An Integer Linear Programming Approach for the Analysis of DTM Strategies

As the number of cores in multicore processors and their operating frequency continue to grow, processor power consumption increases and leads to an escalation in chip temperatures. This escalation causes today’s multicore processors’ performance to be limited by chip thermal constraints rather than process technology or circuit design. As a result, Dynamic Thermal Management (DTM) has an increasingly significant role in the design of new microprocessors. We propose a new way to design and quantitatively analyze DTM strategies. Using Integer Linear Programming, we compute optimal offline DTM strategies that achieve maximal performance and meet thermal constraints. By analyzing the optimal strategies, we are able to calculate the upper bound on DTM performance, to identify optimal strategy patterns and to compare the thermal limitations of several microprocessor layout designs. We employ offline Optimal DTM analysis on the case of Multicore Task Scheduling DTM. We find that this analysis suggests that a layout of many small cores is more thermally efficient than a layout of several large cores. The analysis further suggests specific task scheduling algorithm guidelines that maximize performance under thermal constraints. In addition, we compute and analyze the optimal multicore task scheduling strategies for DVS/DFS-based mechanism and Stop&Go DTMs —————————— ——————————