The deployment of sensors in urban settings is providing data critical for studies involving city planning and policymaking. Typically, the collected data is then sent to the cloud, which provides unlimited computational power for processing. Moving to the cloud, however, can require high network bandwidth and high latency. Edge computing – processing the collected data close to the source of that data – mitigates these disadvantages. For example, a user can process the data in situ and send to the cloud only what is needed for a particular study. Nevertheless, edge computing also has limitations: limited computing and storage resources and unreliable network connectivity to the cloud.
To address these limitations, researchers from Argonne National Laboratory, the University of Chicago and Northwestern University have developed a goal-driven scheduling model (see Fig. 1) that ensures that the user’s application has access to enough resources and can run reliably despite network unreliability. Their research was published in the Journal of Parallel and Distributed Computing in 2022 and was selected for a special issue on “Distributed Intelligence at the Edge for the Future Internet of Things.”
Fundamental to the proposed new scheduler is that it is goal driven. The researchers define a science goal as “one that captures the job description, including objectives, contexts of interest and system- and application-specific requirements.” Typically, in edge computing, a job scheduler is in the cloud and controls execution of jobs at the edge from the cloud. In contrast, the proposed goal-driven model comprises two parts: a cloud scheduler and an edge scheduler. The cloud scheduler accepts user jobs, generates science goals and decides when and where to distribute the jobs to the edge computing nodes. The edge scheduler runs on each edge node and decides what to run at a given time in order to accomplish given science goals.
“Our scheduling model is goal oriented: it delivers jobs whenever the user wants them to run. This makes a big difference when jobs need to be scheduled at the right time to capture events of interest,” said Yongho Kim, a postdoctoral appointee in Argonne’s Mathematics and Computer Science (MCS) division
An equally important feature of the new model is that it is context aware. This means that in scheduling, the model takes into consideration factors such as energy consumption, job priority and edge node locations. Context awareness is triggered by science rules – if-then statements derived from combining scientific facts and ways for accomplishing their scientific goal. For example, an arithmetic science rule might be env.temperature.outside > 35.6 => Hot(outside), and a corresponding logical science rule might be Hot(Outside) => Run(Airconditioner). The Run symbol here is the trigger for the edge scheduler to run the application. Users can design these rules to capture events of interest for their scientific application.
“Edge computing requires more than a simple first come, first served approach,” said Rajesh Sankaran, an experimental systems specialist in Argonne’s MCS division. “The scheduler should be able to respond dynamically, refining the current context using sensors on the edge computing nodes and evaluating the science rules based on new knowledge generated from raw sensor data.”
Smart City Test Results
To test the new model, the research team used a real-world case study involving urban traffic flow. Included in the study were sensor measurements of the number of cars passing an intersection, an estimate of the average speed of the cars, and sample still images. The results validated that the jobs were launched as intended: a science goal was generated and distributed to the edge node; the edge scheduler received the sensor measurements, loaded the science rules and validated the measurements; and the jobs were scheduled and completed successfully.
“The smart city study showed that scientists can specify their science goals using our scheduling model,” said Nicola Ferrier, a senior scientist in Argonne’s MCS division. “But the study also showed that our approach may require some learning practice and a greater understanding of sensor input and output. And it raised several interesting questions: How soon must a particular job be run? Would a tolerance time window make the scheduler more flexible in scheduling multiple jobs? How can one establish priorities between conflicting science goals? Should the science rules be more elaborate to better represent the science goals?
“This work has been a first step,” Ferrier said. “We will start looking at how we can help users better map information and take full advantage of edge computing for scientific applications at the edge.”
For the full paper in the special issue, see Y. Kim, S. Park, S. Shahkarami, R. Sankaran, N. Ferrier, P. Beckman, “Goal-driven scheduling model in edge computing for smart city applications,” Journal of Parallel and Distributed Computing, 167, issue C, Sept. 2022, pp. 97–108, https://doi.org/10.1016/jpdc.2022.04.024.
For information about the special issue, see the journal website.