Automated Analysis of Task-Parallel Execution Behavior Via Artificial Neural Networks

We present an automated analysis technique that leverages artificial neural networks to identify possible causes for sub-optimal execution of task-parallel programs. Performance anomalies in task-parallel programs are often extremely difficult to analyze due to the complexity of the interactions between dynamic runtime systems and hardware. While Hardware Performance Monitoring is a common technique to capture hardware behavior, understanding how the resulting hardware event profiling data relates to task performance is often non-trivial and time-consuming. In this work, we present an automated technique for task-parallel performance analysis that identifies the hardware behaviors that have the greatest impact on task performance. Our technique uses artificial neural networks to model these relationships, allowing for isolation of the specific hardware events that have the most impact to slow down task execution. We show that our technique provides new insights into task-parallel execution behavior, allowing for acceleration of the performance optimization process.