Adapting Application Mapping to Systematic Within-Die Process Variations on Chip Multiprocessors

摘要

Process variations, which lead to timing and power variations across identically-designed components, have been identified as one of the key future design challenges by the semiconductor industry. Using worst case latency/power assumptions is one option to address process variations. This option, while simplifying the problem, is becoming less and less attractive as its performance and power costs keep increasing. As a result, exploring options that allow the software to have knowledge about the actual latency/power consumption values is critical for future systems. Targeting systematic process variations, this paper makes two contributions. First, we discuss how we can assign threads to the cores of a chip multiprocessor (CMP) with process variations in mind and show the energy-delay product (EDP) benefits such a process variation-aware thread mapping can bring. Second, we study the benefits of varying the frequencies on a subset of the cores to increase EDP savings. We propose and evaluate integer linear programming based thread mapping schemes in both studies. While these schemes operate with profile data, they can be made to work with partial profiling as well with the help of curve fitting. We tested our schemes using both sequential and multi-threaded benchmarks from different suites and the results collected indicate that we can achieve EDP savings as much as 73.4%, with an average saving of 37.1% over a process variation agnostic scheme.