algorithms to be more applicable, we extend the 3-phase model to allow programs with multiple paths represented as conditional Directed Acyclic Graphs (DAGs), unlike the previous works that targeted sequential programs. We also introduce a multi-steaming model to exploit the benefits of prefetching by overlapping the memory and computation phases of the same task, which was not allowed in the previous approaches. By combining the automated compilation with the proposed algorithms, we are able to achieve our goal to efficiently manage data SPM in real-time systems.