MODA – A Framework for Memory Centric Performance Characterization

MODA is a memory-centric performance analysis tool specific for Cray XMT supercomputer. The tool was designed and developed at Pacifica Northwest National Laboratory when I interned in 2011. MODA is the first memory resource centric tool on Cray XMT. It can trace program variables and memory references, analyze data types, data  patterns and layouts. It can help to identify the architectural and structural hot spots of large-scale application on Cray XMT supercomputer systems.

There are four phases of using MODA: 1. Code Instrumentation; 2. Execution & Monitoring; 3. Trace Visualization & Post Analysis
; 4. Performance Optimization. MODA first use binary instrument technique and Cray’s compiler tool to instrument memory reference instructions to generate extra memory access information, such as timestamps, memory virtual/physical address, PID variable names etc.  In the execution and monitoring phase, MODA have a runtime to collect memory information and compress data to store to file system. MODA provides a post-analysis and GUI interface to visualize memory behavior of application and help to identify memory hot spots and optimize application.