WoMA: An Input-Based Learning Model to Predict Dynamic Workload of Embedded Applications

Embedded applications often have real-time requirements and thus meeting the user-specified deadline are essential for good user experience. Frequency (down) scaling is a widely used technique to improve energy efficiency but may lead to violations of the user-specified deadline. Thus, a workload prediction model is of critical importance to guide the frequency scaling. In this letter, we propose WoMA , a supervised learning model to predict dynamic workload (execution cycles) of embedded applications based on their input data. WoMA is built on our key observation that the workload of some embedded applications is primarily determined by their input size. Using the "input size" as features, we apply a logistic regression method to construct WoMA trained and tested using five popular embedded applications: 1) adpcm; 2) aes; 3) dfsin; 4) blowfish; and 5) gsm. Since frequency scaling is typically performed at discrete levels, for each application, we classify its workload into five distinct classes. For a given test input, WoMA will predict the class of its corresponding workload. Cycle-accurate simulation results show that 100% of WoMA predictions are accurate. Thus, we use WoMA to guide frequency scaling, resulting in power saving by 9.9%-61.8% across these benchmark applications.