Network-Aware Task Scheduling for Edge Computing
Loading...
Authors
Shrestha, Bibek
Issue Date
2021
Type
Thesis
Language
Keywords
Edge Computing , INT , Programmable data plane , Task scheduling
Alternative Title
Abstract
Edge computing promises low-latency computation by moving data processing closer to the source. Tasks executed at the edge of the network have seen a significant increase in their complexity. The demand for low-latency computation for delay-sensitive applications at the edge is also increasing. To meet the computational demand, task offloading has become a go-to solution where the edge devices offload tasks in part or whole to the edge servers via the network. But the performance fluctuations of the network largely influence the data transfer performance between edge devices and the edge servers, which negatively impacts the overall task execution performance. Hence, monitoring the state of the network is desirable to improve the performance of task offloading at the edge. However, networks are usually dynamic and unpredictable in nature, particularly when the network is being used by multiple other devices and applications simultaneously, resulting in data flows competing with each other for the resources.
In this study, we are leveraging In-band Network Telemetry (INT) to collect fine-grained network information to introduce network awareness in task scheduling for edge computing. Legacy methods of network monitoring that rely on flow-level and port-level statistics are often limited by their collection frequency which is typically in the order of tens of seconds. In contrast, INT can improve the collection frequency by working at the line rate and granularity of information by capturing network telemetry at packet-level directly from the data plane. Such capabilities enable the detection of subtle changes and congestion events in the network, thereby increasing the network visibility while making it more accurate. We implemented a network-aware task scheduler for edge computing that uses high-precision network telemetry for task scheduling. We experimented with different workloads under various congestion scenarios to assess the impact of our network-aware scheduler on the task offloading performance. We observed up to 40% reduction in data transfer time and up to 30% reduction in the overall task execution time by favoring edge servers in uncongested or relatively less congested areas of the network when scheduling the tasks. Our study shows that network visibility is an important factor that can improve task offloading performance. The results so obtained supports our motivation to use INT for obtaining fine-grained high-precision network telemetry to create a network-aware task scheduler for edge computing.