CN113515351A - Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization - Google Patents
Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization Download PDFInfo
- Publication number
- CN113515351A CN113515351A CN202111041530.7A CN202111041530A CN113515351A CN 113515351 A CN113515351 A CN 113515351A CN 202111041530 A CN202111041530 A CN 202111041530A CN 113515351 A CN113515351 A CN 113515351A
- Authority
- CN
- China
- Prior art keywords
- virtual machine
- cloud
- data center
- qos
- cloud task
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/44—Arrangements for executing specific programs
- G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
- G06F9/45533—Hypervisors; Virtual machine monitors
- G06F9/45558—Hypervisor-specific management and integration aspects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/48—Program initiating; Program switching, e.g. by interrupt
- G06F9/4806—Task transfer initiation or dispatching
- G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
- G06F9/4881—Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5083—Techniques for rebalancing the load in a distributed system
- G06F9/5088—Techniques for rebalancing the load in a distributed system involving task migration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/44—Arrangements for executing specific programs
- G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
- G06F9/45533—Hypervisors; Virtual machine monitors
- G06F9/45558—Hypervisor-specific management and integration aspects
- G06F2009/4557—Distribution of virtual machine instances; Migration and load balancing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization. The method comprises the following steps: constructing a cloud task arrival queuing model of multiple virtual machines in a cloud computing data center environment; extracting QoS characteristics of a data center by using a stacking noise reduction automatic encoder technology to obtain a matrix describing QoS characteristic information after dimension reduction, and solving the maximum response time of the current virtual machine so as to perfect the constraint condition of a collaborative optimization objective function; and combining the cloud task arrival queuing model, the collaborative optimization objective function and the Lyapunov optimization method to obtain a resource scheduling algorithm based on the Lyapunov optimization theory, and adopting the resource scheduling algorithm to realize resource scheduling based on energy consumption and QoS collaborative optimization. The invention effectively reduces the energy consumption of the data center while guaranteeing the QoS, and overcomes the interference of the fluctuation of the cloud task arrival to the optimization problem solution in the real scene of the cloud computing data center.
Description
Technical Field
The invention belongs to the field of energy-saving scheduling of cloud computing, and particularly relates to a resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization.
Background
Cloud computing has long been a popular research project in the global IT field by virtue of ITs ultra-large scale service capability. With the continuous development of cloud computing technology, more and more data centers are emerging in the global scope, and the energy consumption generated by the infrastructure also presents an exponential growth situation. The carbon emission of the current global IT industry accounts for 3-5% of the total carbon emission of the world. According to recent reports, Google data centers consume nearly 3 hundred million watts, while Facebook data centers consume 6000 million watts. Data centers consume more power than high energy manufacturing. The international agency for research, mckentin, investigated 70 large data centers and found that the average power consumption for computer operation was only 6% to 12%. With the continuous increase of the scale of the data centers, the influence of the large amount of electric energy consumed by the data centers on the environment is increasingly prominent.
The virtual machine migration technology can improve the resource utilization rate of the physical nodes, and the total energy consumption value of the cloud computing system can be reduced by closing the servers in the idle state of the data center. Conventional energy saving solutions of this type also have significant adverse effects. When the CPU utilization rate approaches 100%, the performance of the virtual machine is significantly reduced, which means that there is inevitably a conflict relationship between two optimization goals, i.e., energy saving and quality of service (QoS). Excessive pursuit of any of these metrics can hinder the optimization of the other metric. In an actual data center scenario, due to load fluctuation and uncertainty factors, dual objectives of energy saving and QoS guarantee become more complex. Previous optimization methods required all entities (users, cloud tasks, service providers, etc.) to meet a single QoS constraint throughout the cloud computing scheduling process, which is impractical in a practical cloud computing environment. And the solving process of the optimization methods is complex, the convergence speed is low, and the real-time scheduling requirement of the large-scale cloud computing data center is difficult to meet.
Disclosure of Invention
In order to realize better balance between energy consumption and QoS optimization, the invention provides a resource scheduling realization method based on energy consumption and QoS cooperative optimization, and energy consumption generated by a data center is minimized on the premise of guaranteeing QoS of a user. The invention constructs a cloud task arrival queuing model of multiple Virtual Machines (VMs), constructs an energy consumption model and a target optimization function of a data center on the basis, and solves a collaborative optimization problem in each scheduling time slice by combining an optimization method of Lyapunov (Lyapunov) stability theory, so that a cloud task queue of the virtual machines can reach an allowable maximum value, the processing capacity of the virtual machines can be fully exerted, and the utilization of resources is realized.
The purpose of the invention is realized by at least one of the following technical solutions.
A resource scheduling implementation method based on energy consumption and QoS collaborative optimization comprises the following steps:
s1, constructing a cloud task arrival queuing model of multiple Virtual Machines (VMs) in a cloud computing data center environment;
s2, extracting QoS characteristics of the data center by using a stacking noise reduction automatic encoder technology to obtain a matrix describing QoS characteristic information after dimension reduction, and obtaining the maximum response time of the current virtual machine through the matrix to perfect the constraint condition of a collaborative optimization objective function;
s3, combining a cloud task arrival queuing model, a collaborative optimization objective function and a Lyapunov (Lyapunov) optimization method, obtaining a cloud task queue length condition which should be possessed by a virtual machine meeting energy consumption and QoS (quality of service) optimization scheduling, obtaining a resource scheduling algorithm based on the Lyapunov (Lyapunov) optimization theory, and solving the resource scheduling problem of energy consumption and QoS collaborative optimization into each time slice by adopting the resource scheduling algorithm to realize resource scheduling based on energy consumption and QoS collaborative optimization.
Further, in step S1, the cloud task arrival queuing model is formed by connecting a host queuing model and a Virtual Machine (VM) queuing model in series, and is used to optimize a relationship between a virtual machine cloud task queue backlog length and system energy consumption;
in the host queuing model, after the cloud task is submitted to the data center, the data center adopts a load balancing strategy of a least load (least loaded) rule to preferentially distribute the cloud task to the host with the least number of uncompleted cloud task requests, and thus the cloud task is formed to the host with the least number of uncompleted cloud task requestsQueuing model with exponentially distributed inter-arrival time and exponentially distributed cloud task execution time (a)A queuing model);
in the Virtual Machine (VM) queuing model, each Virtual Machine (VM) instance forms a cloud task with the interval time of arrival of exponential distribution and the execution time of exponential distribution, and the number of virtual machines isVirtual machine capacity ofQueue model of (1: (Queuing model), the cloud task is submitted to the virtual machine queuing model by the data center after being distributed by the host queuing model, and then is processed in a first-come-first-serve (FCFS) mode.
