next up previous
Next: Energy savings Up: Benchmarks and Experimental Results Previous: Usage trace collection

System power model

The Sharp Zaurus SL-5500 PDA has an Intel StrongARM SA-1110 processor [14], which can run in three modes: RUN, IDLE, and SLEEP. Moreover, its core clock speed can be varied between 59 and 206$MHz$ in discrete steps. It is put into the IDLE mode by the Linux kernel whenever there is nothing to run. We constructed a system power model based on power measurements and available industrial data. The measurement equipment is the same as that used in [28]. The modes of system operation and their corresponding power consumptions are shown in Fig. 3(b). Permitted mode transitions are shown as directed arcs. The delay overhead for mode transitions is as marked on the directed arcs between modes [14]. The front-light is assumed to be always off and the display is assumed to be always on with constant power consumption of $234.4mW$. We assume the power consumption is constant for an operation mode. Notably, the display consumes about half of the system idle power.

RUN corresponds to the SA-1110 RUN mode when the processor is busy executing instructions. Its power is measured when the PDA computes discrete cosine transforms, as detailed in [28]. IDLE206 and IDLE59 correspond to the SA-1110 IDLE mode with a core clock frequency/voltage of 206$MH$z/1.5V and 59$MH$z/1.25V, respectively. IDLE59 is a hypothetical mode for the Zaurus SL-5500, but is real for many other SA-1100 or SA-1110 based systems [9,22,26]. It is estimated as the display power plus the measured non-display power at 206$MHz$ scaled down using the same ratio of the power for IDLE59 to that for IDLE206 in the Itsy system power measurement presented in [26]. SLEEP-Display corresponds to the SA-1110 SLEEP mode when the display is left on. Therefore, its system power consumption is measured when the system is suspended, and the display power is added thereafter.

The energy consumption for system usage is obtained by running its trace through the system power model with DPM/DVS. There are different ways to do DPM/DVS with such a system power model in view of user delays, as discussed next.

The Linux kernel automatically puts the system into the IDLE206 mode whenever there is no process running and returns it to the RUN mode upon interrupts. We take this as the baseline and report the performance of other techniques against it.

The most straightforward DPM/DVS technique would be to put the system into the IDLE59 or the SLEEP-Display mode right after the system finishes responding to the user and put it back into the RUN mode upon a user input. These methods are called the $simple$ and $lazy$ techniques, respectively. Since the IDLE59 to RUN mode-transition delay is small, there is no concern with regard to user productivity. However, that for SLEEP-Display to RUN will most likely be noticed by users and decrease user productivity.

Assuming we can predict user delays absolutely accurately, we can choose to put the system into the SLEEP-Display mode and wake it up right before the next user input, if the delay is long enough. This is called the $perfect$ technique since it gives the upper bound on energy savings based on the system power model.

The proposed user delay models can be used to predict user delays. There are two concerns with respect to prediction errors. First, in the case of overestimation, the system will wake up from the SLEEP-Display mode upon an interrupt generated by the user input and enter the RUN mode directly. Such errors are called lazy errors. If the delay is overestimated by more than the human perceptual threshold, the user will experience a noticeable mode transition delay. Such lazy errors are called serious lazy errors. Second, in the case of underestimation, the system wakes up and transfers to the IDLE59 mode after the predicted delay to get ready for the user input and will not be able to fully exploit the idle time to reduce energy by remaining in the SLEEP-Display mode. Therefore, we report the average error for underestimations. We refer to the DPM/DVS technique based on user delay predictions by the history-based and psychological models as $history$ and $psychological$, respectively.


next up previous
Next: Energy savings Up: Benchmarks and Experimental Results Previous: Usage trace collection
Lin Zhong 2003-12-20