WO2019024445A1 - 地理分布交互服务云资源协同优化方法 - Google Patents

地理分布交互服务云资源协同优化方法 Download PDF

Info

Publication number
WO2019024445A1
WO2019024445A1 PCT/CN2018/072031 CN2018072031W WO2019024445A1 WO 2019024445 A1 WO2019024445 A1 WO 2019024445A1 CN 2018072031 W CN2018072031 W CN 2018072031W WO 2019024445 A1 WO2019024445 A1 WO 2019024445A1
Authority
WO
WIPO (PCT)
Prior art keywords
resource
service
data center
delay
resources
Prior art date
Application number
PCT/CN2018/072031
Other languages
English (en)
French (fr)
Inventor
姚建国
吴家宏
管海兵
Original Assignee
上海交通大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 上海交通大学 filed Critical 上海交通大学
Publication of WO2019024445A1 publication Critical patent/WO2019024445A1/zh

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00Network arrangements or protocols for supporting network services or applications
    • H04L67/50Network services
    • H04L67/51Discovery or management thereof, e.g. service location protocol [SLP] or web services
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00Commerce
    • G06Q30/02Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201Market modelling; Market analysis; Collecting market data
    • G06Q30/0206Price or cost determination based on market factors
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00Network arrangements or protocols for supporting network services or applications
    • H04L67/01Protocols
    • H04L67/10Protocols in which an application is distributed across nodes in the network
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00Network arrangements or protocols for supporting network services or applications
    • H04L67/50Network services
    • H04L67/60Scheduling or organising the servicing of application requests, e.g. requests for application data transmissions using the analysis and optimisation of the required network resources