Further, when the cloud task enters the host queuing model, if at least one available Virtual Machine (VM) is on the host, the data center allocates the cloud task to the Virtual Machine (VM) currently in the idle state for execution; when all the virtual machines on the host computer are occupied by the cloud tasks and fully loaded, the newly arrived cloud tasks enter a queue buffer area of a virtual machine queuing model, and the queue buffer area is a first-in first-out (FIFO) queue specially used for storing the cloud tasks waiting to be executed; after the cloud task is executed and leaves a certain Virtual Machine (VM), the data center distributes the cloud task at the head of the queue buffer area to the Virtual Machine (VM); after the cloud task obtains the access right to the Virtual Machine (VM), the cloud task can be immediately provided with cloud service;
considering the running time of the data center as being composed of a plurality of continuous time slices, wherein the length of each time slice is defined as t; therefore, in time slot t, the queuing model of a single virtual machine: (Queuing model) is defined by the following equation:
wherein the content of the first and second substances,andis the queue length of the kth Virtual Machine (VM) in the jth and jth +1 time slices;is shown in the jth time sliceMiddle k virtual machineThe number of cloud tasks processed;is the jth time sliceMiddle k virtual machineThe number of cloud tasks in the queue; when in useWhen the queue length in a Virtual Machine (VM) is less than its processing powerThe queue length of the Virtual Machine (VM) may reach a minimum value of 0;
the length of a Virtual Machine (VM) queuing model can directly influence the running state of a host in a data center; the operating states of the host can be classified into the following two types:
1) active state: a cloud task queue of a Virtual Machine (VM) on a host is not empty, and a cloud task is waiting to be processed;
2) an idle state: a cloud task queue of a Virtual Machine (VM) on a host is empty, and the host in an idle state can be regarded as being in a sleep mode and in a low power consumption state, and generally speaking, the power consumption generated by the host in the idle state is a constant value;
the power consumption model of a data center can therefore be described in the form:
in the formula (I), the compound is shown in the specification,representing the total energy consumption of the current data center, wherein the first item on the right side of the energy consumption equation represents the total power consumption generated by the physical host machine in the current active state; p is the number of physical hosts in operation,is the firstPower consumption of individual physical hosts in an active state; q is the number of physical hosts in the idle state,is the firstPower consumption of each physical host in an idle state; the power consumption in the idle state can be regarded as a constant; m is the maximum number of Virtual Machines (VMs) currently available in the data center;
when the resource utilization rate is higher, the number of the physical hosts in the idle state is less, and the power consumption is also less; thus equation (2) can be expressed in the following form:
whereinIs as followsThe power consumption of the individual physical hosts in the active state,,is a constant power consumption of the host in the idle state,is in the jth time sliceQueue length of individual virtual machines.
Further, in step S2, processing resource occupation history data of the virtual machine in the data center by using a stacking noise reduction automatic encoder technology, and extracting feature information affecting a QoS index of the data center through robustness of a stacking noise reduction encoder;
using a high-dimensional matrix composed of a parameter set of a virtual machine as data of a network input layer, constructing a target function of a stacking noise reduction automatic encoder, and performing dimension reduction and data reconstruction on the high-dimensional matrix through the target function to obtain optimized characteristics;
in the process of minimizing the loss function of the encoder, the complexity of the stacked noise reduction automatic encoder is also used as one index to participate in the training process and used as the loss functionA constraint term for a number; adding regular terms based on weight attenuation after loss function to improve generalization effect of the stacked noise reduction automatic encoder to avoid overfitting, and adoptingAndcross entropy ofAs a loss function, the robustness of the stacked noise reduction auto-encoder is further enhanced.
Further, in the cloud environment, the requirements of tasks submitted by different users are different, some tasks tend to be completed within the shortest time, and some tasks only need to be completed within a period of time. For tasks with longer completion time limit, the tasks can be considered to be distributed to the virtual machine with weaker performance to be executed, and the virtual machine with stronger performance preferentially ensures the tasks with shorter completion time limit, so that the benefits of all users can be guaranteed as much as possible. The traditional scheduling method requires all entities in the whole cloud computing environment to meet a single QoS constraint target, which often fails to meet the requirements of a real cloud computing environment.
The invention provides a QoS service capability discovery method based on a stacking noise reduction automatic encoder, which is used for constructing a QoS constraint condition in an objective function. The automatic encoder is an unsupervised neural network model, and the basic idea is to enable an encoding layer (a hidden layer) to learn the implicit characteristics of input data, and the learned new characteristics can be used for reconstructing original input data through a decoding layer. Therefore, the automatic encoder performs the work of feature dimension reduction and feature learning.
The stacked noise reduction automatic encoder (SDAE) is an improvement on the basis of an automatic encoder, and aims to learn more robust features; except for a conventional encoding stage and a conventional decoding stage, a noise reduction automatic encoder carries out random damage processing on data information before encoding, and the damage processing mainly adds noise into original input data; obtaining an uncontaminated true input by adding noise to the training data and causing the encoder to learn to remove this noise; this therefore forces the encoder to learn and extract more robust features in the input data.
In the process of executing the cloud task, when the response time of the cloud taskWhen the length is too long, SLA vision can be caused; in the process of training the stacking noise reduction automatic encoder, a high-dimensional matrix formed by a parameter set of the virtual machine is used as data of a network input layer of the stacking noise reduction automatic encoder, and dimension reduction and data reconstruction are carried out on the high-dimensional matrix through the stacking noise reduction automatic encoder to obtain optimized QoS characteristic information;
the parameter set of the virtual machine comprises cloud task response time, queue length, virtual machine CPU utilization rate, bandwidth, CPU number, memory size, memory occupancy rate, virtual machine migration times, virtual machine migration time overhead and SLA violation rate of the virtual machine at the previous moment; when the data center is provided withWhen the virtual machine is in a starting state, in each scheduling period interval, the parameter set of the virtual machine is combined to formHigh dimensional matrix of;
The network structure of the stacked noise reduction automatic encoder comprises an input layer, a damaged input layer, a hidden layer and an output layer; in addition to the conventional encoding and decoding stages, the stacked noise reduction auto-encoder performs a random corruption process on the data before encoding in order to add noise to the training data, thereby forcing the encoder to learn and extract better QoS profile information from the input layer data; order toRepresenting a sample of the original input data,representing corrupted data after the addition of gaussian noise,andrepresenting the weights of the encoder and the decoder respectively,andrepresenting the bias term, the coding function of the stacked noise reduction automatic encoder (SDAE) encodes the original input to obtain a new feature representationThe encoding process is as follows:
whereinIs a function of the sigmoid and is,it is used as a non-linear deterministic mapping; similarly, the decoding function will derive a characterization from the hidden layerInput reconstruction into original inputThe decoding process is as follows:
the goal of training a stacked noise reduction autoencoder is to optimize the parameter setTo minimizeA reconstruction error therebetween; order toRepresenting loss function of stacked noise reduction auto-encoder (SDAE), optimization objective function of noise reduction auto-encoderCan be expressed as:
in the process of minimizing the loss function, the complexity of the stacking noise reduction automatic encoder needs to be considered, and an over-fitting is easily caused by an excessively complex model, so that the complexity of the model is also used as one index to participate in the training process, and the model is further constrained; adding a weight attenuation-based regularization term (L2 Norm) after the loss function to improve the generalization effect of the model to avoid overfitting; the weight attenuation is a coefficient placed in front of the regular term, and the influence of the complexity of the model on the loss function can be adjusted;
by usingAndcross entropy ofAs a loss function;(ii) a WhereinFor stacked noise reduction auto-encodersjAt one time slice, the firstiA squared loss function value of a weight vector of samples, wherein,, Tis the total number of time slices;d is the total number of samples of the input layer of the stacked noise reduction automatic encoder;a weight adjustment coefficient that is a regularization term;(ii) a The resulting loss function is thus obtained as follows:
according to the above equation, the loss optimization function of the stacked noise reduction auto-encoder can be expressed as:
the optimization of the objective function can be solved by using a quasi-Newton method for estimating the parametersBy using the obtained estimated parameters as the parameters of the network model of the stacking noise reduction automatic encoder, a low-dimensional matrix reflecting the QoS characteristic information of the virtual machine can be obtained;
Maximum response time with cloud taskAs an evaluation criterion of the data center service quality; order toRepresenting a maximum response time of a Virtual Machine (VM);represents a maximum response time specified by a Service Level Agreement (SLA);QoS characteristic information matrix representing a stacked noise reduction auto-encoder (SDAE) through a current Virtual Machine (VM)Andcalculating the maximum response time of the cloud task on the virtual machine according to the cosine similarity; in order to ensure the quality of service of the cloud task,should be less thanAnda minimum value in between; thus, overall co-optimization objectivesThe scalar function can be written as:
further, in step S3, analyzing the queue length of the cloud task reaching the queuing model by using Lyapunov (Lyapunov) optimization theory, and fitting the cloud task execution time at the current time by using QoS characteristic information to improve the time constraint condition of data center resource scheduling;
when in useIn time, the resource scheduling algorithm should reduce the number of non-empty task queues of a Virtual Machine (VM) to the maximum extent to reduce the energy consumption generated by a data center; at the same time, whenThe length of the cloud task queue should be kept large enough to preferentially make the utilization rate of the virtual machine as full as possible, and preferentially make the utilization rate of the virtual machine as full as possible, so as to fully release the processing capacity of the virtual machine.
