MODA A Framework for Memory Centric Performance Characterization

In the age of massive parallelism, the focus of performance analysis has switched from the processor and related structures to the memory and I/O resources. Adapting to this new reality, a performance analysis tool has to provide a way to analyze resource usage to pinpoint existing and potential problems in a given application. This requires (1) memory trace collection with minimal perturbation of the application’s behavior; (2) data management of multiple gigabyte and terabyte size trace files; (3) efficient data analysis and visualization of traces; and (4) the introduction of the target architecture’s memory model into the analysis module for a truly memory-centric view. These features enable an application developer to anticipate algorithmic and structural resource bottlenecks on a small scale before a full scale roll out into production. This paper provides an overview of the Memory Observant Data Analysis (MODA) tool, a memory-centric tool first implemented on the Cray XMT supercomputer along with the solution to above mentioned challenges. Throughout the paper, MODA’s capabilities have been showcased with experiments done on matrix multiply and Graph-500 application codes.