Research On Key Technologies Of Real-Time Data Object Scheduling For Cyber-Physical Systems

Cyber-Physical Systems(CPS)provide the possibility of deep integration between the real world and the computing world,which is becoming a new hotspot of next generation information technology.A large part of CPS is to support real-time applications,such as industrial automation system,smart robot control,health monitoring,traffic management and so on.The data objects collected in such CPS usually have real-time properties,that is,the collection,transmission and processing of these data objects needs to be completed in a timely fashion.The practical significance of real-time data objects is to model the current state of the physical world entities.Therefore,a real-time object is only valid within a given period of time.With the rapid development of such new CPS,new opportunities and challenges have been emerged in the field of real-time system design.Firstly,the CPS tends to use a distributed architecture,one immediate advantage is that it can distribute the system workload equally to multiple processing units,so that the overload phenomenon of a single processing unit seldom occurs.However,distributed real-time data processing can also cause some problems.For instant,a large amout of such real-time data has to be handled in embedded devices to reduce data transmission or processing delay.At this point,the increasing number of data collecting sensors have fuelled the big sensor data processing trend,while the data processing units,such as embedded devices,are usually resource-limited.Therefore,reducing the workload of real-time sensor data in the distributed CPS can improve the system performance.Moreover,the network transmission delay bettween distributed embedded processing unit can not be ignored,and such delay may be harmful to the stability of the running system.Therefore,it is more important to make reasonable use of system idle time to ensure the valid of real-time data objects.Secondly,with the increasing scale of CPS,since the battery capacity is limited in embedded devices,especially when the battery is not practicable to recharge or replace,the power comsumption cost in embedded devices deployed in the outside world has become a bottleneck of the CPS development.Unfortunately,simply reducing the update workload of sensor node may not save energy effectively.This means we should resort to well-established energy-efficient techniques,such as DVFS compatible approaches,to maximize energy saving.Hence,to reduce the pressure of heat dissipation and to prolong the operating life of the system,the development of energy-saving technologies to maintain the data freshness in CPS has practical research value.Most existing work focuses on the scheduling of independent systems and assumes that the network latency of sensor update transactions can be ignored,which is difficult to meet the requirements of large real-time data processing in modern CPS.We take a first step to address the dynamic scheduling problem of real-time data objects when transmission delays are considered,and we propose two new algorithms called JB-EDF and JB-EDF*.JB-EDF algorithm first adopts an initial parameters setting of each update transaction,and then adjusts these parameters accordingly.In order to handle more transactions,an improved JB-EDF*method is proposed.The algorithm attempts to quickly assemble a schedulable subset of transactions,and then processes the remaining transactions sequentially.To speed up processing data objects,we also introduce a series of techniques to improve the efficiency of the two algorithms.The experimental results show that JB-EDF and JB-EDF*are less affected by network delay fluctuation,and the runtime stability of both algorithms is better.The acceptance rate of JB-EDF*is much higher than the existing algorithms,so that it also has a wider range of applications.Research on energy-efficient scheduling for real-time data objects in CPS is still in its infancy.In this thesis,DVFS technology is introduced into the underlying scheduling of real-time data objects in CPS,and two energy-saving algorithms called ML-CS and ML-US are proposed for single-core systems.The ML-CS algorithm sets a constant slowdown factor for the update transaction sets,and reduces the execution frequency of each transaction as much as possible,thus reducing the energy consumption in CPS.ML-US algorithm sets different execution frequencies for different real-time transactions at a finer granularity level.By making a trade-off between system idleness and workload cost,it can make better use of the DVFS technology to achieve more energy-saving.The experimental results show that ML-CS and ML-US have obvious energy-saving effect on the whole(up to 60%).Both of the proposed algorithms have higher utilization of computing resources and better freshness for real-time data.With the popularity of multi-core processors,it is necessary to extend the power management methods of real-time data to multi-core CPS platforms.Despite the importance,related work in this field is relatively scarce.This thesis serves as the first attempt to solve the given problem.We prove that the energy-efficient scheduling of real-time data objects on multi-core CPS is also NP-hard problem.Due to the inherent intractability of the problem,we transform and divide it into two sub-problems,namely single-core energy-efficient data objects scheduling problem and transaction to processor mapping problem on multi-core,based on partitioned scheduling.Thus,the existing DVFS methods can be directly applied to each core for saving energy consumption.In addition,a multi-core mapping algorithm TCBM is proposed,which proves that the energy consumption on a single core is related to the density of update transactions assigned to it.It is beneficial for reducing system energy consumption by dispatching update transactions to different processors as equally as possible.Experimental results show that the TCBM algorithm has better energy-saving effect than other traditional multi-core allocation scheme(up to 35%).