Further, Lyapunov (Lyapunov) equation is commonly used to analyze the stability of the queuing model, and the Lyapunov (Lyapunov) equation expression may be expressed in the form of:
whereinIs the queue length of the kth virtual machine in the jth time slice,is the value of the optimization objective function of the stacking noise reduction automatic encoder in the jth time slice;to stack the lyapunov equation values of the optimization objective function of the noise reduction auto-encoder in the jth time slice,,Tis the total number of time slices,, Nfor the maximum number of virtual machines available in the current data center, in each time sliceThe lyapunov drift function is defined as follows:
in order to maintain the stability of the cloud task queue in the cloud computing fluctuation environment, the result of equation (12) should be converged and the drift value should be maintainedThe smaller the value of the final result is, the more stable the cloud task queue is;
on the premise of ensuring the service quality of the cloud task, in order to further reduce the energy consumption of the physical host, namely, the number of the virtual machines running in the virtual machine queuing model is reduced to the minimum; the objective function of the co-optimization problem can be rewritten as:
wherein V is a unit conversion coefficient which is used for converting the energy consumption into corresponding cloud task queue length units; according to the Lyapunov optimization method, a resource scheduling collaborative optimization objective function of the data center is obtained from a formula (13):
the solution to this function is as follows:
transforming equation (15) to obtain an equivalent equation of the form:
whereinIs the firstThe penalty value of the lyapunov drift plus penalty function for each virtual machine, the first solution of equation (16) is:
as shown in the formula (18),(ii) a When in useIn time, the resource scheduling algorithm must satisfy the following conditions:(ii) a Thus whenIn time, the resource scheduling algorithm should reduce the number of non-empty task queues of the VM to the maximum extent; at the same time, whenIn time, the resource scheduling algorithm preferentially enables the utilization rate of the virtual machine to tend to or be in a full load state, so that the processing capacity of the virtual machine is fully released.
Further, a resource scheduling algorithm based on Lyapunov (Lyapunov) optimization theory is executed by the data center in each preset scheduling cycle of the cloud environment, and the execution process of the resource scheduling algorithm is as follows:
if there isThe cloud task arrives at the data center within the period,is the maximum response time of the cloud task,virtual machineThe processing power of (a) is set,for virtual machinesThe current cloud task queuing queue length; queuing a queue for a current cloud taskFinding the queue with the shortest queue length(ii) a If it is notAnd then:
whereinFor the number of cloud tasks currently pending,to be made available to a virtual machineThe number of cloud tasks to process;
if it is notThen give an order(ii) a And slave queues that have been assigned to virtual machinesIn the process of (1), removing the substrate,repeating the steps until the length of the cloud task queuing queue is 0; the resource scheduling algorithm is to arrange the resource utilization rates of the virtual machines with the non-empty queues in an ascending order, so that the virtual machines of the queues meeting the Lyapunov optimization condition are selected from the virtual machines with the non-empty queues to be migrated, the virtual machine with the lowest utilization rate reaches the maximum resource utilization rate limit, and the resource scheduling algorithm can realize the balanced utilization of resources and reduce the number of overloaded hosts.
Compared with the prior art, the invention has the following advantages and technical effects:
1. the resource scheduling framework for energy consumption and QoS collaborative optimization in the cloud computing environment is provided, energy consumption generated by a data center can be minimized on the premise of guaranteeing QoS of a user, the solution of the collaborative optimization problem is specifically carried out in each time slice, and the energy efficiency problem caused by fluctuation of cloud task arrival can be more efficiently dealt with.
2. A multi-virtual-machine cloud task arrival queuing model is constructed, the host queuing model and the virtual machine queuing model are connected in series through an M/M/∞ + M/M/C/N queuing structure, and the efficiency of scheduling and distributing after the cloud task arrives at a data center can be improved.
3. The efficient data center QoS feature extraction method is provided, more feature information influencing the QoS of the data center is excavated through the robustness of the stacked noise reduction encoder, and the cloud task execution time at the current moment is obtained through fitting, so that the time constraint condition of data center resource scheduling is completed.
4. The method combines a queuing model, a deep noise reduction automatic encoder network and a Lyapunov (Lyapunov) stability theory, and provides a resource scheduling algorithm based on the Lyapunov (Lyapunov) optimization theory, so that the QoS is guaranteed, the energy consumption of a data center can be effectively reduced, and the total time overhead of data center scheduling is greatly optimized.
Drawings
FIG. 1 is a framework diagram of collaborative optimization of a cloud computing platform according to an embodiment of the present invention;
fig. 2 is a diagram of a QoS feature extraction framework of a noise reduction-based automatic encoder in a data center according to an embodiment of the present invention.
Detailed Description
In the following description, technical solutions are set forth in conjunction with specific figures in order to provide a thorough understanding of the present invention. This application is capable of embodiments in many different forms than those described herein and it is intended that similar embodiments will be apparent to those skilled in the art without the use of inventive faculty, and that the invention will be within its scope.
The terminology used in the description is for the purpose of describing particular embodiments only and is not intended to be limiting of the description. As used in one or more embodiments of the present specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used in one or more embodiments of the present specification refers to and encompasses any and all possible combinations of one or more of the associated listed items.
It should be understood that although the terms first, second, etc. may be used herein to describe various information in one or more embodiments of the specification, these information should not be limited by these terms, which are used only for distinguishing between similar items and not necessarily for describing a sequential or chronological order of the features described in one or more embodiments of the specification. Furthermore, the terms "having," "including," and similar referents, are intended to cover a non-exclusive scope, such that a process, method, system, article, or apparatus that comprises a list of steps or modules is not necessarily limited to the particular details set forth, but may include other inherent information not expressly listed for such steps or modules.
