Scratchpad Memory Management For Multicore Real-Time Embedded Systems

Multicore systems will continue to spread in the domain of real-time embedded systems due to the increasing need for high-performance applications. This research discusses some of the challenges associated with employing multicore systems for safety-critical real-time applications. Mainly, this work is concerned with providing: 1) efficient inter-core timing isolation for independent tasks, and 2) predictable task communication for communicating tasks. Principally, we introduce a new task execution model, based on the 3-phase execution model, that exploits the Direct Memory Access (DMA) controllers available in modern embedded platforms along with ScratchPad Memories (SPMs) to enforce strong timing isolation between tasks. The DMA and the SPMs are explicitly managed to pre-load tasks from main memory into the local (private) scratchpad memories. Tasks are then executed from the local SPMs without accessing main memory. This model allows CPU execution to be overlapped with DMA loading/unloading operations from and to main memory. We show that by co-scheduling task execution on CPUs and using DMA to access memory and I/O, we can efficiently hide access latency to physical resources. In turn, this leads to significant improvements in system schedulability, compared to both the case of unregulated contention for access to physical resources and to previous cache and SPM management techniques for real-time systems. The presented SPM-centric scheduling algorithms and analyses cover single-core, partitioned, and global real-time systems. The proposed scheme is also extended to support large tasks that do not fit entirely into the local SPM. Moreover, the schedulability analysis considers the case of recovering from transient soft errors (bit flips caused by a single event upset) in several levels of memories, that cannot be automatically corrected in hardware by the ECC unit. The proposed SPM-centric scheduling is integrated at the OS level; thus it is transparent to applications. The proposed scheme is implemented and evaluated on an FPGA platform and a Commercial-Off-The-Shelf (COTS) platform. In regards to real-time task communication, two types of communication are considered. 1) Asynchronous inter-task communication, between either sequential tasks (single-threaded) or parallel tasks (multi-threaded). 2) Intra-task communication, where parallel threads of the same application exchange data. A new task scheduling model for parallel tasks (Bundled Scheduling) is proposed to facilitate intra-task communication and reduce synchronization overheads. We show that the proposed bundled scheduling model can be applied to several parallel programming models, such as fork-join and DAG-based applications, leading to improved system schedulability. Finally, intra-task communication is governed by a predictable inter-core communication platform. Specifically, we propose HopliteRT, a lean and predictable Network-on-Chip that connects the private SPMs