Definitions

  • the present invention relates to an optimization method, and in particular to a geographically distributed interactive service cloud resource collaborative optimization method.
  • Geographically distributed interactive services are a type of delay-sensitive computing service that needs to be deployed across multiple data center areas, such as data center areas in North America, Asia Pacific, and Europe.
  • interactive services such as web search and real-time data analysis
  • the user sends a web request to the data center of the service deployment to obtain data or perform decision analysis.
  • the interactive service relies on the basic data provided by multiple data center areas, and the basic data is globally distributed, it is decided that the geographically distributed interactive service needs to select a suitable data center for deployment, and this set of data centers is composed of At least one data center consisting of data center areas.
  • the data center needs to provide sufficient resources to run the data center workload supporting the interactive service.
  • the deployment location of the geographically distributed interactive service and the data center resource plan need to meet the user's tail delay requirement.
  • the ultimate goal is to minimize the overall computational and communication costs of geographically distributed interactive services.
  • Google Compute Engine Amazon AWS, Microsoft Azure, and Facebook Cloud have deployed a number of data centers globally by region, all of which provide a complete cloud resource usage solution.
  • Google Compute Engine and Microsoft Azure provide a strategy for using data center resources on demand. They provide VM instances with differentiated computing and storage capabilities and corresponding prices. In particular, the actual usage cost of the resource is calculated according to the usage time. For example, if a user rents a VM instance for 10 hours, the total cost is the unit price of the VM instance multiplied by 10 hours.
  • Amazon AWS not only provides a strategy similar to Google Compute Engine and Microsoft Azure on-demand resources, but Amazon AWS also allows users to set up reserved resources contracts, users can reserve a specific configuration of VM instances for 1 or 3 years and pre-pay resources The cost is left, and then the price is calculated at the lower price (about 50% of the on-demand resource price) according to the actual usage time.
  • the price of resources of each cloud service provider is different.
  • the price of data center resources in different regions of the same cloud service provider is also different, and the price of resources is constantly changing according to market demand.
  • Table 1 lists the price information for a cloud service provider partial VM instance.
  • the cloud service provider uses the total amount of the used traffic for the use of the WAN network bandwidth, and the unit price of the traffic is usually stepped. For example, AWS is free for the first 1GB, less than 10G for less than 10TB, and $0.090/GB for charging, and more than 10TB for less than 40TB for $0.085/GB.
  • an object of the present invention is to provide a geographically distributed interactive service cloud resource collaborative optimization method, which aims to minimize the resource allocation cost of a geographically distributed interactive service by a cloudy service provider.
  • a method for collaboratively optimizing cloud resource interaction service cloud resources comprises the following steps:
  • Step 1 Determine the data centers available in each area
  • Step 2 Determine that the interactive service can schedule the distribution of all possible data center combinations
  • Step 3 Design and agree on how data center resources are provided
  • Step 4 Determine a price demand probability distribution model corresponding to the random programming algorithm
  • Step 5 Determine the price model according to the resource provision plan formulated in step 3;
  • Step 6 Define the tail delay of the interactive interactive service and use it as a delay constraint for the optimization model
  • Step 7 Determine a data center resource scheduling optimization model
  • Step 8 Decompose the overall problem into a main problem and a series of sub-problems using the characteristics of stochastic programming
  • Step 9 The optimization algorithm solves, records the resource requirements of each interactive service and the corresponding delay information, determines the relationship between the quantity of resources and the delay, and is used to predict the preliminary workload scheduling; when the service goes online, the resource requirements and the delays correspondingly change, These changes will be used for random planning and ultimately for dynamic adjustment of resource plans.
  • the step 3 specifically includes the following steps:
  • Step 3.1 Only use the reserved resource mode; the data center provides reserved resources, the cloud consumer combines the service deployment cycle and the compromised resource requirements, firstly sets a reservation contract and pays the advance payment; after the interactive service goes online, the service resource is firstly The demand allocates reserved resources. If the reserved resources can meet the quality of service requirements during the deployment period, only the reserved resources satisfy the service quality requirements of the interactive services.
  • Step 3.2 Use only the on-demand resource mode; the actual demand of the service is due to the change of the request rate, the actual demand is greater than the maximum available resource of the lease, resulting in the failure to meet the required quality of service, and additional on-demand leases are reserved on the basis of reserved resources. Resources to meet the quality of service requirements, this situation is called the hybrid phase;
  • Step 3.3 Mixed mode, cloud consumers are more inclined to provide on-demand resources for short-term deployment, and data centers only provide on-demand resources.
  • the step 5 specifically includes the following steps:
  • Step 5.1 Determine the prepaid price of the reserved contract
  • Step 5.2 Determine the dynamic price of the actual running VM instance
  • Step 5.3 Determine the billing method and price of the network resource
  • Step 5.4 superimposing the cost of the computing resource and the network resource to obtain the total price
  • Step 5.5 Determine the dynamic cost of the instance's real-time operation.
  • the step 6 specifically includes the following steps:
  • Step 6.1 Determine the calculation delay and network transmission delay
  • Step 6.2 Determine a delay threshold for each request, and establish a delay threshold matrix
  • Step 6.3 Use the service level agreement to measure whether the tail delay meets the quality of service requirements under the conditions determined by the resource provision plan.
  • the step 6.3 specifically includes the following steps:
  • Step 6.3.1 Probabilistic analysis of historical data using probabilistic statistical methods, and obtain a preliminary relationship between resource provision planning and SLA tail delay.
  • the step 6.3.1 specifically comprises the following steps:
  • Step 6.3.1.1 Statistical analysis of calculation delay and network transmission delay probability
  • Step 6.3.1.2 Calculate the probability delay for each request
  • Step 6.3.1.3 Calculate the probability tail delay.
  • the step 8 comprises the following steps:
  • Step 8.1 Calculate the deployment cost of the candidate data center for each region
  • Step 8.2 Determine the main problem and sub-question of the random plan.
  • the step 9 comprises the following steps:
  • Step 9.1 Workload scheduling
  • Step 9.2 The workload resource plan is adjusted in real time.
  • the step 9.1 comprises the following steps:
  • Step 9.1.1 first generates a probability distribution of VCPU requirements, memory requirements, disk requirements, and network usage based on the request records of the geographically distributed interactive services;
  • Step 9.1.2 solves the formula (13) by using the generated probability distribution as the configuration data, and obtains the corresponding deployment cost of each schedulable data center;
  • Step 9.1.3 Initialize the scheduling decision variable.
  • the step 9.2 comprises the following steps:
  • Step 9.2.1 Obtain the target data center group of all interactive services by scheduling the workload, but the initial resource plan obtained is coarse-grained, and the resource plan needs to be adjusted at an RPP time, so correspondingly occurs in each RPP time.
  • Resource demand probability distribution
  • Step 9.2.2 solves the generated probability distribution as the configuration data, and obtains the resource plan of each workload in the target data center, and the resource plan is fine-grained;
  • Step 9.2.3 Our algorithm needs to select the target data center group for the geographically distributed interactive service at each WDD time and reconfigure the resource plan at each RPP time, where each WDD time contains multiple RPP times.
  • the present invention aims to minimize the resource allocation cost of a geographically distributed interactive service by a cloudy service provider. Because of the uncertainty of data center resource price and the characteristics of interactive service dynamic resource requirements, we design a price model and an SLA delay model. Based on the obtained constraints and combined with random programming, the workload scheduling and resource planning are described separately. The main method and a series of sub-problems are the essence of this method. This optimization design and its constraint method through price modeling and SLA delay modeling have better versatility and scalability for resource scheduling of most geographically distributed interactive services.
  • the present invention saves 24% in cost over service deployment based on on-demand resources. Our design saves 10% in cost over resource-based service deployment.
  • Figure 1 is a schematic diagram of the VCPU demand probability distribution.
  • Figure 2 is a schematic diagram of the probability distribution of memory requirements.
  • Figure 3 is a schematic diagram of the probability distribution of disk demand.
  • Figure 4 is a schematic diagram of the probability distribution of network usage.
  • Geographically distributed interactive services require the support of underlying data, and in order to provide high availability, the underlying data is typically backed up between data centers in the same region. Because of the backup of the underlying data, the interactive service can be deployed in a lower cost data center. However, each cloud service provider has a different price for the same configured VM instance and the same network traffic. In addition, users of interactive services at different times have different pressures on the service request, so the demand for resources of the interactive service is uncertain. Geographically distributed interactive services may change expected resource plans or even redeploy in lower-priced data centers because of changing resource requirements and changing data center resource prices. Therefore, the design principle of the present invention is to optimize the distribution of geographically distributed interactive service workloads separately and adjust corresponding resource plans according to resource requirements.
  • the method for collaboratively optimizing cloud resource interaction service cloud resources of the present invention comprises the following steps:
  • Step 1 Determine which data centers are available for each zone.
  • R data center areas that provide resources for the operation of geographically distributed interactive service workloads.
  • P cloud service providers
  • each cloud service provider has deployed DC i data centers, where i ⁇ P.
  • J r represent the number of data centers in region r.
  • Step 2 Determine that the interactive service can schedule the distribution of all possible data center combinations.
  • ) data centers which are referred to as a target data center group.
  • n data centers are provided by different areas, that is, each area needs to provide at least one data center to complete the operation of the workload. Therefore, let G denote all possible data center groups provided by R regions.
  • Step 3 Design and agree on how data center resources are provided. Because single use of reserved resources or single use of on-demand resources requires higher resource costs, we also consider reserved resources and on-demand resources. There are three resource usage scenarios based on this premise: only reserved, only on demand and mixed. Step 3 specifically includes the following steps:
  • Step 3.1 Only use the reserved resource mode; the data center provides reserved resources, the cloud consumer combines the service deployment cycle and the compromised resource requirements, firstly sets a reservation contract and pays the advance payment; after the interactive service goes online, the service resource is firstly The demand allocates reserved resources. If the reserved resources can meet the quality of service requirements during the deployment period, only the reserved resources can meet the service quality requirements of the interactive services.
  • Step 3.2 Use only the on-demand resource mode; the actual demand of the service may be due to the change of the request rate, and the actual demand is greater than the maximum available resource of the lease, resulting in failure to meet the required service quality. Additional subscriptions may be leased on the basis of reserved resources. Resources are required to meet the quality of service requirements, a situation known as the hybrid phase (reserved and on-demand).
  • Step 3.3 Mixed mode. Cloud consumers are more inclined to provide short-term deployment of services on demand resources, and data centers only provide on-demand resources.
  • Step 4 Determine a price demand probability distribution model corresponding to the random programming algorithm.
  • the price of data center resources is dynamically changing, and the resource requirements of geographically distributed interactive services will dynamically change based on the service request rate.
  • Uncertainty in price and demand can be handled using stochastic programming.