Example (b):
a resource scheduling implementation method based on energy consumption and QoS collaborative optimization comprises the following steps:
s1, constructing a cloud task arrival queuing model of multiple Virtual Machines (VMs) in a cloud computing data center environment;
as shown in fig. 1, the cloud task arrival queuing model is formed by connecting a host queuing model and a Virtual Machine (VM) queuing model in series, and is used for optimizing the relationship between the virtual machine cloud task queue backlog length and the system energy consumption;
in the host queuing model, after the cloud task is submitted to the data center, the data center adopts a load balancing strategy of a minimum load (least loaded) rule to preferentially distribute the cloud task to the host with the minimum number of uncompleted cloud task requests, and thereby, a queuing model (a queuing model with exponentially distributed cloud task arrival interval time and exponentially distributed cloud task execution time) is formedA queuing model);
in the Virtual Machine (VM) queuing model, each Virtual Machine (VM) instance forms a cloud task with the interval time of arrival of exponential distribution and the execution time of exponential distribution, and the number of virtual machines isVirtual machine capacity ofQueue model of (1: (Queuing model), the cloud task is submitted to the virtual machine queuing model by the data center after being distributed by the host queuing model, and then is processed in a first-come-first-serve (FCFS) mode.
When a cloud task enters a host queuing model, if at least one available Virtual Machine (VM) is arranged on a host, the data center allocates the cloud task to the Virtual Machine (VM) in an idle state at present to execute; when all the virtual machines on the host computer are occupied by the cloud tasks and fully loaded, the newly arrived cloud tasks enter a queue buffer area of a virtual machine queuing model, and the queue buffer area is a first-in first-out (FIFO) queue specially used for storing the cloud tasks waiting to be executed; after the cloud task is executed and leaves a certain Virtual Machine (VM), the data center distributes the cloud task at the head of the queue buffer area to the Virtual Machine (VM); after the cloud task obtains the access right to the Virtual Machine (VM), the cloud task can be immediately provided with cloud service;
considering the running time of the data center as being composed of a plurality of continuous time slices, wherein the length of each time slice is defined as t; therefore, in time slot t, the queuing model of a single virtual machine: (Queuing model) is defined by the following equation:
wherein the content of the first and second substances,andis the queue length of the kth Virtual Machine (VM) in the jth and jth +1 time slices;is shown in the jth time sliceMiddle k virtual machineThe number of cloud tasks processed;is the jth time sliceMiddle k virtual machineThe number of cloud tasks in the queue; when in useWhen the queue length in a Virtual Machine (VM) is less than its processing powerThe queue length of the Virtual Machine (VM) may reach a minimum value of 0;
the length of a Virtual Machine (VM) queuing model can directly influence the running state of a host in a data center; the operating states of the host can be classified into the following two types:
1) active state: a cloud task queue of a Virtual Machine (VM) on a host is not empty, and a cloud task is waiting to be processed;
2) an idle state: a cloud task queue of a Virtual Machine (VM) on a host is empty, and the host in an idle state can be regarded as being in a sleep mode and in a low power consumption state, and generally speaking, the power consumption generated by the host in the idle state is a constant value;
the power consumption model of a data center can therefore be described in the form:
in the formula (I), the compound is shown in the specification,representing the total energy consumption of the current data center, wherein the first item on the right side of the energy consumption equation represents the total power consumption generated by the physical host machine in the current active state; p is the number of physical hosts in operation,is the firstPower consumption of individual physical hosts in an active state; q is the number of physical hosts in the idle state,is the firstPower consumption of each physical host in an idle state; the power consumption in the idle state can be regarded as a constant; m is the maximum number of Virtual Machines (VMs) currently available in the data center;
when the resource utilization rate is higher, the number of the physical hosts in the idle state is less, and the power consumption is also less; thus equation (2) can be expressed in the following form:
whereinIs as followsThe power consumption of the individual physical hosts in the active state,,is a constant power consumption of the host in idle state.
S2, extracting QoS characteristics of the data center by using a stacking noise reduction automatic encoder technology to obtain a matrix describing QoS characteristic information after dimension reduction, and obtaining the maximum response time of the current virtual machine through the matrix to perfect the constraint condition of a collaborative optimization objective function;
processing resource occupation historical data of a virtual machine in a data center by using a stacking noise reduction automatic encoder technology, and extracting characteristic information influencing QoS indexes of the data center through robustness of a stacking noise reduction encoder;
using a high-dimensional matrix composed of a parameter set of a virtual machine as data of a network input layer, constructing a target function of a stacking noise reduction automatic encoder, and performing dimension reduction and data reconstruction on the high-dimensional matrix through the target function to obtain optimized characteristics;
in the process of minimizing the loss function of the encoder, the complexity of the stacked noise reduction automatic encoder is also used as one index to participate in the training process and is used as a constraint term of the loss function; adding regular terms based on weight attenuation after loss function to improve generalization effect of the stacked noise reduction automatic encoder to avoid overfitting, and adoptingAndcross entropy ofAs a loss function, the robustness of the stacked noise reduction auto-encoder is further enhanced.
In the process of executing the cloud task, when the response time of the cloud taskWhen the length is too long, SLA vision can be caused; in the process of training the stacking noise reduction automatic encoder, a high-dimensional matrix formed by a parameter set of the virtual machine is used as data of a network input layer of the stacking noise reduction automatic encoder, and dimension reduction and data reconstruction are carried out on the high-dimensional matrix through the stacking noise reduction automatic encoder to obtain optimized QoS characteristic information;
the parameter set of the virtual machine comprises cloud task response time, queue length, virtual machine CPU utilization rate, bandwidth, CPU number, memory size, memory occupancy rate, virtual machine migration times, virtual machine migration time overhead and SLA violation rate of the virtual machine at the previous moment; when the data center is provided withWhen the station virtual machine is in a startup state,in each scheduling period interval, the parameter set of the virtual machine is combined to formHigh dimensional matrix of;
The structure of the stacked noise reduction automatic encoder is shown in fig. 2, and the network structure of the encoder comprises an input layer, a damaged input layer, a hidden layer and an output layer; in addition to the conventional encoding and decoding stages, the stacked noise reduction auto-encoder performs a random corruption process on the data before encoding in order to add noise to the training data, thereby forcing the encoder to learn and extract better QoS profile information from the input layer data; order toRepresenting a sample of the original input data,representing corrupted data after gaussian noise is added, W1 and W2 represent encoder and decoder weights respectively,andrepresenting the bias term, the encoding function of the stacked noise reduction automatic encoder (SDAE) encodes the original input to obtain a new feature representation, and the encoding process is as follows:
whereinIs a function of the sigmoid and is,it is used as a non-linear deterministic mapping; similarly, the decoding function reconstructs the characterizing input from the hidden layer to the original input, the decoding process is as follows:
the goal of training a stacked noise reduction autoencoder is to optimize the parameter setTo minimizeA reconstruction error therebetween; order toRepresenting loss function of stacked noise reduction auto-encoder (SDAE), optimization objective function of noise reduction auto-encoderCan be expressed as:
in the process of minimizing the loss function, the complexity of the stacking noise reduction automatic encoder needs to be considered, and an over-fitting is easily caused by an excessively complex model, so that the complexity of the model is also used as one index to participate in the training process, and the model is further constrained; adding a weight attenuation-based regularization term (L2 Norm) after the loss function to improve the generalization effect of the model to avoid overfitting; the weight attenuation is a coefficient placed in front of the regular term, and the influence of the complexity of the model on the loss function can be adjusted;
by usingAndcross entropy ofAs a loss function;;a weight adjustment coefficient that is a regularization term;(ii) a The resulting loss function can thus be obtained as follows:
according to the above equation, the loss optimization function of the stacked noise reduction auto-encoder can be expressed as:
the optimization of the objective function can be solved by using a quasi-newton method, in this embodiment, a quasi-newton optimization algorithm based on linear search is used for estimating parametersBy using the obtained estimated parameters as the parameters of the network model of the stacking noise reduction automatic encoder, a low-dimensional matrix reflecting the QoS characteristic information of the virtual machine can be obtained;
Maximum response time with cloud taskAs an evaluation criterion of the data center service quality; order toRepresenting a maximum response time of a Virtual Machine (VM);represents a maximum response time specified by a Service Level Agreement (SLA);QoS characteristic information matrix representing a stacked noise reduction auto-encoder (SDAE) through a current Virtual Machine (VM)Andcalculating the maximum response time of the cloud task on the virtual machine according to the cosine similarity; in order to ensure the quality of service of the cloud task,should be less thanAnda minimum value in between; thus, the overall collaborative optimization objective function can be written as:
s3, combining a cloud task arrival queuing model, a collaborative optimization objective function and a Lyapunov (Lyapunov) optimization method, obtaining a cloud task queue length condition which should be possessed by a virtual machine meeting energy consumption and QoS (quality of service) optimization scheduling, obtaining a resource scheduling algorithm based on the Lyapunov (Lyapunov) optimization theory, and solving the resource scheduling problem of energy consumption and QoS collaborative optimization into each time slice by adopting the resource scheduling algorithm to realize resource scheduling based on energy consumption and QoS collaborative optimization.
