TimeTrial: A low-impact performance profiler for streaming data applications

Finding performance bottlenecks in application-specific systems is becoming increasingly labor-intensive as the capabilities of these systems improve. The complex platforms required to meet today's high application performance demands put pressure on developers to sustain current design cycles. Application developers need better tools to diagnose performance issues that arise when utilizing real-world application-specific platforms, from embedded applications to high-performance computational science applications. In this paper, we present TimeTrial, a runtime performance monitoring system that profiles streaming data applications deployed on architecturally diverse computers. TimeTrial is designed to operate with minimal impact on the executing application, exploiting user directives to aggressively perform lossy compression on performance meta-data. We present the design of the TimeTrial performance monitor and demonstrate its use in discovering performance bottlenecks in a production computational biology application.