A UNIFIED HARDWARE/SOFTWARE PRIORITY SCHEDULING MODEL FOR GENERAL PURPOSE SYSTEMS

Abstract

Migrating functionality from software to hardware has historically held the promise of enhancing performance through exploiting the inherent parallel nature of hardware. Many early exploratory efforts in repartitioning traditional software based services into hardware were hampered by expensive ASIC development costs. Recent advancements in FPGA technology have made it more economically feasible to explore migrating functionality across the hardware/software boundary. The flexibility of the FPGA fabric and availability of configurable soft IP components has opened the potential to rapidly and economically investigate different hardware/software partitions. Within the real time operating systems community, there has been continued interest in applying hardware/software co-design approaches to address scheduling issues such as latency and jitter. Many hardware based approaches have been reported to reduce the latency of computing the scheduling decision function itself. However continued adherence to classic scheduler invocation mechanisms can still allow variable latencies to creep into the time taken to make the scheduling decision, and ultimately into application timelines. This dissertation explores how hardware/software co-design can be applied past the scheduling decision itself to also reduce the non-predictable delays associated with interrupts and timers. By expanding the window of hardware/software co-design to these invocation mechanisms, we seek to understand if the jitter introduced by classical hardware/software partitionings can be removed from the timelines of critical real time user processes. This dissertation makes a case for resetting the classic boundaries of software thread level scheduling, software timers, hardware timers and interrupts. We show that reworking the boundaries of the scheduling invocation mechanisms helps to rectify the current imbalance of traditional hardware invocation mechanisms (timers and interrupts) and software scheduling policy (operating system scheduler). We re-factor these mechanisms into a unified hardware software priority scheduling model to facilitate improvements in performance, timeliness and determinism in all domains of computing. This dissertation demonstrates and prototypes the creation of a new framework that effects this basic policy change. The advantage of this approach lies within it's ability to unify, simplify and allow for more control within the operating systems scheduling policy.