WO2018137402A1

WO2018137402A1 - Cloud data centre energy-saving scheduling implementation method based on rolling grey prediction model

Info

Publication number: WO2018137402A1
Application number: PCT/CN2017/113854
Authority: WO
Inventors: 刘发贵; 王彬
Original assignee: 华南理工大学
Priority date: 2017-01-26
Filing date: 2017-11-30
Publication date: 2018-08-02
Also published as: CN106899660A; CN106899660B

Abstract

Disclosed is a cloud data centre energy-saving scheduling implementation method based on a rolling grey prediction model. In the present invention, a cloud data centre energy-saving flow is abstracted as four modules of load prediction, error checking, thermal perception classification and virtual machine scheduling. A working load of the data centre at the next moment is predicted by means of the load prediction module to obtain a load utilization rate of each host. The thermal perception classification module can divide the thermal states of all hosts according to predicted values of the load utilization rates of the hosts, wherein the utilization rate of a host in the state of being hotter is at a higher level, while the utilization rate of a host in the state of being cooler is at a lower level. In order that most of the hosts are maintained in the thermal state of being relatively mild, the virtual machine scheduling module performs operations, such as migration and integration, on a virtual machine on each of the hosts according to a classification result of the thermal states, so as to finally achieve the purposes of guaranteeing the service quality of the data centre and reducing the energy consumption thereof. The present invention overcomes the problem that a traditional grey model has the difficulty of low precision due to the deficiency of some values.

Description

Cloud data center energy-saving scheduling implementation method based on rolling gray prediction model

Technical field

The invention belongs to the field of cloud computing energy-saving scheduling, and particularly relates to a cloud data center energy-saving scheduling implementation method based on a rolling gray prediction model.

Background technique

As an emerging technology in the information and communication technology industry, cloud computing has gradually entered the ranks of thousands of households. It has been favored by enterprises and individual users with its high efficiency, low threshold and high scalability. With the gradual maturity of cloud computing development and the continuous enrichment of user demand, the scale of a series of supporting facilities such as data center servers is also growing rapidly. Large-scale cloud computing data centers with tens of thousands of service nodes have been established around the world, which has saved more computing resources and storage resources in the cloud, but it has also triggered a series of energy consumption problems. . Greenpeace said in its cloud computing report that by 2020, the energy consumption of data centers of major IT operators around the world will be close to 196.3 billion kWh, which will exceed the current total consumption of electricity in Germany, France, Canada and Brazil. sum. At the same time, the Chinese government work report pointed out that the main task in 2008 was to “enhance energy conservation, emission reduction and environmental protection. The Ministry of Science and Technology and the Ministry of Industry and Information Technology formulated the “Special Action Plan for Energy Conservation and Emission Reduction Technology 2014-2015” in 2014. To play a leading role in technology to alleviate the constraints of resources and environment. Increasingly rising energy consumption costs have increased the burden on enterprises, but also increased the carbon emissions affecting climate change on a global scale, which has sounded the alarm for environmental issues.

The rapid increase in energy consumption of these cluster servers has become an important factor affecting the efficiency and development of enterprises. Research in the field of cloud computing energy conservation has also become a key task in the research of emerging technologies at home and abroad. At present, the starting point of cloud computing energy conservation research at home and abroad is different, and there are certain deficiencies. At present, there are still major deficiencies in the workload monitoring and prediction of cloud data centers in the case of traffic bursts. The scheduling integration strategy has drawbacks.

Summary of the invention

In order to more intelligently and effectively reduce the power consumption of the data center and guarantee the quality of service in the case of network traffic bursts, the present invention provides a cloud data center energy-saving scheduling implementation method based on a rolling gray prediction model.

The present invention is specifically achieved by the following technical solutions.

A cloud data center energy-saving scheduling implementation method based on a rolling gray prediction model, which abstracts the energy-saving process of the cloud data center into four modules: load prediction, error checking, thermal sensing classification, and virtual machine scheduling; and predicting the data center workload Therefore, the states of each host are classified, and the virtual machine scheduling algorithm is used to achieve energy saving.

