Computation offloading decisions for reducing completion time

Mobile devices are being widely used in many applications such as image processing, computer vision (e.g. face detection and recognition), wearable computing, language translation, and battlefield operations. However, mobile devices are constrained in terms of their battery life, processor performance, storage capacity, and network bandwidth. To overcome these issues, there is an approach called Computation Offloading, also known as cyber-foraging and surrogate computing. Computation offloading consists of migrating computational jobs from a mobile device to more powerful remote computing resources. Upon completion of the job, the results are sent back to the mobile device. However, a decision must be made; whether to execute a job locally or remotely, and consider the execution time of the job and the time to move the input and output data over a network (e.g. Internet). This dissertation presents a computation offloading decision analysis using our system model to identify the conditions under which computation offloading reduces job completion time. We show the results of our numerical experiments using our system model to identify when computation offloading reduces job completion time. We conduct the same analysis using a real workload trace. Lastly, we present a network delay estimate decision analysis to find the effect of the network delay estimate error on the job completion time, and determine how accurate a network delay estimator has to be. This work is backed up with Monte Carlo simulations and behavioral simulations using the discrete event network simulator NS-3. We have found that the bottleneck transmission rate has to be at least the same order of magnitude as the file size to benefit from computation offloading. We have observed that under poor network connections and using powerful hand-held devices, computation offloading of some applications to remote servers becomes less beneficial; however, cloudlets become more attractive. We have also identified real world applications that benefit from computation offloading from a real workload trace. Finally, we have identified the inflection point that impacts the average decision difference when adding error in the estimate of the communication cost