Managing Information Synchronicity in Real-Time Systems

Description

This project will study the design of real-time systems subject toinformation synchronicityconstraints. In a computing system, when multiple information flows join together (e.g., fordata fusion), the information synchronicity constraint requires that data from different flowscan join only if their source were produced at time points in close propinquity to each other.For example, in autonomous vehicles, perception of the external environment usually relies onfusion of measurements taken from multiple sensors. If measurements from two sensors, e.g.,a camera and a LiDAR, happened at two substantially different time points, the fusion of theirinformation will not be useful to reconstruct an accurate view of the environment.  This project will study how to formally prove the satisfaction of the information synchronicityconstraint, and optimize system design to better meet both information synchronicity andclassical real-time constraints, such as deadlines, tardiness bounds and data freshness bounds.Although the scheduling and analysis problems have been well studied for many classical real-timeconstraints, formal analysis and design optimization oriented to information synchronicityconstraints in real-time systems is still an open problem to our best knowledge.  The information synchronicity performance depends on two factors, (1) timing characteristicsof the generation of source sensor data and (2) the delays experienced by the information flowsinside the computing system. We will first study the problem only considering the 1st factor,and then refine the obtained results to take the 2nd factor into consideration. In this way, wewill have a more structured understanding of the influences by each of them and how to designgood strategies to deal with them. In another dimension, we will first focus on the localdatasynchronization policywhich decides how to match data frames from different informationflows before sending them to the data fusion algorithm, and then extend to the entire system toderive needed timing information and optimize the system design to better meet various timingconstraints. Specifically, we will pursue this project in the following steps:  1.Data synchronization policywithoutother real-time constraints.When the system isnot subject to real-time constraints, we can wait for arbitrarily long time to match data withthe closest timestamps. This excludes the influence of system-level timing behaviours andthus enables us to obtain a fundamental understanding of how to quantify and manage theinfluence of source sensor data’s timing characteristics to information synchronicity.  2.Data synchronization policywithother real-time constraints.When the system is alsosubject to other real-time constraints, the data synchronization policy can only postponethe decision for a limited amount of time. In this case, the information synchronicityperformance also depends on how unbalanced are the delays experienced by input data indifferent information flows. Therefore, we will study how to quantify the interrelationshipamong: (1) information synchronicity, (2) the latency incurred by the data synchronizationpolicy and (3) the delay unbalance among difference flows. 3.System-level timing analysis.As soon as we can quantify the interrelationship among theinformation synchronicity, the latency and the delay unbalance of input flows at each datafusion point, we will look into the big picture of the entire system to study how to analyseboth information synchronicity and other real-time performance of the system.  4.System-level design optimization.Based on the analysis techniques developed above, wewill design good system architectures and explore useful properties specific to optimizationof information synchronicity, to guide us efficiently search for high-quality solutions tomeet both information synchronicity and other real-time constraints of the system.Finally, we will build a tool prototype to implement and evaluate the new analysis techniques,and implement the developed data synchronization policies in ROS and Apollo Cyber RT toevaluate their runtime overheads. We will also collaborate with our industry partner to applythe developed techniques to the design optimization of an autonomous driving system.