Research On Energy-aware Multi-objective Virtual Resource Management In The Cloud Data Center

In recent years,the demand for cloud computing has grown rapidly.To address this problem,cloud computing operators are investing more and more in cloud infrastruc-tures,which is rapidly expanding the quantity and size of cloud data centers.Today,energy consumption has become one of the biggest tests facing the cloud computing industry.It is reported that the electricity bill of a cloud data center accounts for 60%to 70%of its total operating cost.In other words,as long as the energy consumption can be effectively reduced,it can greatly reduce the data center operating costs,and bring huge economic benefits to cloud computing operators.The technology of virtual resource management is an effective means to reduce en-ergy consumption in cloud data centers.This article studies the energy-saving technol-ogy of cloud data centers from the perspective of virtual machine(VM)placement,and virtual resource scheduling.Based on a large number of literature reading,we find that there are still many problems worth discussing and studying in the above areas:the VM placement algorithm in the traffic-intensive environment ignores where the VMs arc placed.If the VMs are placed dispersedly,they will lead to serious energy consump-tion issues;Only a few multi-objective VM placement,algorithms consider the selection of scalarized weight for each objective;Many budget constrained workflow scheduling algorithms do not consider how to meet the budget constraint of the workflow.Some of the above algorithms discuss the success rate;The partial critical path strategy,which has been widely used to solve the deadline constrained workflow scheduling,cannot reflect the changes that occur in the process of scheduling.In real life,virtual resource management problems usually need to consider multiple optimization goals at the same time,so we only consider multi-objective problems in this paper.Most of the virtual resource management problems are NP-hard,so we design algorithms based on ant colony system(ACS)and reinforcement learning(RL).The main research contents are as follows: Excessive traffic requests and limited bandwidth resources are the most impor-tant contradictions in traffic intense data centers.In other words,you cannot place all VMs in the data center at the same time.In order to make full use of bandwidth resources,we can only try to keep the total communication revenue of each time period to the maximum.The previous work only takes into account the communication revenue without considering where the VMs are placed.If these VMs are placed dispersedly,it will waste a lot of energy.In response to the above problems in traffic intense data centers,we propose an ACS-based bandwidth-constrained VM placement(ACS-BVMP)algorithm to maximize the total communication revenue and minimize energy consumption.To solve these problems,we consider communication revenue while placing VMs on the mini-mum number of PMs.The hierarchical tree topology of the data center is too complex,so we simplify the bandwidth constraint based on the property of the topology.In addition,we also put forward a metric for evaluating the quality of the solution.Finally,we compare ACS-BVMP with three multi-objective algo-rithms and two single-objective algorithms.The simulation results show that the performance of ACS-BVMP algorithm is higher than that of other algorithms.To deal with high energy consumption and low resource utilization in the cloud data center,we propose a VM placement algorithm based on multi-objective RL(VMPMORL).VM placement is a typical NP-hard problem,which has attracted much attention from researchers.However,many multi-objective VM placement algorithms ignore the reasonable selection of scalarization weights.Although weight selection is difficult,ignoring this problem will greatly reduce the quality of the solution set.To solve the above problem,we use the Chebyshev scalarization function to scalarize the vector of Q values,which can greatly reduce the difficulty of weight selection.Compared to the existing multi-objective RL algorithms in the field of VM placement,VMPMORL not only considers weight selection but also is designed based on the Pareto set.Finally,we compare VMPMORL with several heuristics.The experimental results show that the VMPMORL algorithm is superior to other algorithms in terms of energy consumption,resource waste,and three quality indicators. Many budget-constrained workflow scheduling algorithms do not consider the budget constraint and the feasibility of solutions.Only a few of them can return feasible solutions.These algorithms distribute the budget constraint of the work-flow to each task,and can only prove that the post-conversion constraints are suficient conditions for the original budget constraints.Besides,how to choose the scalarized weights is often overlooked in these algorithms.To deal with these problems and the increasing energy consumption of the scientific workflow in clouds,we propose a simple way to obtain a feasible solution and designs an energy-aware multi-objective RL(EnMORL)algorithm to reduce the energy con-sumption and makespan.Compared with the existing methods that can meet the budget constraint,our method is completely different,and there is no need for complicated transformation and proof.The characteristics of scientific workflow scheduling and RL are consistent,and the Chebyshev scalarization function can solve the weight selection problem effectively,so we combine them to design the scheduling algorithm.Finally,we compare EnMORL with two state-of-the-art multi-objective scheduling algorithms.The simulation results show that En-MORL outperforms the above two algorithms.The PCP strategy used in the IaaS cloud partial critical paths with deadline dis-tribution(IC-PCPD2)algorithm calculates the sub-deadline of each task before the scheduling algorithm begins.These values will be no longer updated,so they cannot reflect changes that occur during the scheduling process.Besides,the deadline-constrained workflow scheduling algorithms often do not consider the selection of scalarized weights.In response to these problems and the increased energy consumption of scientific workflows in clouds,we propose a modified PCP(MPCP)strategy and a deadline-constrained multi-objective RL(DCMORL)workflow scheduling algorithm.We aim to minimize the energy consumption and cost of the workflow,Unlike PCP,sub-deadlines in MPCP are updated con-stantly,which can accurately reflect changes that occur during scheduling.To solve the problem of weight selection,we design DCMORL based on the Cheby-shev scalarization function.Finally,we compare DCMORL with several state-of-the-art scheduling algorithms.The simulation results show that,the DCMORL algorithm is superior to these algorithms in terms of energy consumption,cost,and hypervolume.