Further, the load prediction is specifically: predicting a workload of the data center by using a rolling gray prediction model, and obtaining load utilization of each host node in the data center at the next moment.

Further, the error check is specifically: performing error check on the load predicted value and the actual workload, determining a deviation value of the current prediction result, and learning based on the error check module to correct the prediction result.

Further, the heat-aware classification is specifically: performing hot-sensing classification on all hosts in the cloud data center according to the current load prediction value of the host, and introducing a service level agreement as a reference indicator to set an upper boundary and a lower bound of the host workload threshold; When the host load utilization is higher than the threshold upper bound, lower than the lower threshold, the upper and lower thresholds, and the load utilization is 0, the four hot states are divided into four different hot states.

Further, the virtual machine scheduling is specifically: scheduling virtual machines according to the hot state of the current host, and solving the problem of overload and no load of the host through the virtual machine scheduling operation, and maintaining the hosts in the data center in a healthy heat. In the state.

Further, the load information of the data center and the performance data of the runtime are intelligently monitored by implementing the rolling gray prediction model algorithm and implementing it into the cloud data center environment in a modular form.

Compared with the prior art, the present invention has the following advantages and technical effects:

The invention mines and analyzes the real-time workload data of the data center, and uses the workload prediction algorithm based on the gray prediction model to establish a load prediction evaluation model, and predicts the load state of each server in the data center at the next moment to avoid overload or empty of the server. The phenomenon of loading, more effectively respond to the current widespread network traffic bursts. By analyzing the relationship between load and energy efficiency, the host's utilization heat perception classification model is established. The host classification standard based on the heat perception mechanism is proposed to solve the energy consumption performance problem caused by the direct load imbalance of each host in the data center.

The invention intelligently monitors the load information of the data center and the performance data of the runtime by implementing the rolling gray prediction model algorithm and implementing it into the cloud data center environment in a modular form. Reduce the energy consumption level in the data center, especially in the case of traffic bursts, reduce the operation and maintenance costs of the cloud service provider, and at the same time improve the data center SLA indicators to ensure the user's cloud service experience.

DRAWINGS

Figure 1 is a diagram of a cloud computing intelligent resource scheduling framework.

Figure 2 shows the energy-saving framework of the cloud data center system.

FIG. 3 is a graph of load prediction experiment results of cloud data center task sequence 1.

4 is a graph of load prediction experiment results of the cloud data center task sequence 2.

Figure 5 is a comparison of the experimental data of the rolling gray prediction model and the ARIMA model.

Figure 6 is a comparison of the average deviation ratio of the rolling gray prediction model and the ARIMA model.

detailed description

In order to make the technical solutions and advantages of the present invention more comprehensible, the following detailed description is made in conjunction with the accompanying drawings.

1. Strategic framework

1.1 Cloud computing resource intelligent scheduling framework

FIG. 1 is an architectural diagram of a cloud computing platform resource intelligent scheduling framework, which is divided into a host layer, a virtual machine layer, a performance evaluation layer, a scheduling layer, and a user layer from bottom to top. The scheduling layer and the evaluation layer are the core of the entire energy-saving strategy framework. Each layer will be explained below.

The host tier refers to all servers in the cloud data center, including all physical host nodes. These hardware devices are the lowest infrastructure in the cloud environment, providing a hardware foundation for energy-efficient scheduling management.

The virtual machine layer is based on the virtualization technology of the host layer. By virtualizing multiple server entities, the resource pool of the virtual machine layer is formed, which enables common computing and resource sharing in the cloud environment.

The performance evaluation layer refers to the collection and evaluation of load data, energy consumption, SLA, PUE and other performance data of the cloud data center. The evaluation layer needs to communicate with the virtual layer to obtain information about the utilization of virtual machine resources and the running status of each virtual machine.

The scheduling layer performs virtual machine initial allocation and virtual machine migration operation based on the data collected by the performance evaluation layer, such as load and energy consumption. The virtual machine is scheduled to ensure that the host can run in a good responsibility. Under the utilization environment.

The user layer refers to all users and service requesters in the cloud computing environment, including individual users, enterprise users, and all users of cloud computing. The user layer will always issue new service requests to the data center.

