A case for compute reuse in future edge systems: An empirical study

Edge computing has been proposed as a solution to the ultra-low latency requirements of novel compute-intensive applications and systems such as autonomous driving, smart cities, etc. However, the scale of such compute-intensive systems coupled with stricter application requirements causes the computation load to exceed the current rate at which we provision the resources at the edge. In this paper, we advocate that reusing (partly or fully) the results of already executed computational tasks among multiple users to calculate new results, which we call "compute reuse", has the potential to significantly reduce resource utilization and lessen the time needed for the completion (i.e., execution) of new tasks. We conduct an experimental study to quantify the performance gain and improved utilization of resources that stem from systems being able to reuse previous computations. Our results demonstrate that such systems can result in up to 50× lower completion times and substantially lower utilization of computing resources. We also discuss alternative design options and tradeoffs of building systems aware of compute reuse semantics.