Research On Thermal Aware Loorplanning And Thread Mapping In CMP Processor

Power consumption increases dramatically with the improvement of semiconductor manufacturing technology. Energy consumed by the microprocessor is converted into heat which causes the processor temperature increasing continuously. When the chip temperature surpasses the threshold, processor becomes instable, the reliability is reduced significantly, and even physical damage might occur. Therefore the temperature becomes an important constraint for nowadays processor design. Researchers in literature exploit Dynamic Thermal Management (DTM) technology to leverage the temperature problem. But at the same time of reducing temperature, DTM will decrease the performance of processor as a result.Chip Multicore Processor (CMP) has become the mainstream in microprocessor manufacturing industry. In CMP processor, multiple threads run concurrently and consume more power than in traditional superscalar processor. Thus the thermal issue becomes more severe then before. Different ways of reducing CMP processor temperature have been proposed in academic and industry interature. Some of them include thread activity migration, temperature aware OS scheduling, efficient temperature throttling mechanism, temperature aware floorplanning, temperature aware compiling, and so on. CMP improves the thread level parallelism of application to thread scheduling, assign workload so even that thermal space is enhanced, reduces the complexity of design, and is able to reuse design.In this paper, a thermal aware floorplanning scheme for CMP processor and thread mapping policies are investigated in order to leverage the thermal problem. The objective is to reduce the time of triggering DTM so as to improve the processor performance as a result.The main contributions of this paper are as follows.First, a benchmark classifing scheme based on heat spreading model is proposed in this paper. Using our scheme, twenty-six programs in SPEC CPU2000 suite are classified into three categories according to the level of peak temperature of steady state: Hot, Warm, and Cool. Then the relationship between the thermal characteristic of benchmark and the program behavior is discussed. Experimental results show that benchmark classification based on heat spreading model reflects the thermal characteristic and program behavior of each benchmark correctly.Second, a thermal aware floorplanning method for CMP processor is proposed through rotating a single core's fioorplan. And the relationship between floorplan and tempearature is studyed. Experimental results show that reduction of peak temperature can be up to 0.31℃,0.09℃,0.36℃respectively when two threads of workload, three threads of workload, four threads of workload run on a Alpha-ev6 like quad-core processor.At last, the difference of peak temperature can be up to 3～4℃for the different mappings when an identical workload running on CMP. A themal aware thread mapping algorithm which can evenly distribute heat energy among the chip is proposed. We compare the peak temperature of the proposed mapping with more than hundred different randomly generated mappings and find that our scheme is effective to obtain a mapping which is close to lower bound of the range.