Static Performance Evaluation for Memory-Bound Computing: The MBRAM Model

We present the MBRAM model for static evaluation of the performance of memory-bound programs. The MBRAM model predicts the actual running time of a memory-bound program directly from pseudo-code. This means that the final running time can be predicted even before the program has been developed and benchmarked. In contrast to the "Big Oh" complexity model, which measures the time solely by counting the number of instructions executed, the MBRAM model predicts running times based on memory accesses, cache parameters, RAM bandwidth, and other important architectural parameters. As a result, the MBRAM model correctly ranks orders the actual running times for implementations of seven different O(n log n) sorting algorithms. In our suite of benchmarks, the MBRAM model consistently underestimates the actual running times, with errors ranging from 10% to 44%.

• Type: Conference
• Year: 2004
• Where: PDPTA
• Authors: Gene Cooperman, Xiaoqin Ma, Viet Ha Nguyen
• Keywords: Actual Running Times, MBRAM Model, PDPTA 2004, PDPTA 2007, Running Times