The Edge computing paradigm seeks to bring Cloud-like compute capabilities close to the Edge of the network, next to where the data is generated, so as to minimize the data communication latency. Edge applications such as autonomous driving, surveillance for accident and crime detection, and robotics are latency sensitive. To ensure low end-to-end latency, the impact on latency of all layers of the computing stack needs to be considered. In this thesis, we investigate the impact of multiple applications sharing the memory on the compute latency. We present a comprehensive experimental evaluation of the impact of different types of co-located memory applications on the latency sensitive application. We choose YOLOv3, a deep learning based object recognition system as an example of a latency sensitive application. We synthesize microbenchmarks that capture the various characteristics of memory intensive background applications. We show that at a high memory utilization due to the co-located microbenchmark, YOLOv3 suffers a latency degradation of 20x compared to low memory utilization situations. To mitigate the impact on latency due to memory intensive applications, we propose and evaluate latency control strategies based on the recently available Pressure Stall Information feature in the Linux kernel. We show that using \textit{latd} our proposed user space latency controller, the latency constraints of YOLOv3 are not significantly violated despite the high memory pressure exerted by background memory intensive microbenchmarks. The thesis thus makes it possible to practically deploy latency sensitive applications along with memory intensive background applications on the same physical machine at the Edge while efficiently utilizing the memory.