Dynamic Stabilization Of MIPS Rates For Out-of-Order Processors

Embedded multimedia systems are soft real-time systems. Their computing processeshave to be predictable so that they can be adjusted adaptively. Out-of-order processorslaunch and execute multiple instructions in a cycle, by executing instructions in out-of-order ways, which makes it difficult to predict the computing times. This confines theirapplications in embedded real-time systems. Therefore, it is urgent to develop methods tostabilize the MIPS rates of out-of-order processors, which can be used to improve thepredictability of programs’ execution time.Based on analyzing limitations of dynamic stabilization of MIPS rates by PIDmethods with fixed coefficients, this thesis proposes a method based on the variableintegration of PID to adjust MIPS rates for out-of-order processors dynamically. Althoughout-of-order mechanisms can improve efficiency of instructions execution and resourceutilization, it also makes the procedure of instruction execution more complicated andunpredictable. It can be found that MIPS rates are highly variability and the procedure ofinstruction execution has great uncertainty. For example, the maximum MIPS rate is4times as high as the minimum one. Furthermore, it is rather common that the MIPS rateusually has a dramatic change when running a program. The methods based on PIDfeedback with fixed coefficients can stabilize MIPS rates when a program only changessmoothly, however when the program has sudden changes, the MIPS rate will notconverge to the target rate and just oscillate nearby the target value, due to theaccumulation of the system’s error caused by fixed integration coefficients. In order tosolve this problem, a PID feedback controller which uses variable integration is proposed.It adaptively adjusts the values of integration coefficients to get the correspondingfeedback values, which can be used to stabilize programs’ MIPS rates with non-stationarychanges for out-of-order processors through the mechanism of dynamic voltage andfrequency scaling (DVFS).Experimental results show that the proposed PID feedback control method withvariable integration can not only adjust programs with stationary MIPS rates, but alsomake the MIPS rate converge to the target value for the non-stationary MIPS rates after arelatively short period of time, which also achieves a good stabilization procedure. Bystabilizing the MIPS rates for out-of-order processors, the execution time of applicationscan be predicted more accurately. It has promising applications for both providing real-time requirement and improving effectiveness of adaptive scaling.