PDL Abstract

Towards Edge-caching for Image Recognition

First Workshop on Smart Edge Computing and Networking (SmartEdge) '17, held in conjunction with PerCom 2017, March 13 - 17, 2017, Hawaii, USA.

Utsav Drolia, Katherine Guo*, Jiaqi Tan, Rajeev Gandhi, Priya Narasimhan

Carnegie Mellon University
* Bell Labs, Nokia


With the available sensors on mobile devices and their improved CPU and storage capability, users expect their devices to recognize the surrounding environment and to provide relevant information and/or content automatically and immediately. For such classes of real-time applications, user perception of performance is key. To enable a truly seamless experience for the user, responses to requests need to be provided with minimal user-perceived latency.

Current state-of-the-art systems for these applications require offloading requests and data to the cloud. This paper proposes an approach to allow users’ devices and their onboard applications to leverage resources closer to home, i.e., resources at the edge of the network. We propose to use edge-servers as specialized caches for image-recognition applications.We develop a detailed formula for the expected latency for such a cache that incorporates the effects of recognition algorithms’ computation time and accuracy. We show that, counter-intuitively, large cache sizes can lead to higher latencies. To the best of our knowledge, this is the first work that models edge-servers as caches for compute-intensive recognition applications.