Machine Learning Based Fast Thermal Analysis And Optimization For Multi-core Systems

Power density of integrated multi-core systems keeps increasing with technology scaling,causing severe thermal related problems,including system reliability and performance degradation issues.Many DTM(dynamic thermal management)method were proposed to solve the thermal problem.Runtime full chip temperature prediction is significant to dynamic thermal management.The existing study mostly focus on full chip thermal estimation,which strongly replied on the dynamic power,rather than full chip thermal prediction.And the proposed thermal prediction work can only predict local hotspot.There are few study can achieve full chip temperature prediction without dynamic power information.In order to solve the problem,we proposed a new fast thermal analysis technology called AFTP(ARMA based full-chip temperature prediction)and LFTP(LSTM based full-chip temperature prediction).The fast thermal analysis technology are consisted of sensor temperature prediction,predicted sensor temperature based thermal recovery and sensor placement.The only different between AFTP and LFTP is the sensor temperature prediction model.AFTP use ARMA(autoregressive moving average)based sensor temperature prediction model and LFTP use LSTM(long short-term memory)based sensor temperature model.As for thermal recovery,we use EigenMaps.We use sensor placement algorithm to find optimal number and position of thermal sensor,which can improve the accuracy of thermal recovery.AFTP and LFTP can achieve high quality full chip temperature prediction without dynamic power information and cost low computing overhead.After we get predicted thermal map of multi-core system,we still need to solve thermal control problem.In order to solve the nonlinear control problem,we proposed a new thermal control optimization method called compact piecewise linear(PWL)model based predictive control.First,a compact PWL thermal model,which takes dynamic power as input,is built by combining multiple local compact linear thermal models expanded at several Taylor expansion points.These local compact linear thermal models are obtained by sampling based model order reduction(MOR)with high accuracy.Their Taylor expansion points are selected by a systematic scheme which exploits the thermal behavior property of the multi-core chips.Based on the compact PWL thermal model,a new predictive control method is proposed to compute the future power recommendation for DTM.By approximating the nonlinearity accurately with the compact PWL thermal model and being equipped with predictive control technique,the new DTM achieves an overall high quality temperature management with smooth and accurate temperature tracking.Experimental results show the new method outperforms the linear model predictive control based method and the echo state network based predictive thermal management method in temperature management quality with lower computing overhead.