CASITA: A Tool for Identifying Critical Optimization Targets in Distributed Heterogeneous Applications

Programming of high performance computing systems has become more complex over time. Several layers of parallelism need to be exploited to efficiently utilize the available resources. To support application developers and performance analysts we propose a technique for identifying the most performance critical optimization targets in distributed heterogeneous applications. We have developed CASITA, a tool which uses an execution trace and the knowledge about the programming models MPI, OpenMP and CUDA as well as their hierarchy among each other to build a distributed event dependency graph. After locating wait states in this graph, we detect their root cause and compute the critical path, an important property for performance optimizations. Compared to existing analysis approaches, we incorporate the hierarchy of multiple programming models and derive a metric from both the time an activity spends on the critical path and the waiting time it caused. For the purpose of visualization, CASITA enriches the input trace with additional counter information so that results can be inspected in the Vampir trace viewer.