Replay-based Performance Analysis of Nonblocking Collective MPI Operations

The efficient use of large-scale high-performance computing systems requires identifying and eliminating communication bottlenecks, which often manifest as wait states. The Scalasca trace tools automatically detect and quantify such wait states through a replay-based analysis of event traces recorded by Score-P, tracing them back to their root causes and computing the critical path of the execution. While this analysis is well established for blocking collective MPI operations, it does not cover their nonblocking counterparts introduced with MPI-3.This thesis extends Scalasca's replay-based analysis to nonblocking collective MPI operations. The wait-state detection is implemented in Scalasca, and a theoretical concept for synchronization point tracking and the delay and critical-path analyses is developed. The extended tool is evaluated across a range of test cases, demonstrating its capability to identify wait states in nonblocking collective operations and showcasing its practical benefit.