1.2 Cloud Data Center System Energy Saving Framework

2 is a framework diagram of an energy-saving architecture of a cloud computing data center system, which is divided into four modules from top to bottom, namely a load prediction module, an error checking module, a thermal sensing classification module, and a virtual machine scheduling module. Each module will be explained below.

(1) Load prediction module: The data center processes thousands of service requests per second. The load prediction module can continuously monitor the workload data of the physical machines in the data center and analyze the effective historical load data. Predict the CPU utilization of each PM at a future time. Load forecasting module can help us in effective areas The server is overloaded and unloaded in the current state of the data center.

(2) Error check module: After the load predictor completes the prediction process, the error check module calculates the deviation between the actual value and the predicted value, and optimizes the future prediction result by analyzing and calculating their relative errors.

(3) Thermal sensing classification module: According to the workload prediction value obtained by the above module, we divide the physical machine into four categories, called boiling point, warming point, cooling point, Freezing point. We introduced a Service Level Agreement (SLA) as an important reference indicator under this energy-saving architecture, and used SLA to set the upper and lower bounds of the PM workload threshold. When the current load utilization of PM is higher than the upper bound of the threshold, we call it the boiling point; when its utilization is lower than the threshold lower bound (non-zero), we call it the cold spot; when the load utilization is 0, it is the freezing point; when PM is used; When the load utilization is between the upper and lower limits of the threshold, we call it the temperature point. The thermal perceptual classification module can well understand and regulate the load situation inside the current data center by dividing the category area of the PM.

(4) Virtual machine scheduling module: The purpose of cloud computing energy saving is to enable the entire data center to operate at a higher utilization rate and to maximize the quality of the user's cloud service. So we need to make as many PMs as possible to operate as a warm spot. Therefore, we need to convert more boiling point PM into the temperature point and integrate some cool or freezing point PM as much as possible. The virtual machine scheduling module will migrate some of the virtual machines running at the boiling point to the cool spot, and some of the cool spots with sufficient capacity will be dynamically integrated into one.

2. Load forecasting

2.1 Data Center Energy Efficiency Indicators

The energy efficiency indicators in the cloud data center environment presented in this paper are more abundant. Table 1 lists the performance indicator data that needs to be obtained. There are three main types of indicators, namely time type, environment type and performance indicator.

Table 1

The time type is a significant attribute that determines the running time of the data center and the scheduling time interval during the test. The environment type defines the specific configuration parameters of each host and virtual machine in the data center, and the length of the currently processed cloud task list. These indicators together determine the objective state of the data center operation process. The performance indicators mainly reflect the current data center operation, including the utilization rate and energy consumption data of the host at various times, the energy efficiency of the data center in a certain time interval, the proportion of the data center that meets and violates the service level agreement, and the scheduling process. The number of hosts that were shut down, the total number of virtual machine migrations that occurred during the scheduling process, and so on.

2.2 rolling gray prediction model

In the face of traffic bursts, it is very challenging to analyze and process massive amounts of network traffic data in real time. When the traffic burst occurs, the frequency of boiling point and cold spot in the data center will be more frequent, and the overload phenomenon of the server is more serious. This paper proposes a data center workload prediction method to deal with traffic bursts.

The workload changes in the data center have certain volatility and uncertainty. The short-term prediction model for establishing workload is a small data modeling problem, while the gray model is more suitable for solving the problem of less data and poor information state prediction than other models. The residual state prediction model of the workload can be established with a small amount of data, which is suitable for the establishment of short-term prediction models for data center workloads.

The workload prediction method is based on the improved gray model, and the gray prediction method is considered as a time series pre- The model is a good alternative strategy. In the cloud environment, we use the historical workload value of each PM in the data center as the historical data sequence. The gray model uses a differential equation to describe the whole system establishment process.

