high performance computing on graphics processing units: hgpu.org

Deep Learning (DL) model-based AI services are increasingly offered in a variety of predictive analytics services such as computer vision, natural language processing, speech recognition. However, the quality of the DL models can degrade over time due to changes in the input data distribution, thereby requiring periodic model updates. Although cloud data-centers can meet the computational requirements of the resource-intensive and time-consuming model update task, transferring data from the edge devices to the cloud incurs a significant cost in terms of network bandwidth and are prone to data privacy issues. With the advent of GPU-enabled edge devices, the DL model update can be performed at the edge in a distributed manner using multiple connected edge devices. However, efficiently utilizing the edge resources for the model update is a hard problem due to the heterogeneity among the edge devices and the resource interference caused by the co-location of the DL model update task with latency-critical tasks running in the background. To overcome these challenges, we present Deep-Edge, a load- and interference-aware, fault-tolerant resource management framework for performing model update at the edge that uses distributed training. This paper makes the following contributions. First, it provides a unified framework for monitoring, profiling, and deploying the DL model update tasks on heterogeneous edge devices. Second, it presents a scheduler that reduces the total re-training time by appropriately selecting the edge devices and distributing data among them such that no latency-critical applications experience deadline violations. Finally, we present empirical results to validate the efficacy of the framework using a real-world DL model update case-study based on the Caltech dataset and an edge AI cluster testbed.