Design and Analysis of Dynamic Thermal Management in Chip Multiprocessors (CMPs)

Chip Multiprocessors (CMPs) have been prevailing in the modern microprocessor market. As the significant heat is converted by the ever-increasing power density and current leakage, the raised operating temperature in a chip has already threatened the system?s reliability and led the thermal control to be one of the most important issues needed to be addressed immediately in chip designs. Due to the cost and complexity of designing thermal packaging, many Dynamic Thermal Management (DTM) schemes have been widely adopted in modern processors. In this study, we focus on developing a simple and accurate thermal model, which provides a scheduling decision for running tasks. And we show how to design an efficient DTM scheme with negligible performance overhead. First, we propose an efficient DTM scheme for multimedia applications that tackles the thermal control problem in a unified manner. A DTM scheme for multimedia applications makes soft realtime scheduling decisions based on statistical characteristics of multimedia applications. Specifically, we model application execution characteristics as the probability distribution of the number of cycles required to decode frames. Our DTM scheme for multimedia applications has been implemented on Linux in two mobile processors providing variable clock frequencies in an Intel Pentium-M processor and an Intel Atom processor. In order to evaluate the performance of the proposed DTM scheme, we exploit two major codecs, MPEG-4 and H.264/AVC based on various frame resolutions. Our results show that our DTM scheme for multimedia applications lowers the overall temperature by 4 degrees C and the peak temperature by 6 degrees C (up to 10 degrees C), while maintaining frame drop ratio under 5% compared to existing DTM schemes for multimedia applications. Second, we propose a lightweight online workload estimation using the cumulative distribution function and architectural information via Performance Monitoring Counters (PMC) to observe the processes dynamic workload behaviors. We also present an accurate thermal model for CMP architectures to analyze the thermal correlation effects by profiling the thermal impacts from neighboring cores under the specific workload. Hence, according to the estimated workload characteristics and thermal correlation effects, we can estimate the future temperature of each core more accurately. We implement a DTM scheme considering workload characteristics and thermal correlation effects on real machines, an Intel Quad-Core Q6600 system and Dell PowerEdge 2950 (dual Intel Xeon E5310 Quad-Core) system, running applications ranging from multimedia applications to several benchmarks. Experiments results show that our DTM scheme reduces the peak temperature by 8% with 0.54% performance overhead compared to Linux Standard Scheduler, while existing DTM schemes reduce peak temperature by 4% with up to 50% performance overhead.