Analyzing the queue length of the cloud task reaching a queuing model by utilizing a Lyapunov (Lyapunov) optimization theory, fitting the cloud task execution time at the current moment by utilizing QoS (quality of service) characteristic information, and improving the time constraint condition of data center resource scheduling;
when in useIn time, the resource scheduling algorithm should reduce the number of non-empty task queues of a Virtual Machine (VM) to the maximum extent to reduce the energy consumption generated by a data center; at the same time, whenThe length of the cloud task queue should be kept large enough to preferentially make the utilization rate of the virtual machine as full as possible, so as to fully release the processing capacity of the virtual machine.
The Lyapunov (Lyapunov) equation is commonly used to analyze the stability of the queuing model, and the Lyapunov (Lyapunov) equation expression may be expressed in the form:
wherein at each time sliceIn general, the Lyapunov (Lyapunov) drift function may be defined as follows:
in order to maintain the stability of the cloud task queue in the cloud computing fluctuation environment, the result of equation (12) should be converged and the drift value should be maintainedThe smaller the value of the final result is, the more stable the cloud task queue is; in this embodiment, the stability of the cloud task queue model under the condition of the formula (10) is firstly proved, that is, the maximum value of the formula (12) can be converged; the following expression can be obtained by equation (12):
based on the following inequality relationships:
the maximum cloud task queue length that a virtual machine can accommodate may be expressed in the form:
wherein t represents the execution time of a single cloud task; integrating the formula, when the constraint condition of the collaborative optimization objective function needs to be satisfied, the number of the cloud tasks scheduled on the virtual machine needs to satisfy the following condition:
thus, equation (13) can be rewritten in the form:
the first term to the right of the inequality number of formula (17) is a constant term; the second term is a convex quadratic function, which means that the function can obtain a finite maximum value; therefore, it is proved that the Lyapunov (Lyapunov) drift function can make the cloud computing environment maintain stability under the condition constraint of the formula (10).
On the premise of ensuring the service quality of the cloud task, in order to further reduce the energy consumption of the physical host, namely, the number of the virtual machines running in the virtual machine queuing model is reduced to the minimum; the objective function of the co-optimization problem can be rewritten as:
wherein V is a unit conversion coefficient which is used for converting the energy consumption into corresponding cloud task queue length units; according to the Lyapunov (Lyapunov) optimization method, the resource scheduling objective optimization function of the data center can be obtained from formula (18):
the solution process of the optimization function is as follows:
by performing a transformation on equation (20), an equivalent equation of the form:
whereinIs the penalty value of the Lyapunov (Lyapunov) drift-plus-penalty function, the first solution of equation (21) is:
as shown in the formula (23),(ii) a When in useIn time, the resource scheduling algorithm must satisfy the following conditions:(ii) a Thus can be properly obtainedIn time, the resource scheduling algorithm should reduce the number of non-empty task queues of the VM to the maximum extent so as to reduce the energy consumption generated by the data center;at the same time, whenThe resource scheduling algorithm will preferentially make the virtual machine utilization as fully loaded as possible, so that the processing power of the virtual machine is fully released.
The resource scheduling algorithm based on the Lyapunov (Lyapunov) optimization theory is executed by the data center in each preset scheduling period of the cloud environment, and the execution process of the resource scheduling algorithm is as follows:
if there isThe cloud task arrives at the data center within the period,is the maximum response time of the cloud task,virtual machineThe processing power of (a) is set,for virtual machinesThe current cloud task queuing queue length; queuing a queue for a current cloud taskFinding the queue with the shortest queue length(ii) a If it is notAnd then:
whereinFor the number of cloud tasks currently pending,to be made available to a virtual machineThe number of cloud tasks to process;
if it is notThen give an order(ii) a And slave queues that have been assigned to virtual machinesRemoving the queue and repeating the steps until the queue length is 0; the resource scheduling algorithm is to arrange the resource utilization rates of the virtual machines with the non-empty queues in an ascending order, so that the virtual machines of the queues meeting the Lyapunov optimization condition are selected from the virtual machines with the non-empty queues to be migrated, the virtual machine with the lowest utilization rate reaches the maximum resource utilization rate limit, and the resource scheduling algorithm can realize the balanced utilization of resources and reduce the number of overloaded hosts.
In this embodiment, there isA virtual machine, an orderIndicating by performing on each time slice tThe strategy obtained by the line drift penalty function,the unit conversion coefficient is expressed and used for converting the system energy consumption into corresponding cloud task queue length units;andrespectively representing a Lyapunov (Lyapunov) drift value and a penalty value obtained by performing a drift-plus-penalty function; by analyzing the co-optimization objective function, the following equation can be derived:
wherein the content of the first and second substances,is a time sliceTo middleThe length of the queue for each virtual machine,as time slicesThe k-th virtual machine drift value,as a mapping function of the drift value and the drift policy for the kth virtual machine,is a constant value, and is characterized in that,;stacking parameter sets of the noise reduction automatic encoder for the time t;
meanwhile, expected values are taken from two sides of the formula, and according to an expected standard calculation mode, the following results can be obtained:
since the approximate value of the optimal solution is obtained by solving the quasi-Newton method, the optimal solution is obtainedThe optimal estimation strategy obtained for implementing the drift penalty function,the function E is a solving function for the expected value, which is constant.