Let x ⁽⁰⁾ = (x ⁽⁰⁾ (1), x ⁽⁰⁾ (2), ..., x ⁽⁰⁾ (n)) be the original workload sequence, where n is the length of the data sequence, 1 time The cumulative generation sequence is x ⁽¹⁾ = (x ⁽¹⁾ (1), x ⁽¹⁾ (2), ..., x ⁽¹⁾ (n)), where

Define the grey derivative of x ⁽¹⁾ as

d(k)=x ⁽⁰⁾ (k)=x ⁽¹⁾ (k)-x ⁽¹⁾ (k-1).

Let z ⁽¹⁾ generate a sequence of neighbors of the sequence x ⁽¹⁾ , ie

z ⁽¹⁾ (k)=αx ⁽¹⁾ (k)+(1-α)x ⁽¹⁾ (k-1),

Where α is the coefficient of variation, α∈(0,1).

Then the gray differential equation model defining GM(1,1) is

d(k)+az ⁽¹⁾ (k)=b, or x ⁽⁰⁾ (k)+az ⁽¹⁾ (k)=b,

x ⁽⁰⁾ (k)+az ⁽¹⁾ (k)=b,

In formula (1), x ^{(0) (k)} is called the derivative of gray, a factor known as development, z ^{(1) (k)} referred whitening background value, b referred to the amount of ash effect.

Substituting the timetable k=2,3,...,n into equation (1) and introducing the matrix vector notation:

Then the GM(1,1) model can be expressed as a matrix Y=Bu. where u is a matrix of fitting coefficients of a and b, and B is an information matrix provided by the original sequence.

Using regression analysis to obtain the estimated values of a and b, then the corresponding whitening model of the series is

Solutions have to

Then get the predicted value

Therefore, the next-time workload prediction value is obtained accordingly:

The traditional GM (1,1) model is likely to fall into the dilemma of low prediction accuracy due to the lack of data and partial values. After adopting the rolling GM (1,1) model, the prediction model will be re-established for each next prediction. The model will continue to use newer data for prediction, and the old data will be discarded. Using the historical data of the first 100 scheduling periods to predict the results of the 101st cycle and continuously updating them can greatly improve the accuracy of prediction. We will dynamically update the value of the translation transformation constant C in the model based on the historical load of the data center to ensure that the gray model is always available during the prediction process. Iteratively select the sequence

The maximum value of λ(k), Maxλ(k) and the minimum value Minλ(k), ensure that both Minλ(k) and Maxλ(k) are in the available space.

Inside.

3. Thermal sensing based virtual machine classification strategy

The traditional virtual machine scheduling strategy is mostly based on heuristic algorithms (genetic algorithm, simulated annealing, particle swarm optimization, etc.), and the heuristic algorithm can not avoid the problem of falling into the local optimal solution. The heuristic algorithm is essentially a kind of greed. The strategy that leads to an optimal solution that does not meet the greedy rules will be missed. The heat-aware virtual machine classification scheduling policy redistributes the virtual machines between physical nodes according to the "hot and cold" state of the host, which can effectively reduce the number of overloaded and no-load hosts in the data center, and achieve the balance between energy consumption and SLA. .

The thermal sensing model continuously monitors the thermal state information of the physical nodes of the data center, and divides the current physical nodes into four types according to the "hot and cold" state, which are boiling point, warming point, and cooling point. ), freezing point, and use SLA to set the upper and lower bounds of the PM workload threshold. When the current load utilization of PM is higher than the upper bound of the threshold, we call it the boiling point; when its utilization is lower than the threshold lower bound (non-zero), we call it the cold spot; when the load utilization is 0, it is the freezing point; when PM is used; When the load utilization is between the upper and lower limits of the threshold, we call it the temperature point. We need to make as many PMs as possible to operate as a warm spot. So we need to convert more The boiling point of the PM is between the temperature points and integrates some cool spots or PMs with freezing points as much as possible. The virtual machine scheduling module will migrate some of the virtual machines running at the boiling point to the cool spot. Some of the cool spots with sufficient capacity will be dynamically integrated into one, and the physical machine at the freezing point will be shut down to save energy.

4. Experimental description

For the rolling gray prediction model, we first need to let the data center run for a period of time t, and obtain a certain amount of historical load data as input samples for the rolling gray prediction model learning. The larger the time t, the more accurate the prediction result of the rolling gray prediction model. Here, the experiment t takes 100 sampling period lengths.

