Efficiently and Transparently Maintaining High SIMD Occupancy in the Presence of Wavefront Irregularity

Demand is increasing for high throughput processing of irregular streaming applications; examples of such applications from scientific and engineering domains include biological sequence alignment, network packet filtering, automated face detection, and big graph algorithms. With wide SIMD, lightweight threads, and low-cost thread-context switching, wide-SIMD architectures such as GPUs allow considerable flexibility in the way application work is assigned to threads. However, irregular applications are challenging to map efficiently onto wide SIMD because data-dependent filtering or replication of items creates an unpredictable data wavefront of items ready for further processing. Straightforward implementations of irregular applications on a wide-SIMD architecture are prone to load imbalance and reduced occupancy, while more sophisticated implementations require advanced use of parallel GPU operations to redistribute work efficiently among threads. This dissertation will present strategies for addressing the performance challenges of wavefront- irregular applications on wide-SIMD architectures. These strategies are embodied in a developer framework called Mercator that (1) allows developers to map irregular applications onto GPUs ac- cording to the streaming paradigm while abstracting from low-level data movement and (2) includes generalized techniques for transparently overcoming the obstacles to high throughput presented by wavefront-irregular applications on a GPU. Mercator forms the centerpiece of this dissertation, and we present its motivation, performance model, implementation, and extensions in this work