Resource Management for Reconfigurable Computing Systems

Next generation computing devices, from cell-phones and personal entertainment devices to medical devices and supercomputers, will continue to demand increased computation, reduced energy consumption, and improved user interfaces. In the past, these demands were met with microprocessor improvements; process technology scaling, improved single thread CPU performance, and multiple CPU cores per chip all contributed to performance gains. However, computing devices increasingly require resources specialized for compute intensive code such as digital signals processors (DSPs), graphics processing units (CPUs), application specific integrated circuits (ASICs), and reconfigurable hardware (RH) to meet performance, power and interactivity demands.

In multi-tasking systems, multiple applications compete for access to these resources, and thus sharing the resources effectively is a key component of meeting goals of system performance, allocation fairness, quality of service, energy consumption, or some combination of these. To efficiently support the changing needs of many, competing applications, the system may dynamically bind computation to either CPU or reconfigurable resources --- choosing at runtime whether the system is better served by executing the computation on the CPU or in reconfigurable logic. Dynamic binding allows the system considerable flexibility to optimize the resource utilization, but also adds complexity --- the ordering and frequency of changing the reconfigurable resource allocation affects the system benefit that can be achieved.

Efficient management and use of the reconfigurable resources is influenced by allocation quality and frequency. In turn, allocation quality is influenced by the system's ability to efficiently change allocations, the timeliness (and usefulness) of the allocations, and the contention for access to the resources. This thesis proposes several important components of resource management for reconfigurable computing systems that improve system performance by specifically targeting these three factors that affect allocation quality. To support efficient multi-tasking resource management, we present methods to provide preemption-like behavior that avoids expensive context save and restore in RH execution when sharing RH amongst multi-tasking applications, RH allocators that improve allocation quality through awareness of application context switches, and application thread schedulers that improve RH allocation quality by co-scheduling applications to reduce contention for the RH resources.