University of California, Santa Barbara
When predicting application performance, allocating resources, and scheduling jobs, an accurate estimate of the required resources is essential. Although CPU and disk performance is relatively well understood, memory performance is often ignored or considered constant. The authors' research shows that memory bandwidth can vary by up to two orders of magnitude depending upon access pattern, read/write ratio, and number of cores accessing the memory. They believe that resource management can be improved by accounting for these factors, especially for data-intensive applications.