Low Delay Communication And Computation Resource Scheduling For Heterogeneous Distributed Computing

With the popularization of mobile devices and the development of the Internet of Things,many emerging applications are computation-intensive and latency-sensitive.The limitations of server computing resources and network communication resources have greatly promoted the integration and development of distributed computing and communication networks.Mobile Edge Computing(MEC)as a flexible and efficient distributed computing technology has received the widespread attention from all walks of life.User terminals can choose to offload computing tasks to edge servers due to their insufficient computing capabilities.After receiving the offloading tasks,the edge servers cooperate to perform distributed processing on them,which relieves the computing pressure of the terminal and greatly improves the efficiency of task execution.In view of the above research scenarios,how to jointly consider communication and computation to obtain the low-latency benefits is the main content of this paper.Motivated by this,our paper has carried out the research of resource scheduling strategy for heterogeneous distributed computing.First,the thesis based on the MEC computing offloading scenario is composed of multiple servers,considering the limited computing resources of edge servers and the limited network communication resources,a delay-optimal offloading strategy including the task assignment and the offloading order is proposed.First,considering the communication transmission and the computation process jointly,we construct a min-max offloading model with the optimal completion delay.Then,the optimal offloading scheduling polices are derived according to the task assignment is fair or adjustable.In the case of fair task allocation scenario,the cycles of tasks assigned to each MEC server are the same.The paper proposes the “Lowest-Computation-First”rule as the optimal off-loading order.Next,with the adjustable task assignment,the offloading scheduling policy jointly optimizes the task assignment and the offloading order.According to the optimality condition that all tasks can complete the computation at the same time,the optimal closed-form of task assignment is obtained.The objective function is transformed to the wasted computation resources,and the optimal offloading order with the “Highest-Communication-First”rule is proposed.Compared with other offloading scheduling algorithms,the proposed offloading scheduling policy with the joint optimization can obtain the best delay performance.Next,the thesis focuses on the collaborative computing process based on the Map Reduce framework between different edge servers,and a delay-optimal task allocation strategy which combines the Map and the Reduce is derived.First,three execution phases of Map,Shuffle,and Reduce require multiple communications and computations when processing the distributed tasks,where the computation load and the communication load will be affected by the task allocation at the same time.The thesis constructs a min-max task allocation model with the joint optimization of multiple phases,and proposes an iterative algorithm framework according to the competitiveness and monotonicity of the problem.Next,the eigenvalue equations for the problem are established and the iterative method is derived based on the Perron theorem.The optimal Map and Reduce task allocation for each edge server can be obtained when the algorithm reaches the convergence.The algorithm minimizes the overall task execution delay by optimizing the tradeoff between the computation delay and the communication delay.Compared with the other four benchmark algorithms,it has a great advantage in delay performance.Finally,for the delay-optimal resource scheduling algorithm under the Map Reduce framework proposed in Chapter 3,the thesis based on the Map Reduce architecture in Hadoop verifies the resource scheduling algorithm by executing the Word Count task,and clarifies the validity of the proposed algorithm in the distributed computing scenario.The thesis establishes a heterogeneous cluster environment composed of multiple computing nodes with different resource configurations.Then,the task allocation rule in Map Reduce is determined by the output results of the scheduling algorithm.Moreover,we run the Word Count instances of different sizes and record the corresponding delay.Simulation results shows that the proposed delay-optimal resource scheduling algorithm can obtain a better delay performance compared with other two benchmarks.