Accumulating the time slices according to the above deduction conclusion, so as to obtain:
whereinIs an estimate of the lyapunov drift plus a penalty function penalty value,is the total number of time slices,for optimizing the objective function during a time slice TThe value of the lyapunov equation of (a),optimization objective function for stacking noise reduction auto-encoders at time tThe value of (a). Since L (Θ (T)). gtoreq.0, one can find:
observing the right end of the formula, only one adjustable parameter V can be found, and the rest are fixed values; thus, for the above formula, the objective optimization function for cloud environment energy consumption can be expressed as:
similarly, analyzing the average length of the cloud task queue, the following inequality may be obtained: ,
in the above-mentioned formula,represents a constant obtained according to the constant term of the formula (17) in the embodiment when the average length of the cloud task queue is calculated,representing the Lyapunov expectation and penalty scaling factor,representing the maximum penalty value achieved after execution of the drift penalty function,represents the minimum penalty value obtained after the execution of the drift penalty function;
v is used as a unique variable parameter to determine the backlog length of the cloud task queue; v determining the speed at which the time-averaged constraint of the original problem converges to a non-negative number; thus, the cloud task queue backlog length for the average duration may be represented by o (v); according to the conclusion obtained by the analysis, it is obvious that the unit conversion coefficient V has opposite influence on the energy consumption optimization of the cloud system and the length of the cloud task queue backlog; this illustrates that there is a trade-off between algorithm execution time (number of iterations) and algorithm efficiency (solution accuracy) [ O (V), O (1/V) ]. Because the Lyapunov optimization method belongs to the model-free learning category, and the optimization problem can be solved and completed in each independent time slice, the processing capacity of the virtual machine can be fully exerted, the QoS can be guaranteed, the energy consumption of the data center can be effectively reduced, and the time overhead of data center scheduling can be greatly reduced.
The above-mentioned procedures are preferred embodiments of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention shall be covered by the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. A resource scheduling implementation method based on energy consumption and QoS collaborative optimization is characterized by comprising the following steps:
s1, constructing a cloud task arrival queuing model of the multiple virtual machines in a cloud computing data center environment;
s2, extracting QoS characteristics of the data center by using a stacking noise reduction automatic encoder technology to obtain a matrix describing QoS characteristic information after dimension reduction, and obtaining the maximum response time of the current virtual machine through the matrix to perfect the constraint condition of a collaborative optimization objective function;
s3, combining the cloud task arrival queuing model, the collaborative optimization objective optimization function and the Lyapunov optimization method, obtaining the cloud task queue length condition which should be possessed by the virtual machine meeting the energy consumption and QoS optimization scheduling, obtaining a resource scheduling algorithm based on the Lyapunov optimization theory, and solving the resource scheduling problem of the energy consumption and QoS collaborative optimization into each time slice by adopting the resource scheduling algorithm to realize the resource scheduling based on the energy consumption and QoS collaborative optimization.
2. The method for implementing resource scheduling based on energy consumption and QoS collaborative optimization according to claim 1, wherein in step S1, the cloud task arrival queuing model is composed of a host queuing model and a virtual machine queuing model in series, and is used for optimizing a relationship between a virtual machine cloud task queue backlog length and system energy consumption;
in the host queuing model, after the cloud tasks are submitted to the data center, the data center adopts a load balancing strategy of a minimum load criterion to preferentially distribute the cloud tasks to the hosts with the minimum unfinished cloud task request number, and thus, a queuing model with the cloud task arrival interval time in exponential distribution and the cloud task execution time in exponential distribution is formed;
in the virtual machine queuing model, each virtual machine instance forms an exponential distribution of the cloud task arrival interval time and the cloud task execution time, and the number of the virtual machines isVirtual machine capacity ofAccording to the queuing model, the cloud tasks are distributed by the host queuing model and then submitted to the virtual machine queuing model by the data center, and then are processed in a first-come-first-serve mode.
3. The method for realizing resource scheduling based on energy consumption and QoS collaborative optimization according to claim 2, wherein when a cloud task enters a host queuing model, if at least one virtual machine is used on a host, the data center allocates the cloud task to the virtual machine currently in an idle state for execution; when all the virtual machines on the host are occupied by the cloud tasks and fully loaded, the newly arrived cloud tasks enter a queue buffer area of a virtual machine queuing model, and the queue buffer area is a first-in first-out queue specially used for storing the cloud tasks waiting to be executed; after the cloud task is executed and leaves a certain virtual machine, the data center distributes the cloud task at the head of the queue buffer area to the virtual machine; after the cloud task obtains the access right to the virtual machine, the cloud task can immediately provide cloud service for the virtual machine;
considering the running time of the data center as being composed of a plurality of continuous time slices, wherein the length of each time slice is defined as t; therefore, in the time slot t, the queuing model of the single virtual machine is defined by the following formula:
wherein the content of the first and second substances,andis the queue length of the kth Virtual Machine (VM) in the jth and jth +1 time slices;is shown in the jth time sliceMiddle k virtual machineThe number of cloud tasks processed;is the jth time sliceMiddle k virtual machineThe number of cloud tasks in the queue; when in useThe queue length in the virtual machine is less than its processing capacityThe queue length of the virtual machine can reach a minimum value of 0.
4. The method for realizing resource scheduling based on energy consumption and QoS collaborative optimization according to claim 3, wherein the length of the virtual machine queuing model directly affects the running state of a host in the data center; the operating states of the host are classified into the following two types:
1) active state: the cloud task queue of the virtual machine on the host is not empty, and the cloud task is waiting to be processed;
2) an idle state: the cloud task queue of the virtual machine on the host is empty, the host in an idle state can be regarded as being in a sleep mode and in a low power consumption state, and the power consumption generated by the host in the idle state is a constant value;
the power consumption model of a data center is described in the form:
in the formula (I), the compound is shown in the specification,representing the total energy power consumption of the current data center, with the energy consumption equation to the rightThe first item represents the total power consumption generated by the physical host currently in the active state; p is the number of physical hosts in operation,is the firstPower consumption of individual physical hosts in an active state; q is the number of physical hosts in the idle state,is the firstPower consumption of each physical host in an idle state; the power consumption in the idle state is regarded as a constant; m is the maximum number of virtual machines currently available in the data center;
equation (2) is expressed in the following form:
5. The method for realizing resource scheduling based on energy consumption and QoS collaborative optimization according to claim 1, wherein in step S2, historical data of resource occupation of virtual machines in a data center are processed by using a stacking noise reduction automatic encoder technology, and feature information affecting QoS indexes of the data center is extracted through robustness of a stacking noise reduction encoder;
using a high-dimensional matrix composed of a parameter set of a virtual machine as data of a network input layer, constructing a target function of a stacking noise reduction automatic encoder, and performing dimension reduction and data reconstruction on the high-dimensional matrix through the target function to obtain optimized characteristics;
in the minimization process of the loss function of the encoder, the complexity of the stacked noise reduction automatic encoder is also used as one index to participate in the training process and is used as a constraint term of the loss function; adding regular terms based on weight attenuation after loss function to improve generalization effect of the stacked noise reduction automatic encoder to avoid overfitting, and adoptingAndcross entropy ofAs a loss function, the robustness of the stacked noise reduction auto-encoder is further enhanced.