The calculation method of the average absolute deviation value and the average absolute percentage error of the experimental test of the rolling gray prediction model is as follows:

Where x ⁽⁰⁾ (k) is the actual value of the load utilization,

(k) is the predicted result of the rolling gray prediction model.

We collected the load data in the data center for 100 sampling periods as the initial input sequence of the rolling gray prediction model, and then runs the rolling gray prediction model algorithm. The algorithm calculates and predicts the 101st data by calculating the data sequence of the first 100 sampling periods. The load data at the cycle time is compared with the average absolute deviation and the average absolute percentage error of the actual and predicted workload values at the 101st cycle time to adjust the original rolling model and update the input data sequence. The workload of the 102nd, 103rd and subsequent cycles is then predicted accordingly.

Figure 3 and Figure 4 show the experimental results of workload prediction for two cloud task sequences. The cloud task sequence of Figure 3 has a traffic incident at 1-5 hours and 22-24 hours. The cloud of Figure 4 A traffic incident occurred at the 10th hour and the 20th hour of the task sequence. The experimental rolling grey prediction model is compared with the autoregressive moving average (ARIMA) model and the actual load value. Both sets of experimental results show that the rolling gray prediction model is more accurate than the actual value for the data center workload (Actual). Compared with the ARIMA model, the rolling gray prediction model is more timely for feedback in the case of traffic bursts.

FIG. 5 is an experimental result after running the cloud task sequence shown in FIG. 4. The result shows that the average deviation ratio of the rolling gray prediction model in the load prediction process of the cloud data center is 6.93%, and the average deviation ratio of the ARIMA model is 10.35. %, the overall prediction effect of the rolling gray prediction model is better. Fig. 6 is a deviation ratio area chart after the rolling gray prediction model and the ARIMA model experiment. If the area enclosed by the line and the X axis is smaller, the smaller the prediction deviation is, the higher the precision is. As shown The deviation of the ARIMA model shown is much larger than that of the rolling gray prediction model, which highlights the advantage of the rolling gray prediction model in cloud data center load forecasting.

Claims

The cloud data center energy-saving scheduling implementation method based on the rolling gray prediction model is characterized in that the energy-saving process of the cloud data center is abstracted into four modules: load prediction, error checking, thermal sensing classification and virtual machine scheduling; The prediction is performed to classify each host, and the virtual machine scheduling algorithm is used to achieve energy saving.
The cloud data center energy-saving scheduling implementation method based on the rolling gray prediction model according to claim 1, wherein the load prediction is specifically: predicting a workload of the data center by using a rolling gray prediction model, and obtaining data at a next moment. The load utilization of each host node in the center.
The cloud data center energy-saving scheduling implementation method based on the rolling gray prediction model according to claim 1, wherein the error verification is specifically: performing error checking on the load prediction value and the actual workload, and determining the current prediction result. The deviation value is calculated based on the error check module to correct the prediction result.
The cloud data center energy-saving scheduling implementation method based on the rolling gray prediction model according to claim 1, wherein the heat-aware classification is: performing heat-sensing classification on all hosts in the cloud data center according to the current load prediction value of the host. And introduce a service level agreement as a reference indicator to set the upper and lower bounds of the host workload threshold; when the host load utilization is above the upper threshold, below the threshold lower bound, between the upper and lower thresholds, and the load utilization is 0. These four cases are divided into four different thermal states.
The cloud data center energy-saving scheduling implementation method based on the rolling gray prediction model according to claim 1, wherein the virtual machine scheduling is specifically: scheduling virtual machines according to a hot state of the current host, and scheduling operations through the virtual machine. Solve the problem of host overload and no load, and maintain the hosts in the data center in a healthy hot state.
The cloud data center energy-saving scheduling implementation method based on the rolling gray prediction model according to any one of claims 1 to 5, characterized in that the method is implemented in a cloud data center environment by using a rolling gray prediction model algorithm and in a modular form. To intelligently monitor the load information of the data center and the performance data of the runtime.