  • Stochastic programming uses a set of uncertain parameters described by probability distributions to decompose the problem into a main problem and a series of probability-related sub-problems. Solving the main problem can get a preliminary result, and then solve the sub-problem one by one to make up for the error of the preliminary solution of the main problem.
  • the step 4 includes the following steps: determining a probability distribution of each sub-question of the random plan.
  • the probability distribution of ⁇ is finite.
  • Step 5 Determine the price model based on the resource provision plan developed in step 3. Step 5 specifically includes the following steps:
  • Step 5.1 Determine the prepaid price for the reserved contract. Assuming that the cloud consumer creates a reserved resource contract k for the workload i in the data center j, the contract k describes a specific configured VM instance, Indicates the VM instance for the resource type Demand Indicates the resource type Unit price. Adding the price of the resource type required by the VM instance, you can get the prepaid price of the reserved contract k in the data center j.
  • Step 5.2 Determine the dynamic price of the actual running VM instance.
  • the VM instance is actually used in the case of ⁇
  • the reserved resource and the actual use price of the on-demand resource in time t can be obtained.
  • t is a billing cycle and also an RPP period, as shown in the following formula (4):
  • Step 5.3 Determine the billing method and price of the network resource.
  • WAN network resources are billed based on monthly traffic.
  • the WAN network resources will be charged at different unit prices.
  • K n represents the relationship between the amount of data transferred and the unit price.
  • Step 5.4 Superimpose the cost of the computing resource and the network resource to obtain the total price. If a geographically distributed interactive service is distributed to a data center group, a sufficient number of VM instances must be allocated for the workload (eg, 16 core CPU, 64G memory, and 500G SSD x 10 per VM instance). Therefore, we calculate the deployment cost of the geographically distributed interactive service according to the following formula, as shown in the following formula (5):
  • S represents the request source of m geographically distributed interactive services
  • I represents the workload set of each interactive service
  • J g represents all data center sets of the data center group g (g ⁇ G).
  • workload i the number of reserved VM instances that reserve contract k is set in data center j.
  • cloud consumers can make multiple reservation contracts in one data center.
  • X ij represents a binary variable equal to 1 if the workload i is assigned to data center j, otherwise 0.
  • Step 5.5 Determine the dynamic cost of the instance's real-time operation. It can be known from formula (5) that the resource allocation cost includes three parts: reserved Prepaid cost for a number of VM instances; WAN network usage cost; the cost of actually running a reserved VM instance and running an on-demand VM instance. Because the VM starts running according to actual needs, the cost of actually running the VM instance is dynamically changing. We use C Y to represent this dynamic cost. If the VM instance is retained Run cannot meet the tail delay requirement of the interactive service Then we need extra The number of VM instances causes the interactive service to meet the tail delay requirement, as shown in the following equation (6):
  • Step 6 Define the tail delay of the interactive interactive service and use it as a delay constraint for the optimization model.
  • the step 6 specifically includes the following steps:
  • Step 6.1 Determine the computation delay and network transmission delay. make with Represents data processing delay and WAN network transmission delay, respectively. Calculate the total delay of the service using Equation (7) below.
  • Step 6.2 Determine the delay threshold for each request and establish a delay threshold matrix.
  • Each service request has a corresponding delay requirement, which is expressed as a delay threshold.
  • Step 6.3 Use the service level agreement to measure whether the tail delay meets the QoS (Quality of Service) requirements under the conditions determined by the resource provision plan. Due to the very high cost of real-time prediction of the delay of each service request and the final reflection of the resource requirements, especially the delay prediction of a large number of interactive service requests is difficult to implement in reality. Therefore, we estimate the high percentage tail delay as an SLA (Service Level Agreement) assessment of the resource requirements of the interactive service, and this assessment is based on historical requests. For example, if we set x to a high percent SLA tail delay constraint, the probability that the service delay does not exceed the threshold must not be less than x%, otherwise the quality of service is not guaranteed.
  • SLA Service Level Agreement
  • Step 6.3 specifically includes the following steps:
  • Step 6.3.1 Probabilistic analysis of historical data by using a probabilistic statistical method, and obtaining a preliminary relationship between the resource providing plan and the SLA tail delay. Considering that a request source sends ⁇ requests to data center j for completion of workload i during t time, the probability tail delay is calculated using the following steps.
  • Step 6.3.1 specifically includes the following steps:
  • Step 6.3.1.1 Statistical analysis of computational delay and network transmission delay probability. make Indicates the probability that the computational delay of workload i in data center j does not exceed the calculated delay threshold, Indicates the probability that the WAN network transmission delay does not exceed the transmission delay threshold.
  • operating represents a method for calculating the probability of processing a data center processing when The result is 1, otherwise 0.
  • operation Used to estimate the network probability delay as shown in the following equation (9):
  • the average delay probability value obtained from a large number of requested delayed records should be close to the expected value of the delay.
  • the estimation of the delay probability will be more accurate when more request records are taken into account in the estimation.
  • Step 6.3.1.2 Calculate the probability delay for each request.
  • the delay of interactive services mainly includes two parts: a) data center processing delay. b) WAN network transmission delay.
  • the convolution function mathematically produces a third function that represents the superposition of the first two functions, so it is often seen here as a superposition of computational delay and network delay.
  • the convolution function can be used to obtain the expected probability delay for each service request sent to data center j, as in the following equation (10):
  • the operator "*" indicates the convolution method.
  • Step 6.3.1.3 Calculate the probability tail delay.
  • the interactive service needs to send the request to a specific data center group to complete the corresponding workload, so the probability tail delay of all service requests for an interactive service s should be averaged between the data center groups, as shown in the following equation (11) :
  • F sg (t) represents the probability tail delay of the Internet request sent by the interactive service s to the data center group g
  • F sg (t) acts as a constraint function for the scheduling decision of the interactive service and the resource provision plan.
  • Step 7 Determine the data center resource scheduling optimization model.
  • the scheduling of interactive services and resource planning can be obtained by solving the following optimization equations as shown in the following equation (12):
  • (11a) indicates that the workload of each interactive service needs to be assigned to a target data center
  • (12b) indicates that the tail delay of the interactive service should meet the tail delay SAL requirement
  • (12c) indicates that the allocated resource must be less than or equal to the data center.
  • the maximum resource constraint (12d) indicates that the number of reserved VM instances actually used does not exceed the maximum constraint of the reserved contract
  • (12e-12h) indicates that the number of VM instances is a natural number.
  • Step 8 Decompose the overall problem into a main problem and a series of sub-questions using the characteristics of stochastic programming.
  • the step 8 includes the following steps:
  • Step 8.1 Calculate the deployment cost of the candidate data center for each region using Equation (13). Taking the initial cost obtained as the weighting factor of the workload scheduling, using the integer programming to select the least cost data center group, the interactive service scheduling to an optimized data center group, and interacting with the service resident data in each WDD time interval. At the same time, at each RPP time, according to the change of demand, the resource plan of the target data center group is solved by using random programming to solve the time period as follows (13):
  • Step 8.2 Determine the main problem and sub-question of the random plan.
  • the main problem dispatches the interactive service to the lowest cost data center group according to the weight value; a series of sub-questions about the price and demand probability distribution dynamically adjust the resource provision plan according to the dynamic resource demand, as shown in the following formula (14):
  • Step 9 The optimization algorithm solves, records the resource requirements of each interactive service and the corresponding delay information, determines the relationship between the quantity of resources and the delay, and is used to predict the preliminary workload scheduling; when the service goes online, the resource requirements and the delays correspondingly change, These changes will be used for random planning and ultimately for dynamic adjustment of resource plans.
  • Step 9 includes the following steps:
  • Step 9.1 Workload scheduling.
  • Step 9.1 includes the following steps:
  • step 9.1.1 the probability distribution of VCPU requirements, memory requirements, disk requirements, and network usage is first generated based on the request records of the geographically distributed interactive services.
  • Step 9.1.2 using the generated probability distribution as the configuration data, solving the formula (13), and obtaining the corresponding deployment cost of each schedulable data center;
  • Step 9.1.3 initialize the scheduling decision variable, use the initial cost obtained in step 9.1.2 as the weight value, use formula (14) to select the data center with the lowest price as the target data center group, and the distribution of the workload needs to meet the maximum data center. Processing capacity constraints.
  • Step 9.2 The workload resource plan is adjusted in real time.
  • Step 9.2 includes the following steps:
  • Step 9.2.1 Through the scheduling of the workload, we can get the target data center group of all interactive services, but the initial resource plan obtained is coarse-grained, and the resource plan needs to be adjusted at one RPP time, so at each RPP time. Generate corresponding resource demand probability distributions;
  • Step 9.2.2 uses the generated probability distribution as the configuration data, and solves the sub-problem defined by the formula (15), and obtains the resource plan of each workload in the target data center, and the resource plan is fine-grained;
  • Step 9.2.3 Our algorithm needs to select the target data center group for the geographically distributed interactive service at each WDD time and reconfigure the resource plan at each RPP time, where each WDD time contains multiple RPP times.
  • the invention selects the most experimental configuration data of multiple data centers of the three cloud service providers of Google Compute Engine, Microsoft Azure and Amazon AWS, and selects three interactive services of Web Search, Real-time Data Analysis and Big Query and related services. Request records as the base test data. Experiments were conducted on the scheduling of workloads and the resource planning of interactive services.
  • the workflow scheduling process is as follows: First, according to the request record of the geographically distributed interactive service, the probability distribution of the VCPU requirement, the memory requirement, the disk requirement, and the network usage rate is generated. The experimental results are shown in Figure 1 to Figure 4; using the generated probability distribution as the configuration data, the formula (13) is solved, and the corresponding deployment cost of each schedulable data center is obtained; the initial scheduling decision variables are used. The cost is used as the weight value, and the data center with the lowest price is selected as the target data center group, and the distribution of the workload needs to meet the constraint of the maximum processing capacity of the data center.
  • the workflow of the workload resource planning is as follows: Through the scheduling of the workload, the target data center group of all interactive services can be obtained, but the initial resource plan obtained is coarse-grained, and the resource plan needs to be adjusted at an RPP time, so Each RPP time generates a corresponding resource demand probability distribution; using the generated probability distribution as the configuration data, and solving the sub-problem defined by the formula (15), the resource plan of each workload in the target data center can be obtained, and the resource plan is obtained. It is fine-grained; the target data center group needs to be selected for the geographically distributed interactive service at each WDD time, and the resource plan is reconfigured at each RPP time, where each WDD time contains multiple RPP times.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Accounting & Taxation (AREA)
  • Signal Processing (AREA)
  • Development Economics (AREA)
  • Finance (AREA)
  • Strategic Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Game Theory and Decision Science (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Economics (AREA)
  • Data Mining & Analysis (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明提供了一种地理分布交互服务云资源协同优化方法,包括以下步骤:步骤1:确定每个区域可用的数据中心;步骤2:确定交互服务可以调度分布所有可能数据中心组合;步骤3:设计并且约定数据中心资源提供方式;步骤4:确定与随机规划算法对应的价格需求概率分布模型;步骤5:根据步骤3制定的资源提供计划确定价格模型;步骤6:定义交互交互服务的尾延迟,并作为优化模型的延迟约束;步骤7:确定数据中心资源调度优化模型;步骤8:利用随机规划的特点把总体问题分解为一个主问题和一系列的子问题求解;步骤9:优化算法求解。本发明旨在通过多云服务提供商最小化地理分布交互服务的资源配置成本。

Description

地理分布交互服务云资源协同优化方法 技术领域
本发明涉及一种优化方法,具体地,涉及一种地理分布交互服务云资源协同优化方法。
背景技术
地理分布交互服务是一类需要部署到多个数据中心区域(如:北美、亚太以及欧洲等数据中心区域)的延迟敏感计算服务。对于web search、real-time data analysis等交互服务,用户发送Web请求到服务部署的数据中心获取数据或者进行决策分析。因为交互服务依赖于多个数据中心区域提供的基础数据,而这些基础数据是全球分布的,所以决定了地理分布交互服务需要选择一组合适的数据中心进行部署,而且这一组数据中心由各数据中心区域提供的至少一个数据中心组成。另外为了满足交互服务的部署要求,需要数据中心提供足够的资源运行支撑交互服务的数据中心工作负载。所以由于基础数据的依赖以及数据中心资源的需求,地理分布交互服务的部署位置以及数据中心资源计划需要满足用户的尾延迟要求。最终目标是最少化地理分布交互服务的总体计算和通信成本。
现今的云计算服务已经非常成熟。Google Compute Engine、Amazon AWS、Microsoft Azure以及阿里云等在全球范围按区域部署了一定数量的数据中心,这些数据中心均能提供完备的云资源使用方案。Google Compute Engine和Microsoft Azure提供了按需使用数据中心资源的策略。它们提供差异化的计算和存储能力的VM实例以及相应的价格。特别地,资源的实际使用花费是按照使用时间计算,例如某用户租用了一个VM实例10小时,那么总费用为VM实例的单价乘10小时。Amazon AWS不仅提供了类似Google Compute Engine和Microsoft Azure按需使用资源的策略,同时Amazon AWS还允许用户制定预留资源合约,用户可以预留1年或者3年特定配置的VM实例并且预先支付资源预留费用,然后以较低的价格(按需资源价格的50%左右)按照实际使用时间进行计价。一方面各个云服务提供商的资源价格各异,另一方面同一个云服务提供商不同区域的数据中心资源价格也不同,并且 根据市场需求资源价格在不断的变化。表1列举了云服务提供商部分VM实例的价格信息。因为交互服务通过Web请求请求访问数据中心获取计算服务,所以对于WAN网络带宽的使用,云服务提供商均按照使用流量的总量进行计价,并且流量的单价通常呈现阶梯型。例如AWS对于前1GB免费,大于1G少于10TB按照$0.090/GB收费,大于10TB少于40TB按照$0.085/GB收费。
因为地理分布交互服务是一类延迟敏感的Web服务,所以对于不同的服务请求率需要适当数量的数据中心资源提供用户要求期望的服务质量。在GoogleCompute Engine对于交互服务在不同时段请求率的统计信息中,可以得到服务请求率差异巨大,最低请求率与最高请求相差10 3倍以上。所以对于交互服务的部署,资源计划适应需求的不确定性是一个亟待解决的问题。尽管目前云服务提供商提供了基于预留或者按需使用资源的策略,但是都暴露了一些缺点。假如用户使用基于保留的资源,为了保证服务质量,用户必须使用预留大量的资源保证在请求率最大的时候交互服务依然能够满足用户的服务质量需求。但是因为需求的不确定性以及巨大的差异,这样就会造成大量资源的浪费,而且花费很高但很多时候没有发挥相应的作用。当用户仅使用按需资源时,虽然可以动态满足服务的资源需求,但是因为按需资源的价格比保留资源价格高50%以上,所以与利用预留资源策略的资源花费差异不大。基于这个前提需要设计一种能够混合预留资源与按需资源的策略、节以省地理分布交互服务的部署费用。
发明内容
针对现有技术中的缺陷,本发明的目的是提供一种地理分布交互服务云资源协同优化方法,其旨在通过多云服务提供商最小化地理分布交互服务的资源配置成本。
根据本发明的一个方面,提供一种地理分布交互服务云资源协同优化方法,其特征在于,包括以下步骤:
步骤1:确定每个区域可用的数据中心;
步骤2:确定交互服务可以调度分布所有可能数据中心组合;
步骤3:设计并且约定数据中心资源提供方式;
步骤4:确定与随机规划算法对应的价格需求概率分布模型;
步骤5:根据步骤3制定的资源提供计划确定价格模型;
步骤6:定义交互交互服务的尾延迟,并作为优化模型的延迟约束;
步骤7:确定数据中心资源调度优化模型;
步骤8:利用随机规划的特点把总体问题分解为一个主问题和一系列的子问题求解;
步骤9:优化算法求解,记录每一个交互服务的资源需求以及相应的延迟信息,确定资源数量与延迟的关系并用于预测初步的工作负载调度;当服务上线以后,资源需求以及延迟相应的变化,这些变化将用于随机规划并最终作用于资源计划的动态调整。
优选地,所述步骤3具体包括以下步骤:
步骤3.1:仅使用预留资源模式;数据中心提供预留资源,云消费者结合服务部署周期以及折中的资源需求首先制定预留合同和支付预付金;交互服务上线后,首先按照服务资源实际需求分配预留资源,如果预留资源能够在部署周期均能满足服务质量要求,则仅预留资源就满足交互服务的服务质量需求;
步骤3.2:仅使用按需资源模式;服务的实际需求因为请求率的变化而出现实际需求大于租约最大可用资源数量导致达不到要求的服务质量,在预留资源的基础上租用额外的按需资源以达到服务质量的要求,这种情况称为混合阶段;
步骤3.3:混合模式,云消费者更倾向于按需资源进行服务的短期部署,数据中心仅提供按需资源。
优选地,所述步骤5具体包括以下步骤:
步骤5.1:确定预留合约预付价格;
步骤5.2:确定实际运行VM实例的动态价格;
步骤5.3:确定网络资源计费方式与价格;
步骤5.4:将计算资源与网络资源的花费叠加得到总价格;
步骤5.5:确定实例实时运行的动态成本。
优选地,所述步骤6具体包括以下步骤:
步骤6.1:确定计算延迟与网络传输延迟;
步骤6.2:确定每个请求的延迟阈值,建立延迟阈值矩阵;
步骤6.3:利用服务等级协议衡量在资源提供计划确定条件下尾延迟是否满足服务质量要求。
优选地,所述步骤6.3具体包括以下步骤:
步骤6.3.1:利用概率统计方法,对历史数据进行概率分析,得到资源提供计划与SLA尾延迟的初步关系。
优选地,所述步骤6.3.1具体包括以下步骤:
步骤6.3.1.1:计算延迟与网络传输延迟概率统计分析;
步骤6.3.1.2:计算每个请求的概率延迟;
步骤6.3.1.3:计算概率尾延迟。
优选地,所述步骤8包括以下步骤:
步骤8.1:计算每一个区域的候选数据中心的部署成本;
步骤8.2:确定随机规划主问题与子问题。
优选地,所述步骤9包括以下步骤:
步骤9.1:工作负载调度;
步骤9.2:工作负载资源计划实时调整。
优选地,所述步骤9.1包括以下步骤:
步骤9.1.1首先根据地理分布交互服务的请求记录生成VCPU需求,内存需求,磁盘需求以及网络使用率的概率分布;
步骤9.1.2以生成的概率分布作为配置数据,对公式(13)进行求解,并得到每个可调度数据中心的相应部署成本;
步骤9.1.3初始化调度决策变量。
优选地,所述步骤9.2包括以下步骤:
步骤9.2.1通过对工作负载的调度,得到所有交互服务的目标数据中心组,但是得到的初步资源计划是粗粒度的,需要在么一个RPP时间调整资源计划,因此在每一个RPP时间产生相应的资源需求概率分布;
步骤9.2.2以生成的概率分布作为配置数据进行求解,得到每一个工作负载在目标数据中心的资源计划,该资源计划是细粒度的;
步骤9.2.3我们算法需要在每一个WDD时间为地理分布交互服务选择目标数据中心组,并在每一个RPP时间重新配置资源计划,其中每个WDD时间包含多个RPP时间。
与现有技术相比,本发明具有如下的有益效果:本发明旨在通过多云服务提供商最小化地理分布交互服务的资源配置成本。因为数据中心资源价格的不确定性,以及交互服务动态资源需求的特点,我们设计了一个价格模型以及一个SLA延迟模型,基于得到的约束条件并且结合随机规划把工作负载的调度以及资源计划分别描述为主方法以及一系列子问题是本方法的精髓。这种通过价格建模以及SLA延迟建模而得到的优化设计及其约束方法,对大部分的地理分布交互服务的资源调度有较 好的通用性与扩展性。本发明比基于按需资源的服务部署节省了24%的成本。我们的设计比基于预留资源的服务部署节省了10%的成本。我们的设计能够满足延迟的SLA约束,即能够满足基于协商的尾延迟约束。另外对于全球分布的交互服务还缺少相关的旨在减少服务部署成本的研究,某些相关的研究往往面向单个数据中心或者非延迟敏感应用的部署。因此我们开创性地研究了如何优化全球分布的交互服务部署成本,并设计了一种算法达到了相应的预期效果。
附图说明
通过阅读参照以下附图对非限制性实施例所作的详细描述,本发明的其它特征、目的和优点将会变得更明显:
图1为VCPU需求概率分布示意图。
图2为内存需求概率分布示意图。
图3为磁盘需求概率分布示意图。
图4为网络使用量概率分布示意图。
具体实施方式
下面结合具体实施例对本发明进行详细说明。以下实施例将有助于本领域的技术人员进一步理解本发明,但不以任何形式限制本发明。应当指出的是,对本领域的普通技术人员来说,在不脱离本发明构思的前提下,还可以做出若干变形和改进。这些都属于本发明的保护范围。
地理分布交互服务需要基础数据的支持,而为了提供高可用性,基础数据通常会在同一个区域的数据中心之间进行备份。因为基础数据的备份,交互服务可以选择价格较低的数据中心进行部署。但是,各个云服务提供商对于相同配置的VM实例以及相同的网络流量的价格是不同的。另外,不同时刻交互服务的用户对于服务的请求压力是变化的,因此造成交互服务对于资源的需求是不确定的。因为变化的资源需求以及变化的数据中心资源价格,地理分布交互服务可能会改变预期的资源计划甚至重新部署在价格更低的数据中心。因此本发明的设计原则是分别优化地理分布交互服务工作负载的分布以及根据资源需求调整相应的资源计划。
本发明地理分布交互服务云资源协同优化方法包括以下步骤:
步骤1:确定每个区域可用的数据中心。假设有R个数据中心区域为地理分布交 互服务工作负载的运行提供资源。在每个区域,存在P个云服务提供商并且每个云服务提供商已经部署了DC i个数据中心,其中i∈P。令J r表示区域r的数据中心数量,我们可以用下面的公式计算区域r中所有可用数据中心,如下式(1):
Figure PCTCN2018072031-appb-000001
步骤2:确定交互服务可以调度分布所有可能数据中心组合。假设存在m个地理分布交互服务,每个服务都包含多于|R|个工作负载,这些工作负载需要分布到一个数据中心组。例如,用户向n(n≥|R|)个数据中心发送服务请求,这n个数据中心称为一个目标数据中心组。特别地,n个数据中心由不同的区域提供,即每个区域都需要提供至少一个数据中心完成工作负载的运行。因此,令G表示R个区域提供的所有可能数据中心组。我们使用组合数学理论计算所有可用的数据中心组,如下式(2):
Figure PCTCN2018072031-appb-000002
步骤3:设计并且约定数据中心资源提供方式。因为单一使用预留资源或者单一使用按需资源都需要较高的资源成本,因此我们同时考虑预留资源和按需资源。基于这个前提有三种资源使用情况:仅预留,仅按需和混合。步骤3具体包括以下步骤:
步骤3.1:仅使用预留资源模式;数据中心提供预留资源,云消费者结合服务部署周期以及折中的资源需求首先制定预留合同和支付预付金;交互服务上线后,首先按照服务资源实际需求分配预留资源,如果预留资源能够在部署周期均能满足服务质量要求,则仅预留资源就可以满足交互服务的服务质量需求。
步骤3.2:仅使用按需资源模式;服务的实际需求因为请求率的变化而出现实际需求大于租约最大可用资源数量导致达不到要求的服务质量,可以在预留资源的基础上租用额外的按需资源以达到服务质量的要求,这种情况称为混合阶段(预留和按需)。
步骤3.3:混合模式。云消费者更倾向于按需资源进行服务的短期部署,数据中心仅提供按需资源。
步骤4:确定与随机规划算法对应的价格需求概率分布模型。数据中心资源的 价格是动态变化的,另外地理分布交互服务的资源需求将根据服务请求率动态变化。利用随机规划可以处理价格与需求的不确定性。随机规划采用一组由概率分布描述的不确定参数,把问题分解为一个主问题以及一系列概率相关的子问题。求解主问题可以得到一个初步的结果,然后逐一求解子问题弥补主问题初步求解的误差。所述步骤4包括以下步骤:确定随机规划各子问题的概率分布。根据上述的三种资源提供计划,我们分析价格以及需求的概率属性。令Λ t表示每个RPP(资源计划)时间t中的价格和需求情况的集合,对于一个WDD(工作负载调度)时间T中的所有价格与需求情况,Λ被定义为如下式(3):
Figure PCTCN2018072031-appb-000003
Λ的概率分布是有限的,例如集合Λ代表有限数量的价格与需求情况,用P λ∈[0,1]表示每一种情况出现的概率,其中λ是一个复合变量,λ=(λ 1,λ 2,…,λ |T|)∈Λ。
步骤5:根据步骤3制定的资源提供计划确定价格模型。步骤5具体包括以下步骤:
步骤5.1:确定预留合约预付价格。假设云消费者为工作负载i在数据中心j制定预留资源合约k时,合约k描述了一种特定配置的VM实例,令
Figure PCTCN2018072031-appb-000004
表示该VM实例对于资源类型
Figure PCTCN2018072031-appb-000005
的需求量,令
Figure PCTCN2018072031-appb-000006
表示资源类型
Figure PCTCN2018072031-appb-000007
的单位价格。把VM实例需要的资源类型价格相加,即可得到工作负载i在数据中心j制定预留合约k的预付价格
Figure PCTCN2018072031-appb-000008
步骤5.2:确定实际运行VM实例的动态价格。当在λ情况下实际使用该VM实例,可以得到预留资源以及按需资源在时间t内的实际使用价格
Figure PCTCN2018072031-appb-000009
Figure PCTCN2018072031-appb-000010
因为实际使用资源需要按照使用时间计费,其中t为一个计费周期同时也是一个RPP周期,如下式(4):
Figure PCTCN2018072031-appb-000011
步骤5.3:确定网络资源计费方式与价格。WAN网络资源根据每月传输流量进行计费。当数据传输超过网络价格策略中设置的最大流量时,WAN网络资源将以不 同的单价收费。令K n表示传输的数据量与单位价格之间的关系。例如,Amazon EC2的WAN网络价格策略为:K n={“前1GB”:“$0.000/GB”,“大于1G少于10TB”:“$0.090/GB”,“大于10T少于40TB”:“$0.085/GB”,“大于40TB少于100TB”:“$0.070/GB”,“大于100TB少于350TB”:“$0.050/GB”,...}。因此,令
Figure PCTCN2018072031-appb-000012
表示WAN网络流量的单价。
步骤5.4:将计算资源与网络资源的花费叠加得到总价格。如果将地理分布交互服务分发到数据中心组,则必须为工作负载分配足够数量的VM实例(例如,16核CPU,64G内存和每个VM实例500G SSD×10)。因此我们根据下面的公式计算地理分布互动服务的部署成本,如下式(5):
Figure PCTCN2018072031-appb-000013
式中,S表示m个地理分布交互服务的请求源,I表示每个交互服务的工作负载集合,J g表示数据中心组g(g∈G)的所有数据中心集合。
Figure PCTCN2018072031-appb-000014
表示为工作负载i在数据中心j制定预留合约k的保留VM实例数量,另外云消费者可以在一个数据中心制定多个预留合约。
Figure PCTCN2018072031-appb-000015
表示每个网络资源计费周期的网络数据传输量,其中T n通常表示为一年12个网络资源计费周期。X ij表示一个二进制变量,如果工作量i分配给数据中心j,则等于1,否则为0。
步骤5.5:确定实例实时运行的动态成本。由公式(5)可以知道,资源配置成本包括三部分:预留
Figure PCTCN2018072031-appb-000016
数量VM实例的预付成本;WAN网络使用成本;实际运行预留VM实例和运行按需VM实例的成本。因为VM按照实际需求启动运行,所以实际运行VM实例的成本是动态变化的。我们用C Y表示该动态成本。如果保留的VM实例
Figure PCTCN2018072031-appb-000017
的运行不能满足交互服务的尾延迟要求
Figure PCTCN2018072031-appb-000018
那么我们需要额外
Figure PCTCN2018072031-appb-000019
数量的VM实例使交互服务达到尾延迟要求,如下式(6):
Figure PCTCN2018072031-appb-000020
步骤6:定义交互交互服务的尾延迟,并作为优化模型的延迟约束。所述步骤6具体包括以下步骤:
步骤6.1:确定计算延迟与网络传输延迟。令
Figure PCTCN2018072031-appb-000021
Figure PCTCN2018072031-appb-000022
分别表示数据处理延 迟和WAN网络传输延迟。利用下面的公式(7)计算服务的总延迟。
Figure PCTCN2018072031-appb-000023
步骤6.2:确定每个请求的延迟阈值,建立延迟阈值矩阵。每个服务请求都有相应的延迟要求,具体表现为延迟阈值。我们独立地考虑每一个服务请求,如果所有的服务请求都满足各自的延迟要求,我们认为地理分布交互服务可以满足服务尾延迟要求。因此,令
Figure PCTCN2018072031-appb-000024
Figure PCTCN2018072031-appb-000025
分别表示处理延迟阈值和WAN网络传输延迟阈值,令L(t)表示发送到n(n=|J g|)个数据中心的服务请求的延迟约束阈值,我们可以得到下面的延迟阈值矩阵,如下式(8):
Figure PCTCN2018072031-appb-000026
步骤6.3:利用服务等级协议衡量在资源提供计划确定条件下尾延迟是否满足QoS(服务质量)要求。由于对每一个服务请求的延迟进行实时预测并最终反映为资源要求的开销非常大,特别对于大量的交互服务请求的延迟预测在现实中难以实现。因此,我们将高百分比尾延迟作为SLA(服务等级协议)对交互服务的资源需求进行评估预测,而且这种评估是基于历史请求的。例如,如果我们将x设置为高百分位SLA尾延迟约束,服务延迟不超过阈值的概率不得低于x%,否则服务质量不能保证。高百分位SLA尾延迟提供了一种让地理分布交互服务满足服务质量要求的方法。步骤6.3具体包括以下步骤:步骤6.3.1:利用概率统计方法对历史数据进行概率分析,得到资源提供计划与SLA尾延迟的初步关系。考虑一个请求源在t时间内为完成工作负载i的执行将Ω个请求发送到数据中心j,利用下面的步骤计算概率尾延迟。步骤6.3.1具体包括以下步骤:
步骤6.3.1.1:计算延迟与网络传输延迟概率统计分析。令
Figure PCTCN2018072031-appb-000027
表示工作负载i在数据中心j的计算延迟不超过计算延迟阈值的概率,令
Figure PCTCN2018072031-appb-000028
表示WAN网络传输延迟不超过传输延迟阈值的概率。操作
Figure PCTCN2018072031-appb-000029
表示一个用于计算数据中心处理概率延迟的方法,当
Figure PCTCN2018072031-appb-000030
时,结果为1,否则为0。类似地,操作
Figure PCTCN2018072031-appb-000031
用于估算网络概率延迟,如下式(9):
Figure PCTCN2018072031-appb-000032
根据大数法则(LLN),从大量请求的延迟记录获得的平均延迟概率值应接近延迟的预期值。当在估计中考虑到更多的请求记录时,延迟概率的估算将更准确。
步骤6.3.1.2:计算每个请求的概率延迟。交互服务的延迟主要包括两部分:a)数据中心处理延迟。b)WAN网络传输延迟。卷积函数在数学上产生第三个函数表示前两个函数的叠加,所以在这里通常被视为计算延迟和网络延迟的叠加。对于工作负载i,利用卷积函数可以得到发送到数据中心j的每个服务请求的预期概率延迟,如下式(10):
Figure PCTCN2018072031-appb-000033
其中运算符“*”表示卷积方法。
步骤6.3.1.3:计算概率尾延迟。交互式服务需要把请求发送到特定的数据中心组中完成相应的工作负载,因此对于一个交互服务s的所有服务请求的概率尾延迟应在数据中心组之间取平均概率,如下式(11):
Figure PCTCN2018072031-appb-000034
其中F sg(t)表示交互服务s的互联网请求发送到数据中心组g的概率尾延迟,F sg(t)作为交互服务的调度决策和资源提供计划的约束函数。
步骤7:确定数据中心资源调度优化模型。通过价格建模以及延迟约束建模,交互服务的调度以及资源计划可以通过求解下面的优化方程得到如下式(12)等:
Figure PCTCN2018072031-appb-000035
Figure PCTCN2018072031-appb-000036
Figure PCTCN2018072031-appb-000037
Figure PCTCN2018072031-appb-000038
Figure PCTCN2018072031-appb-000039
Figure PCTCN2018072031-appb-000040
Figure PCTCN2018072031-appb-000041
Figure PCTCN2018072031-appb-000042
Figure PCTCN2018072031-appb-000043
Figure PCTCN2018072031-appb-000044
其中(11a)表示每一个交互服务的工作负载需要分配到一个目标数据中心,(12b)表示交互服务的尾延迟应该满足尾延迟SAL要求,(12c)表示分配的资源必须小于或者等于数据中心的最大资源约束,(12d)表示实际使用的预留VM实例数量不超过预留合约的最大约束,(12e-12h)表示VM实例的数量为自然数。
步骤8:利用随机规划的特点把总体问题分解为一个主问题和一系列的子问题求解。所述步骤8包括以下步骤:
步骤8.1:利用公式(13)计算每一个区域的候选数据中心的部署成本。把得到的初步成本作为工作负载调度的权重因子,利用整数规划选择成本最少的数据中心组,交互服务调度到一个最优化的数据中心组,并且在每一个WDD时间区间内交互服务常驻这个数据中心组;同时在每一个RPP时间,根据需求的变化利用随机规划求解该时间段在目标数据中心组的资源计划如下式(13)等:
Figure PCTCN2018072031-appb-000045
Figure PCTCN2018072031-appb-000046
Figure PCTCN2018072031-appb-000047
步骤8.2:确定随机规划主问题与子问题。主问题根据权重值把交互服务调度到花费最低的数据中心组;一系列关于价格与需求概率分布的子问题根据动态的资源需求动态地调整资源提供计划,如下式(14)等:
Figure PCTCN2018072031-appb-000048
Figure PCTCN2018072031-appb-000049
Figure PCTCN2018072031-appb-000050
Figure PCTCN2018072031-appb-000051
Figure PCTCN2018072031-appb-000052
因为价格的不确定性,需要在每一个WDD时间重复上面的算法进行工作负载的调度以及资源计划。
步骤9:优化算法求解,记录每一个交互服务的资源需求以及相应的延迟信息,确定资源数量与延迟的关系并用于预测初步的工作负载调度;当服务上线以后,资源需求以及延迟相应的变化,这些变化将用于随机规划并最终作用于资源计划的动态调整。步骤9包括以下步骤:
步骤9.1:工作负载调度。步骤9.1包括以下步骤:
步骤9.1.1,首先根据地理分布交互服务的请求记录生成VCPU需求,内存需求,磁盘需求以及网络使用率的概率分布。
步骤9.1.2,以生成的概率分布作为配置数据,对公式(13)进行求解,并得到每个可调度数据中心的相应部署成本;
步骤9.1.3,初始化调度决策变量,利用步骤9.1.2得到的初步成本作为权重值,利用公式(14)选择价格最低的数据中心作为目标数据中心组,而且工作负载的分布需要满足数据中心最大处理能力的约束。
步骤9.2:工作负载资源计划实时调整。步骤9.2包括以下步骤:
步骤9.2.1通过对工作负载的调度,我们可以得到所有交互服务的目标数据中心组,但是得到的初步资源计划是粗粒度的,需要在么一个RPP时间调整资源计划,因此在每一个RPP时间产生相应的资源需求概率分布;
步骤9.2.2以生成的概率分布作为配置数据,利用公式(15)定义的子问题进行求解,可以得到每一个工作负载在目标数据中心的资源计划,该资源计划是细粒度的;
步骤9.2.3我们算法需要在每一个WDD时间为地理分布交互服务选择目标数据中心组,并在每一个RPP时间重新配置资源计划,其中每个WDD时间包含多个RPP时间。
本发明选取了Google Compute Engine,Microsoft Azure和Amazon AWS三个云服务提供商的多个数据中心最为实验配置数据,选择了Web Search,Real-time Data Analysis和Big Query三种交互服务以及相关的服务请求记录作为基础测试数据。分别对工作负载的调度以及交互服务的资源计划进行了实验。
其中,工作负载调度的流程如下:首先根据地理分布交互服务的请求记录生成VCPU需求,内存需求,磁盘需求以及网络使用率的概率分布。实验结果如图1至图4所示;以生成的概率分布作为配置数据,对公式(13)进行求解,并得到每个可调度数据中心的相应部署成本;初始化调度决策变量,利用到的初步成本作为权重值,利用选择价格最低的数据中心作为目标数据中心组,而且工作负载的分布需要满足数据中心最大处理能力的约束。
而工作负载资源计划的流程如下:通过对工作负载的调度,可以得到所有交互服务的目标数据中心组,但是得到的初步资源计划是粗粒度的,需要在么一个RPP时间调整资源计划,因此在每一个RPP时间产生相应的资源需求概率分布;以生成的概率分布作为配置数据,利用公式(15)定义的子问题进行求解,可以得到每一个工作负载在目标数据中心的资源计划,该资源计划是细粒度的;需要在每一个WDD时间为地理分布交互服务选择目标数据中心组,并在每一个RPP时间重新配置资源计划,其中每个WDD时间包含多个RPP时间。
以上对本发明的具体实施例进行了描述。需要理解的是,本发明并不局限于上述特定实施方式,本领域技术人员可以在权利要求的范围内做出各种变形或修改,这并不影响本发明的实质内容。

Claims (10)

  1. 一种地理分布交互服务云资源协同优化方法,其特征在于,包括以下步骤:
    步骤1:确定每个区域可用的数据中心;
    步骤2:确定交互服务可以调度分布所有可能数据中心组合;
    步骤3:设计并且约定数据中心资源提供方式;
    步骤4:确定与随机规划算法对应的价格需求概率分布模型;
    步骤5:根据步骤3制定的资源提供计划确定价格模型;
    步骤6:定义交互交互服务的尾延迟,并作为优化模型的延迟约束;
    步骤7:确定数据中心资源调度优化模型;
    步骤8:利用随机规划的特点把总体问题分解为一个主问题和一系列的子问题求解;
    步骤9:优化算法求解,记录每一个交互服务的资源需求以及相应的延迟信息,确定资源数量与延迟的关系并用于预测初步的工作负载调度;当服务上线以后,资源需求以及延迟相应的变化,这些变化将用于随机规划并最终作用于资源计划的动态调整。
  2. 根据权利要求1所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤3具体包括以下步骤:
    步骤3.1:仅使用预留资源模式;数据中心提供预留资源,云消费者结合服务部署周期以及折中的资源需求首先制定预留合同和支付预付金;交互服务上线后,首先按照服务资源实际需求分配预留资源,如果预留资源能够在部署周期均能满足服务质量要求,则仅预留资源就满足交互服务的服务质量需求;
    步骤3.2:仅使用按需资源模式;服务的实际需求因为请求率的变化而出现实际需求大于租约最大可用资源数量导致达不到要求的服务质量,在预留资源的基础上租用额外的按需资源以达到服务质量的要求,这种情况称为混合阶段;
    步骤3.3:混合模式,云消费者更倾向于按需资源进行服务的短期部署,数据中心仅提供按需资源。
  3. 根据权利要求1所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤5具体包括以下步骤:
    步骤5.1:确定预留合约预付价格;
    步骤5.2:确定实际运行VM实例的动态价格;
    步骤5.3:确定网络资源计费方式与价格;
    步骤5.4:将计算资源与网络资源的花费叠加得到总价格;
    步骤5.5:确定实例实时运行的动态成本。
  4. 根据权利要求1所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤6具体包括以下步骤:
    步骤6.1:确定计算延迟与网络传输延迟;
    步骤6.2:确定每个请求的延迟阈值,建立延迟阈值矩阵;
    步骤6.3:利用服务等级协议衡量在资源提供计划确定条件下尾延迟是否满足服务质量要求。
  5. 根据权利要求4所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤6.3具体包括以下步骤:
    步骤6.3.1:利用概率统计方法,对历史数据进行概率分析,得到资源提供计划与SLA尾延迟的初步关系。
  6. 根据权利要求5所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤6.3.1具体包括以下步骤:
    步骤6.3.1.1:计算延迟与网络传输延迟概率统计分析;
    步骤6.3.1.2:计算每个请求的概率延迟;
    步骤6.3.1.3:计算概率尾延迟。
  7. 根据权利要求1所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤8包括以下步骤:
    步骤8.1:计算每一个区域的候选数据中心的部署成本;
    步骤8.2:确定随机规划主问题与子问题。
  8. 根据权利要求1所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤9包括以下步骤:
    步骤9.1:工作负载调度;
    步骤9.2:工作负载资源计划实时调整。
  9. 根据权利要求8所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤9.1包括以下步骤:
    步骤9.1.1首先根据地理分布交互服务的请求记录生成VCPU需求,内存需求,磁盘需求以及网络使用率的概率分布;
    步骤9.1.2以生成的概率分布作为配置数据,对公式(13)进行求解,并得到每个可调度数据中心的相应部署成本;
    步骤9.1.3初始化调度决策变量。
  10. 根据权利要求8所述的地理分布交互服务云资源协同优化方法,其特征在于,所述步骤9.2包括以下步骤:
    步骤9.2.1通过对工作负载的调度,得到所有交互服务的目标数据中心组,但是得到的初步资源计划是粗粒度的,需要在么一个RPP时间调整资源计划,因此在每一个RPP时间产生相应的资源需求概率分布;
    步骤9.2.2以生成的概率分布作为配置数据进行求解,得到每一个工作负载在目标数据中心的资源计划,该资源计划是细粒度的;
    步骤9.2.3我们算法需要在每一个WDD时间为地理分布交互服务选择目标数据中心组,并在每一个RPP时间重新配置资源计划,其中每个WDD时间包含多个RPP时间。
PCT/CN2018/072031 2017-07-31 2018-01-10 地理分布交互服务云资源协同优化方法 WO2019024445A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710639715.5A CN107395733B (zh) 2017-07-31 2017-07-31 地理分布交互服务云资源协同优化方法
CN201710639715.5 2017-07-31

Publications (1)

Publication Number Publication Date
WO2019024445A1 true WO2019024445A1 (zh) 2019-02-07

Family

ID=60342409

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/072031 WO2019024445A1 (zh) 2017-07-31 2018-01-10 地理分布交互服务云资源协同优化方法

Country Status (2)

Country Link
CN (1) CN107395733B (zh)
WO (1) WO2019024445A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110888687A (zh) * 2019-09-27 2020-03-17 华北水利水电大学 基于合约设计的移动边缘计算任务卸载最优合约设计方法
CN111459670A (zh) * 2020-03-30 2020-07-28 中科边缘智慧信息科技(苏州)有限公司 一种在边缘计算不同层次进行协同处理的方法
CN112363829A (zh) * 2020-11-03 2021-02-12 武汉理工大学 一种基于弹性尺度聚合的用户资源动态分配方法
CN112783649A (zh) * 2021-01-19 2021-05-11 苏州博纳讯动软件有限公司 一种面向云计算的交互感知的容器化微服务资源调度方法
CN112860407A (zh) * 2021-03-02 2021-05-28 东南大学 一种云边协作的序列依赖流水车间任务调度执行优化方法
CN113282413A (zh) * 2021-05-20 2021-08-20 南京航空航天大学 车辆边缘计算网络中QoS需求自适应的资源配置方法
CN113626162A (zh) * 2021-07-09 2021-11-09 西安电子科技大学 基于动态资源共享的数据中心任务混合部署方法及系统
CN114697209A (zh) * 2022-03-30 2022-07-01 广州穗华能源科技有限公司 一种云边协同计算资源配置方法及配置系统
CN115189999A (zh) * 2022-07-20 2022-10-14 贵州电网有限责任公司 用于对云计算服务进行管理的系统及方法

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107395733B (zh) * 2017-07-31 2020-08-04 上海交通大学 地理分布交互服务云资源协同优化方法
CN108848514B (zh) * 2018-06-20 2021-08-03 中国联合网络通信集团有限公司 数据通信优化方法及数据通信优化器
CN109325266B (zh) * 2018-08-29 2023-11-10 天津大学 面向在线云服务的响应时间分布预测方法
CN111507554B (zh) * 2019-01-31 2022-07-22 北京骑胜科技有限公司 一种服务资源调度方法、装置、设备及存储介质
TWI723410B (zh) 2019-05-31 2021-04-01 伊雲谷數位科技股份有限公司 雲端資源管理系統、雲端資源管理方法以及非暫態電腦可讀取記錄媒體
CN111988388B (zh) * 2020-08-13 2023-05-16 北京达佳互联信息技术有限公司 流量分配的方法、装置、电子设备及存储介质
CN113642806A (zh) * 2021-08-31 2021-11-12 黄杉 一种基于云计算的优化设计服务系统和方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103002005A (zh) * 2011-09-07 2013-03-27 埃森哲环球服务有限公司 云服务监测系统
US20130151688A1 (en) * 2011-12-07 2013-06-13 Alcatel-Lucent Usa Inc. Optimization mechanisms for latency reduction and elasticity improvement in geographically distributed data centers
CN103530801A (zh) * 2013-10-29 2014-01-22 东南大学 一种基于动态定价策略的多数据中心成本优化方法
US20140282716A1 (en) * 2013-03-15 2014-09-18 Brightroll, Inc. Geo, segment, uniques distributed computing system
CN104065663A (zh) * 2014-07-01 2014-09-24 复旦大学 一种基于混合云调度模型的自动伸缩、费用优化的内容分发服务方法
CN107395733A (zh) * 2017-07-31 2017-11-24 上海交通大学 地理分布交互服务云资源协同优化方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103002005A (zh) * 2011-09-07 2013-03-27 埃森哲环球服务有限公司 云服务监测系统
US20130151688A1 (en) * 2011-12-07 2013-06-13 Alcatel-Lucent Usa Inc. Optimization mechanisms for latency reduction and elasticity improvement in geographically distributed data centers
US20140282716A1 (en) * 2013-03-15 2014-09-18 Brightroll, Inc. Geo, segment, uniques distributed computing system
CN103530801A (zh) * 2013-10-29 2014-01-22 东南大学 一种基于动态定价策略的多数据中心成本优化方法
CN104065663A (zh) * 2014-07-01 2014-09-24 复旦大学 一种基于混合云调度模型的自动伸缩、费用优化的内容分发服务方法
CN107395733A (zh) * 2017-07-31 2017-11-24 上海交通大学 地理分布交互服务云资源协同优化方法

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110888687B (zh) * 2019-09-27 2023-03-14 华北水利水电大学 基于合约设计的移动边缘计算任务卸载最优合约设计方法
CN110888687A (zh) * 2019-09-27 2020-03-17 华北水利水电大学 基于合约设计的移动边缘计算任务卸载最优合约设计方法
CN111459670A (zh) * 2020-03-30 2020-07-28 中科边缘智慧信息科技(苏州)有限公司 一种在边缘计算不同层次进行协同处理的方法
CN112363829A (zh) * 2020-11-03 2021-02-12 武汉理工大学 一种基于弹性尺度聚合的用户资源动态分配方法
CN112363829B (zh) * 2020-11-03 2024-03-29 武汉理工大学 一种基于弹性尺度聚合的用户资源动态分配方法
CN112783649A (zh) * 2021-01-19 2021-05-11 苏州博纳讯动软件有限公司 一种面向云计算的交互感知的容器化微服务资源调度方法
CN112783649B (zh) * 2021-01-19 2024-03-01 江苏博云科技股份有限公司 一种面向云计算的交互感知的容器化微服务资源调度方法
CN112860407A (zh) * 2021-03-02 2021-05-28 东南大学 一种云边协作的序列依赖流水车间任务调度执行优化方法
CN113282413A (zh) * 2021-05-20 2021-08-20 南京航空航天大学 车辆边缘计算网络中QoS需求自适应的资源配置方法
CN113282413B (zh) * 2021-05-20 2024-03-05 南京航空航天大学 车辆边缘计算网络中QoS需求自适应的资源配置方法
CN113626162A (zh) * 2021-07-09 2021-11-09 西安电子科技大学 基于动态资源共享的数据中心任务混合部署方法及系统
CN114697209A (zh) * 2022-03-30 2022-07-01 广州穗华能源科技有限公司 一种云边协同计算资源配置方法及配置系统
CN114697209B (zh) * 2022-03-30 2023-12-22 广州穗华能源科技有限公司 一种云边协同计算资源配置方法及配置系统
CN115189999A (zh) * 2022-07-20 2022-10-14 贵州电网有限责任公司 用于对云计算服务进行管理的系统及方法
CN115189999B (zh) * 2022-07-20 2023-08-22 贵州电网有限责任公司 用于对云计算服务进行管理的系统及方法

Also Published As

Publication number Publication date
CN107395733B (zh) 2020-08-04
CN107395733A (zh) 2017-11-24

Similar Documents

Publication Publication Date Title
WO2019024445A1 (zh) 地理分布交互服务云资源协同优化方法
CN111480145B (zh) 用于根据基于信用的机制来调度工作负载的系统和方法
Samaan A novel economic sharing model in a federation of selfish cloud providers
Lucas-Simarro et al. Scheduling strategies for optimal service deployment across multiple clouds
CN110858161A (zh) 资源分配方法、装置、系统、设备和介质
KR101903623B1 (ko) 버스트 모드 제어
US20130111027A1 (en) Accessing physical resources in a cloud computing environment
Cong et al. A survey of profit optimization techniques for cloud providers
US10886743B2 (en) Providing energy elasticity services via distributed virtual batteries
CN104065663A (zh) 一种基于混合云调度模型的自动伸缩、费用优化的内容分发服务方法
Jie et al. Online task scheduling for edge computing based on repeated Stackelberg game
CN104038540A (zh) 一种应用代理服务器自动选择方法及系统
Lucas‐Simarro et al. Cost optimization of virtual infrastructures in dynamic multi‐cloud scenarios
Jiao et al. Cost optimization for online social networks on geo-distributed clouds
Ang et al. Dynamic pricing scheme for resource allocation in multi-cloud environment
Wanis et al. Efficient modeling and demand allocation for differentiated cloud virtual-network as-a service offerings
Jayanthi Literature review: Dynamic resource allocation mechanism in cloud computing environment
Fan et al. Knative autoscaler optimize based on double exponential smoothing
Subramanian et al. A study on optimized resource provisioning in federated cloud
CN112583609A (zh) 一种云计费方法、装置、云管理平台、系统及存储介质
Yao et al. Optimal overbooking policy for cloud service providers: Profit and service quality
Vanmechelen et al. A simulation framework for studying economic resource management in grids
Li et al. A research of resource scheduling strategy for cloud computing based on Pareto optimality M× N production model
Depoorter et al. Advance reservation, co-allocation and pricing of network and computational resources in grids
Sampaio et al. Enhancing reliability of compute environments on Amazon EC2 spot instances

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18842159

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC , EPO FORM 1205A DATED 29.06.2020.

122 Ep: pct application non-entry in european phase

Ref document number: 18842159

Country of ref document: EP

Kind code of ref document: A1