6. The method as claimed in claim 5, wherein the method is implemented in cloud computing, and is characterized in that the method is implemented in cloud computingIn the process of executing the service, when the response time of the cloud taskWhen the length is too long, SLA vision can be caused; in the process of training the stacking noise reduction automatic encoder, a high-dimensional matrix formed by a parameter set of the virtual machine is used as data of a network input layer of the stacking noise reduction automatic encoder, and dimension reduction and data reconstruction are carried out on the high-dimensional matrix through the stacking noise reduction automatic encoder to obtain optimized QoS characteristic information;
the parameter set of the virtual machine comprises cloud task response time, queue length, virtual machine CPU utilization rate, bandwidth, CPU number, memory size, memory occupancy rate, virtual machine migration times, virtual machine migration time overhead and SLA violation rate of the virtual machine at the previous moment; when the data center is provided withWhen the virtual machine is in a starting state, in each scheduling period interval, the parameter set of the virtual machine is combined to formHigh dimensional matrix of。
7. The method for implementing resource scheduling based on energy consumption and QoS collaborative optimization of claim 6, wherein the network structure of the stacked noise reduction automatic encoder comprises an input layer, a damaged input layer, a hidden layer and an output layer; the stacking noise reduction automatic encoder performs random damage processing on data before encoding, and adds noise into training data, so that the encoder is forced to learn and extract better QoS characteristic information from input layer data; order toRepresenting the original inputInto a data sample, whereinAnd D is the total number of samples of the input layer of the stacked noise reduction automatic encoder;representing corrupted data after the addition of gaussian noise,andrepresenting the weights of the encoder and decoder, respectively;andrepresenting the bias term, the coding function of the stacked noise reduction automatic encoder (SDAE) encodes the original input to obtain a new feature representationThe encoding process is as follows:
whereinIs a function of the sigmoid and is,it is used as a non-linear deterministic mapping; decoding function characterizes from hidden layerInput reconstruction into original inputThe decoding process is as follows:
the goal of training a stacked noise reduction autoencoder is to optimize the parameter setTo minimizeA reconstruction error therebetween; order toRepresenting loss function of stacked noise-reducing automatic encoder, optimized objective function of noise-reducing automatic encoderExpressed as:
in the process of minimizing the loss function, the complexity of the stacking noise reduction automatic encoder is considered, the complexity of the model is also used as one index to participate in the training process, and then the model is constrained; adding a weight attenuation-based regularization term after the loss function to improve the generalization effect of the model to avoid overfitting; the weight attenuation is a coefficient placed in front of the regular term and is used for adjusting the influence of the complexity of the model on the loss function;
by usingAndcross entropy ofAs a loss function;(ii) a WhereinFor stacked noise reduction auto-encodersjAt one time slice, the firstiA squared loss function value of a weight vector of samples, wherein,, Tis the total number of time slices;d is the total number of samples of the input layer of the stacked noise reduction automatic encoder;a weight adjustment coefficient that is a regularization term;(ii) a The resulting loss function is thus obtained as follows:
according to the above equation, the loss optimization function of the stacked noise reduction auto-encoder can be expressed as:
solving for estimating parameters by using quasi-Newton methodWhereinIs an estimate of the encoder weight(s),is an estimate of the decoder weights,is a bias termIs determined by the estimated value of (c),is a bias termObtaining a low-dimensional matrix reflecting the QoS characteristic information of the virtual machine by using the obtained estimation parameters as the parameters of the network model of the stacking noise reduction automatic encoder;
Maximum response time with cloud taskAs an evaluation criterion of the data center service quality; order toRepresenting a maximum response time of the virtual machine;represents a maximum response time specified by a service level agreement;QoS characteristic information matrix representing passage of stacking noise reduction automatic encoder through current virtual machineAndcalculating the maximum response time of the cloud task on the virtual machine according to the cosine similarity;should be less thanAnda minimum value in between; the overall collaborative optimization objective function is as follows:
8. the method for realizing resource scheduling based on energy consumption and QoS collaborative optimization according to claim 7, wherein in step S3, a Lyapunov optimization theory is used to analyze the queue length of a cloud task to a queuing model, and QoS characteristic information is used to fit the cloud task execution time at the current moment, so as to improve the time constraint condition of data center resource scheduling;
when in useIn time, the resource scheduling algorithm should reduce the number of non-empty task queues of a Virtual Machine (VM) to the maximum extent to reduce the energy consumption generated by a data center; at the same time, whenThe length of the cloud task queue should remain large enough to allow the processing power of the virtual machine to be fully freed.
9. The method for realizing resource scheduling based on energy consumption and QoS collaborative optimization of claim 8, wherein the Lyapunov equation is expressed in the form of:
whereinIs the queue length of the kth virtual machine in the jth time slice,is the value of the optimization objective function of the stacking noise reduction automatic encoder in the jth time slice;to stack the lyapunov equation values of the optimization objective function of the noise reduction auto-encoder in the jth time slice,,Tis the total number of time slices,, Nfor the maximum number of virtual machines available in the current data center, in each time sliceThe lyapunov drift function is defined as follows:
wherein the content of the first and second substances,for the size of the drift value, the stability of the cloud task queue needs to be maintained under the environment of cloud computing fluctuation, the result of the formula (12) is required to be converged, and the drift valueThe smaller the value of the final result is, the more stable the cloud task queue is;
on the premise of ensuring the service quality of the cloud task, the energy consumption of the physical host is further reduced, namely the number of the virtual machines running in the virtual machine queuing model is reduced to the minimum; the objective function of the collaborative optimization problem is rewritten as:
wherein V is a unit conversion coefficient which is used for converting the energy consumption into corresponding cloud task queue length units; according to the Lyapunov optimization method, a resource scheduling collaborative optimization objective function of the data center is obtained from a formula (13):
the solution to this function is as follows:
transforming equation (15) to obtain an equivalent equation of the form:
whereinIs the firstThe penalty value of the lyapunov drift plus penalty function for each virtual machine, the first solution of equation (16) is:
as shown in the formula (18),(ii) a When in useIn time, the resource scheduling algorithm must satisfy the following conditions:(ii) a Thus whenIn time, the resource scheduling algorithm should reduce the number of non-empty task queues of the VM to the maximum extent; at the same time, whenIn time, the resource scheduling algorithm preferentially enables the utilization rate of the virtual machine to tend to or be in a full load state, so that the processing capacity of the virtual machine is fully released.
