The proposed work aims to present the functional components of MATE (Monitoring, Analysis and Tuning Environment) and DMA (Dynamic Modeler and Analyzer).

MATE supports dynamic automatic tuning of parallel/distributed applications. During run-time MATE automatically instruments a running application to gather information about the application’s behavior (through 2 components: Monitor and Dynamic Monitoring library DMLib). The analysis phase receives events, searches for bottlenecks by applying a performance model and determines solutions for overcoming the bottlenecks (Analyzer component). Finally, the application is dynamically tuned by applying the encountered solution (Tuner component). Moreover, we provide a separate Dynamic Tracer tool based on Monitor and DMLib that generates trace files for MPI applications.

The DMA tool models the execution flow through high-level application structures such as loops and communication operations and characterizes them with statistical execution profiles. The tool automatically extracts models of MPI applications using DynInst. The modeling is performed while the application runs and it is scalable to thousands of nodes. To track the communication, the tool uses a MPI piggyback component to pass additional information over MPI messages. The automated analysis observes this online model to quickly identify the most important performance problems and correlate them with application source code. This technique is also able to discover causal dependences between the problems and infer root causes of synchronization inefficiencies.