10. The method for realizing resource scheduling based on energy consumption and QoS collaborative optimization according to any one of claims 1 to 9, wherein a resource scheduling algorithm based on the Lyapunov optimization theory is executed by the data center in each preset scheduling cycle of the cloud environment, and the execution process of the resource scheduling algorithm is as follows:
if there isThe cloud task arrives at the data center within the period,is the maximum response time of the cloud task,virtual machineThe processing power of (a) is set,for virtual machinesThe current cloud task queuing queue length; queuing a queue for a current cloud taskFinding the queue with the shortest queue length(ii) a If it is notAnd then:
whereinFor the number of cloud tasks currently pending,to be made available to a virtual machineThe number of cloud tasks to process;
if it is notThen give an order(ii) a And slave queues that have been assigned to virtual machinesRemoving the cloud task queue, and repeating the steps until the length of the cloud task queue is 0; the resource scheduling algorithm is to arrange the resource utilization rates of the virtual machines with the non-empty queues in an ascending order, so that the virtual machines of the queues meeting the Lyapunov optimization conditions are selected from the virtual machines with the non-empty queues to be migrated, and the virtual machine with the lowest utilization rate reaches the maximum resource utilization rate limit.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111041530.7A CN113515351B (en) | 2021-09-07 | 2021-09-07 | Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111041530.7A CN113515351B (en) | 2021-09-07 | 2021-09-07 | Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113515351A true CN113515351A (en) | 2021-10-19 |
CN113515351B CN113515351B (en) | 2022-03-29 |
Family
ID=78062825
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111041530.7A Active CN113515351B (en) | 2021-09-07 | 2021-09-07 | Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113515351B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113722076A (en) * | 2021-10-29 | 2021-11-30 | 华南理工大学 | Real-time workflow scheduling method based on QoS and energy consumption collaborative optimization |
CN114460892A (en) * | 2021-12-20 | 2022-05-10 | 北京科技大学 | Task control method based on cloud programmable logic controller |
CN115150277A (en) * | 2022-06-13 | 2022-10-04 | 燕山大学 | Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center |
CN115167641A (en) * | 2022-07-07 | 2022-10-11 | 凯德技术长沙股份有限公司 | Inertia self-adaptive temperature control method and system |
CN116155835A (en) * | 2023-01-17 | 2023-05-23 | 哈尔滨工业大学(威海) | Cloud resource service quality assessment method and system based on queuing theory |
CN116302451A (en) * | 2023-05-18 | 2023-06-23 | 广州豪特节能环保科技股份有限公司 | Offline energy-saving scheduling method and system for cloud computing data center |
CN117170886A (en) * | 2023-11-03 | 2023-12-05 | 北京邮电大学 | Continuous learning resource scheduling method and device for large-connection video analysis |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150131504A1 (en) * | 2013-11-14 | 2015-05-14 | National Chiao Tung University | Power-saving data scheduling system in lte and method thereof |
CN104811491A (en) * | 2015-04-17 | 2015-07-29 | 华南理工大学 | Cloud computing resource scheduling method based on genetic algorithm |
CN106528941A (en) * | 2016-10-13 | 2017-03-22 | 内蒙古工业大学 | Data center energy consumption optimization resource control algorithm under server average temperature constraint |
CN109445903A (en) * | 2018-09-12 | 2019-03-08 | 华南理工大学 | Cloud computing energy-saving distribution implementation method based on the discovery of QoS feature |
CN112860409A (en) * | 2021-03-09 | 2021-05-28 | 合肥工业大学 | Mobile cloud computing random task sequence scheduling method based on Lyapunov optimization |
-
2021
- 2021-09-07 CN CN202111041530.7A patent/CN113515351B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150131504A1 (en) * | 2013-11-14 | 2015-05-14 | National Chiao Tung University | Power-saving data scheduling system in lte and method thereof |
CN104811491A (en) * | 2015-04-17 | 2015-07-29 | 华南理工大学 | Cloud computing resource scheduling method based on genetic algorithm |
CN106528941A (en) * | 2016-10-13 | 2017-03-22 | 内蒙古工业大学 | Data center energy consumption optimization resource control algorithm under server average temperature constraint |
CN109445903A (en) * | 2018-09-12 | 2019-03-08 | 华南理工大学 | Cloud computing energy-saving distribution implementation method based on the discovery of QoS feature |
CN112860409A (en) * | 2021-03-09 | 2021-05-28 | 合肥工业大学 | Mobile cloud computing random task sequence scheduling method based on Lyapunov optimization |
Non-Patent Citations (3)
Title |
---|
ROHAIZA YUSOFF ET AT ALL: "Energy-Efficient Resource Allocation Scheduler with QoS Aware Supports for Green LTE Network", 《2015 IEEE 6TH CONTROL AND SYSTEM GRADUATE RESEARCH COLLOQUIUM (ICSGRC)》 * |
李磊等: "基于李雅普诺夫优化的容器云队列在线任务和资源调度设计", 《计算机应用》 * |
袁非牛等: "自编码神经网络理论及应用综述", 《计算机学报》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113722076A (en) * | 2021-10-29 | 2021-11-30 | 华南理工大学 | Real-time workflow scheduling method based on QoS and energy consumption collaborative optimization |
CN113722076B (en) * | 2021-10-29 | 2022-03-29 | 华南理工大学 | Real-time workflow scheduling method based on QoS and energy consumption collaborative optimization |
CN114460892A (en) * | 2021-12-20 | 2022-05-10 | 北京科技大学 | Task control method based on cloud programmable logic controller |
CN115150277B (en) * | 2022-06-13 | 2023-09-15 | 燕山大学 | Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center |
CN115150277A (en) * | 2022-06-13 | 2022-10-04 | 燕山大学 | Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center |
CN115167641A (en) * | 2022-07-07 | 2022-10-11 | 凯德技术长沙股份有限公司 | Inertia self-adaptive temperature control method and system |
CN115167641B (en) * | 2022-07-07 | 2023-06-20 | 凯德技术长沙股份有限公司 | Inertial self-adaptive temperature control method and system |
CN116155835A (en) * | 2023-01-17 | 2023-05-23 | 哈尔滨工业大学(威海) | Cloud resource service quality assessment method and system based on queuing theory |
CN116155835B (en) * | 2023-01-17 | 2024-04-16 | 哈尔滨工业大学(威海) | Cloud resource service quality assessment method and system based on queuing theory |
CN116302451A (en) * | 2023-05-18 | 2023-06-23 | 广州豪特节能环保科技股份有限公司 | Offline energy-saving scheduling method and system for cloud computing data center |
CN116302451B (en) * | 2023-05-18 | 2023-08-08 | 广州豪特节能环保科技股份有限公司 | Offline energy-saving scheduling method and system for cloud computing data center |
CN117170886A (en) * | 2023-11-03 | 2023-12-05 | 北京邮电大学 | Continuous learning resource scheduling method and device for large-connection video analysis |
CN117170886B (en) * | 2023-11-03 | 2024-02-02 | 北京邮电大学 | Continuous learning resource scheduling method and device for large-connection video analysis |
Also Published As
Publication number | Publication date |
---|---|
CN113515351B (en) | 2022-03-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113515351B (en) | Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization | |
CN110096349B (en) | Job scheduling method based on cluster node load state prediction | |
CN108182115B (en) | Virtual machine load balancing method in cloud environment | |
Guo et al. | Cloud resource scheduling with deep reinforcement learning and imitation learning | |
WO2023184939A1 (en) | Deep-reinforcement-learning-based adaptive efficient resource allocation method for cloud data center | |
CN109885397B (en) | Delay optimization load task migration algorithm in edge computing environment | |
CN109324875B (en) | Data center server power consumption management and optimization method based on reinforcement learning | |
CN111444021B (en) | Synchronous training method, server and system based on distributed machine learning | |
CN104168318A (en) | Resource service system and resource distribution method thereof | |
CN111431961A (en) | Energy-saving task allocation method for cloud data center | |
CN110955463B (en) | Internet of things multi-user computing unloading method supporting edge computing | |
CN111182637A (en) | Wireless network resource allocation method based on generation countermeasure reinforcement learning | |
CN111611062B (en) | Cloud-edge collaborative hierarchical computing method and cloud-edge collaborative hierarchical computing system | |
CN111752708A (en) | Storage system self-adaptive parameter tuning method based on deep learning | |
CN103699433B (en) | One kind dynamically adjusts number of tasks purpose method and system in Hadoop platform | |
CN112685138B (en) | Multi-workflow scheduling method based on multi-population hybrid intelligent optimization in cloud environment | |
CN112114960B (en) | Scheduling strategy for remote sensing image parallel cluster processing adapting to internet scene | |
CN111813506A (en) | Resource sensing calculation migration method, device and medium based on particle swarm algorithm | |
CN112835698A (en) | Heterogeneous cluster-based dynamic load balancing method for request classification processing | |
US20220156633A1 (en) | System and method for adaptive compression in federated learning | |
CN115408152A (en) | Adaptive resource matching obtaining method and system | |
CN115951989B (en) | Collaborative flow scheduling numerical simulation method and system based on strict priority | |
CN115237568A (en) | Mixed weight task scheduling method and system for edge heterogeneous equipment | |
WO2024037368A1 (en) | Scheduling optimization method of scheduling apparatus, scheduling apparatus and storage medium | |
Dong et al. | Quantum particle swarm optimization for task offloading in mobile edge computing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |