CN109815009B

CN109815009B - Resource scheduling and optimizing method under CSP

Info

Publication number: CN109815009B
Application number: CN201811625775.2A
Authority: CN
Inventors: 张栋梁; 刘会会; 张中军; 叶海琴; 谭永杰; 李纲; 陈立勇; 王倩
Original assignee: Zhoukou Normal University
Current assignee: Zhoukou Normal University
Priority date: 2018-12-28
Filing date: 2018-12-28
Publication date: 2022-01-25
Anticipated expiration: 2038-12-28
Also published as: CN109815009A

Abstract

The invention discloses a resource scheduling and optimizing method under CSP (chip scale package), which belongs to the technical field of cloud computing, and comprises the following steps that firstly, sub-schedules of all application programs are generated in parallel and assembled into a pre-scheduling set; taking each pre-scheduling as a root in a pre-scheduling set for processing, executing GMaP in parallel, finally generating a plurality of scheduling results, and selecting the best scheduling result as a solution; GMAP mainly addresses energy issues but needs to meet deadlines; if all deadlines are met, GMaP only focuses on energy consumption unless a new deadline occurs; under the GMaP framework, scheduling policies can become too power or time consuming, but resource requests and runtime of the GMaP can be adjusted in two ways based on the power combination of the target cloud environment: the number of roots can be adjusted according to any natural number; the size of each search tree can be adjusted individually. The GMAP of the invention has flexible control on the size of a search space and the running time of an algorithm.

Description

Resource scheduling and optimizing method under CSP

Technical Field

The invention belongs to the technical field of cloud computing, and particularly relates to a resource scheduling and optimizing method under CSP.

Background

Cloud computing is considered as a next-generation computing technology model due to the advantages of on-demand self-service, ubiquitous network access, location-independent resource pools, risk transfer and the like. Cloud computing moves data computing and resource storage from a network to the "cloud" from where users can access the required resources anytime and anywhere. Meanwhile, a large data center and a server cluster-Cloud Service Provider (CSP) providing services for cloud users are also in the wake, which minimizes the operation cost and maximizes the operation efficiency of providing cloud services, and attracts more and more users to access the cloud service center.

The virtualization technology is one of important driving forces for cloud computing development, a server, a data center CPU, memory resources and the like are reconstructed on a basic unit virtualization platform (VM) for cloud computing deployment and management, a user calls the VM to construct a service and application platform through a CSP (chip scale service), and the CSP determines how to allocate physical resources to each user for cooperative use; as is well known, most of the operating costs of CSP are due to dynamic and static energy consumption, which is up to 50% of the peak power when large cloud servers are idle. To this end, selectively shutting down idle cloud servers or balancing resource utilization across all active servers may minimize energy consumption. To address the user's concern about potential adverse consequences, a Service Level Agreement (SLA) is used to normalize the quality of service agreed upon by both parties, including deadlines, privacy and security specifications, and the like. The CSP must unconditionally satisfy the user's personal requests while managing large-scale heterogeneous elastic cloud resources to maximize energy efficiency. To achieve this, the following three principles, namely the 3A criterion, must be followed:

1) accurate cloud platform modeling the cloud platform modeling only considers physical factors, is vital to model optimization, can reduce the computational complexity, and can keep enough model precision.

2) Appropriate user workload modeling currently, there are two workload model frameworks, "batch mode scheduling" and "dependent mode scheduling" in cloud computing. The key difference between them is the difference in the dependency of the processed tasks. Under the coarse-grained environment, each user workload is used as an atomic task and is independent of other tasks, whether a task set forms a batch processing task or not can be judged according to the Poisson arrival rate and the response time, and the complexity of CSP task scheduling is greatly reduced by the batch processing task. Under a fine-grained environment, the workload of each user is a more accurate task graph, tasks in the graph are mutually dependent and mutually supported, and CSP task scheduling is completed together.

3) Acceptable complexity for mode-dependent scheduling, the CSP ensures that the optimization process itself does not generate lengthy run times and high energy costs.

Disclosure of Invention

The invention aims to provide a resource scheduling and optimizing method under CSP. The method provides a universal scheduling and optimizing framework for the CSP, aims to improve the energy efficiency to the maximum extent and simultaneously meet the deadline of all users, and the framework is enough to bear the large-scale workload of multiple users processed under a large-scale cloud computing platform.

The technical scheme is as follows:

a resource scheduling and optimizing method under CSP (cloud service provider) starts a scheduling framework which is not limited by deadline and saves energy, and optimizes application programs violating the deadline and total energy consumption cost. First, the child schedules for each application are generated in parallel and assembled into a pre-schedule set. And taking each pre-scheduling as a root in the pre-scheduling set for processing, executing GMaP in parallel, finally generating a plurality of scheduling results, and selecting the best scheduling result as a solution. GMAP (Guided Migrate and Pack Guided migration and integration) mainly solves the energy problem but needs to meet the deadline. If all deadlines are met, GMaP only focuses on energy consumption unless a new deadline occurs. Under the GMaP framework, scheduling policies can become too power or time consuming, but resource requests and runtime of the GMaP can be adjusted in two ways based on the power combination of the target cloud environment:

1) the number of roots can be adjusted according to any natural number;

2) the size of each search tree can be adjusted individually.

Stage 1: generating child schedules

The sub-scheduling of user a is based on U^aG of (A)_aScheduling, assume that one VM is instantiated for each requested VM (virtual machine) type. While all virtual machines map to a single virtual server. Regardless of the scheduling algorithm, the sub-scheduling algorithms can be generated in parallel in all the application programs, and an improved greedy algorithm is adopted.

And (2) stage: application features

At this stage, each application has different feature parameters. The first characteristic parameter is G_aIs recorded as a sub-schedule length of

The second parameter is the deadline ambiguity parameter

The obfuscated application may better handle CSP energy costs.

Finally, assuming infinite VM resources, the PARa is subtracted from the timing length of the genetic algorithm, and three ordered lists of applications are generated in ascending order according to these parameters: l is_seed[↑]，PAR[↑]And SLACK [ stag [ °]。

And (3) stage: preliminary scheduling generation

The prescheduled set is generated by overlaying the child schedule onto the server.

And (4) stage: optimization

The core of GMaP is program optimization for directed search. This stage will migrate tasks from one virtual machine to another to meet deadlines and maximize energy efficiency. GMaP utilizes an evolutionary algorithm to achieve global optimality. Each iterative optimization goes through two steps, migration and integration.

Migrate () (migration function) functions to locate the source server D currently_xType g virtual machine on to another destination server D_yThe same type of virtual machine. The source server and the target server may be different, but in most cases they should be located in the same server farm in order to avoid high communication delays.

Further, three important decisions are made per migration attempt, which together determine the solution quality, namely:

1) which user migration should be selected;

2) which task migration should be selected;

3) to which server the task should be migrated.

The invention has the beneficial effects that:

the GMAP of the invention has flexible control on the size of a search space and the running time of an algorithm. Experimental results show that when GMaP is deployed for CSP, the global energy consumption is improved by more than 23% when the CSP provides services for 30-50 users, and is improved by more than 16% when the CSP provides services for 60-100 users.

Drawings

FIG. 1 is an environmental field view;

FIG. 2 is an application model;

FIG. 3 is a server farm distribution graph;

FIG. 4 is a graph of dynamic energy consumption as a function of CPU (Central processing Unit) utilization;

FIG. 5 is a virtual machine scheduling diagram;

FIG. 6 is a graph showing changes in COSP (total amount of energy consumption) values;

FIG. 7 is a diagram of the virtual machines assigned per user for experiment 6.

Detailed Description

The technical solutions of the present invention will be described in further detail with reference to the accompanying drawings and the detailed description.

The invention adopts the dependent mode scheduling of the workload, and provides a new cloud resource allocation mode, namely a task scheduling model and an energy consumption optimization framework, wherein the framework has the following characteristics:

(1) the workload is modeled as a set of multitask graphs with output dependencies. The cloud system optimization framework based on the workload model comprises: nephele, Pegasus, VGrADS.

(2) The cloud platform is modeled as a weighted graph representing heterogeneous servers with different resource capacities, energy efficiencies, and communication limits.

(3) The user requests a virtual machine (e.g., amazon EC2) in a pay-per-demand billing protocol, but without regard to resource allocation and task scheduling issues.

(4) The CSP integrally handles the problems of deadline, resource allocation, virtual machine arrangement, task scheduling, energy consumption cost optimization and the like.

(5) Under the optimized cloud resource condition, the scheduling algorithm completely realizes parallelization processing.

The task scheduling of fig. 1 is performed by the CSP selectively accepting workload requests through admission control policies, then allocating an appropriate number of virtual machines to each workload request, allocating the virtual machines to corresponding physical servers, and merging the virtual machines if necessary. On the premise of meeting the SLA deadline, all requests are processed, and energy consumption cost is reduced to the maximum extent. The workload model not only has the parallelization advantage of the task graph, but also has the global optimization characteristic; currently only applicable to batch workload frameworks, the algorithm is offline, but is easily converted to an adaptive online algorithm by recursive triggering when new users enter.

A related work

The first step in cloud service optimization for CSP is resource allocation, which is essentially the process of allocating computing resources in the form of physical servers and virtual machines. The simplest way is to consider the cloud platform as homogeneous servers that are disjoint from each other, while the workload is an independent request pattern for a given arrival rate. In this case, the resource allocation is solved by the following method: modifying a packing algorithm or an adaptive workload prediction algorithm inspired by queuing theory. A more accurate hardware modeling approach is related to communication capabilities. In the prior art, authors propose a graphics model based on multiple service cloud environments, and solve the server configuration and message routing problems by means of MILP (mixed integer linear programming), and introduce other variants, such as bundled virtual machine requests, virtual machine performance variability, multi-server sleep states, or price auctions.

The resource configuration is followed by the mapping of the application or virtual machine to the physical server. The goal of this process is to maintain a near-optimal utilization level for each server to achieve energy efficiency. This problem is similar to the classical load balancing problem in internet services, which can be solved simultaneously during resource allocation of independent workloads. For example, the binning algorithm in the prior art can not only reduce the number of deployed servers to the maximum extent, but also prevent the servers from being unloaded or overloaded, and the virtual machine configuration and layout problems are customized as a satisfaction function in the prior art. With other classical solutions, such as MCMF (minimum cost maximum flow), when workload is not predictable in advance, dynamic workload migration and virtual machine reallocation becomes very important at this point. The change in electricity prices in time and space keeps the load balanced. Related methods include MILP, original double, bargaining games and probability prediction.

In a cloud computing system, a graph of scheduling tasks for a dependency-based workload is different from batch task scheduling. At a higher level, this problem is similar to chip multiprocessing (CMP chip multiprocessor) scheduling for parallel/clustered computing communities. However, the technology developed by it is not directly applied to cloud users, mainly due to the opacity of public clouds; due to the coexistence of multiple competing users, more importantly, the elastic characteristics of the underlying hardware of the cloud computing are not suitable for the CSP, the improved scheduling framework is optimized from the perspective of individual users, for example, Nephele and Pegasus fully consider the dynamic characteristics of computing resources in the cloud computing. Due to lack of understanding of other competing users and the entire cloud resource map, they cannot capture global CSP management mechanisms such as admission control, virtual machine placement and consolidation, with the result that they are rated only in terms of resource utilization and deadlines for individual users, rather than overall energy consumption.

Second, user workload model

The invention uses directed acyclic graphs (DAG directed acyclic graphs) to model the user workload, and the whole workload is expressed as N disjoint DAG sets: { G₁(V₁,E₁),G₂(V₂,E₂),...,G_N(V_N,E_N) Each directed acyclic graph G_a(1. ltoreq. a. ltoreq.N) represents a workload request, G_aEach vertex V in_i ^a(1≤i≤|V_a|) represents a task. In general, it is assumed that each workload request is an application program belonging to a separate user. Thus, each application is equivalent to a user workload request. G_aFrom

To T_j ^aIs a line segment of (A) represents T_j ^aOutput dependence of

Weights on arrow line segments

Representing a previous task

Successfully delivered to T_j ^aThe amount of data required. An example is shown in figure 2.

Task features

Tasks are all processed by virtual machines, which are classified into K types: { VM₁,VM₂......,VM_K}, each virtual machine VM_g(1. ltoreq. g. ltoreq.K) is coupled to a two-dimensional integer set representing the amount of CPU and memory resources required by the virtual machine, i.e. the amount of CPU and memory resources required by the virtual machine

And

each task

Is coupled as a set of dyadic integers

First parameter in binary integer set

Representing tasks

Can only run on that virtual machine type. Second parameter

Is a task

In type

The longest execution time on the virtual machine, the set of binary integers is the data input part of the optimization algorithm. We consider the longest execution time as a common factor affecting all scheduling algorithms. The team of the Nephele project conceived a learning mechanism that enabled the cloud operator to summarize this execution time from past experience.

Resource request

In addition to workload requests, the user must also get the computing resources of the CSP. Of course, the CSP charges the user for the resource request according to a predetermined charging contract. In current cloud systems, resource requests are tied to virtual machine types. Each user can only specify the type of virtual machine required, but cannot specify the number of virtual machines of the required type, so the user does not have to be concerned with the details of resource allocation.

The virtual machine request is represented as: each application G_aIn relation to a binary array, the binary array U^aThere are K elements (K is the number of virtual machines):

if it is not

Representing a VM_gRequested by user a, otherwise

U^aIt must be ensured that all tasks are mapped onto the virtual machine, i.e. when

It must conform to the VM_gBelong to { VM₁,VM₂,......,VM_KAre multiplied by

Application per user is requested by workload G_aAnd virtual machine request U^aDefine, but the user does not perform scheduling. The reasons are three: i) the cloud platform is not transparent to the user, ii) the user does not have corresponding computing power; iii) CSPs have overall scheduling authority to achieve greater efficiency.

End period

Although a user cannot request more virtual machines of the same type, nor can he schedule his workload, workload performance can be controlled by specifying deadlines in the SLAs. User a (G)_a) The workload deadline of is noted

In general, when user a gives a smaller deadline, the CSP will be G_aAllocate more VM resources to G_aThe tasks in (1) can be executed in parallel and completed as soon as possible. However, the workload of the user is constrained by the admission control policy, and therefore applications that exceed the deadline will be denied scheduling or discarded during scheduling.

Third, cloud platform model

The cloud is composed of M servers: { D₁,D₂,……,D_MComposed of, and modeled as an undirected graph of M vertices, each representing a server, each edge's weight B_x,yRepresents (D)_x,D_y) The communication capacity between the servers, a plurality of adjacent servers constituting a locally connected server cluster, the server clusters communicating with each other via a high-speed channel, the distance between a server and the channel bandwidth being given by the value B_x,yAnd (4) showing. B is_x,xBy default, ∞, i.e. tasks executed on the same server do not incur any communication overhead. Furthermore, assuming that there is a high speed path between any two servers, whether through direct connection or multiple hops, the multi-hop path will be abstracted to a low value of B_x,yThe connecting edge of (2). As in fig. 3, the cloud platform consists of 9 servers and constitutes two server farms, one with 6 servers and the other with 3 servers, and the local connections may also be heterogeneous, for clarity, andnot all server connections are shown.

Virtual machine configuration and resource utilization

During operation, each server D_xAnd an array of integers having K elements

Therein are disclosed

Presentation Server D_xRun as type g VMs (VM)_g) Number of virtual machines, Q^xIs dynamic in that it changes due to the VM stopping and starting. Q^xValue of Q at time t^x(t) each server D_xInvolving a limited number of resources, i.e. CPU numbers respectively

And amount of memory

Obviously, D_xVM configuration of (2) is based on total amount of resources, i.e.

Satisfy the requirement of

And

server D_xPower consumption at time t includes static power consumption

And dynamic power consumption

Both of these factors are related to D_xResource utilization Util at time t_x(t) related, Util_x(t) consideration of Q alone^x(t) CPU utilization on a number of VMsRegardless of whether the virtual machine is running or idle, since during idle, the CPU also needs to be running. Util_x(t) is represented as follows:

when Util_x(t)>At the time of 0, the number of the first,

is a constant, otherwise it is 0. The present invention defines server D as having an optimal utilization per watt of performance as in the prior art_xHas an optimal utilization rate of Opt_x. Opt of the present Server_xAbout 0.7 when Util_x(t)<Opt_xWhen time comes, power consumption increases faster. Parameter(s)

And beta_xRespectively represent servers D_xIn Util_x(t)<Opt_xAnd Util_x(t)≥Opt_xThe power consumption increase ratio of the time is different from server to server even if the utilization rate is the same.

The calculation formula of (a) is as follows:

FIG. 4 depicts the performance of a CPU under different CPU utilization conditions

β_x10, and Opt_xWhen equal to 0.7

Graph of the variation of (c).

Assume that the maximum scheduled length upper bound for all applications is L_maxThe total amount of energy Consumption (COSP) is then the sum of the power consumption at all servers for the entire runtime:

admission control strategy

The purpose of access control is to judge and solve the problem of user workload that consumes resources excessively. In the cloud, some users request a large amount of VMs to be reserved, so that resources are occupied, and difficulties are brought to demands of other users and resource scheduling. In the invention, based on the deadline of a user, a secondary admission control strategy is adopted to screen and filter the user.

Under the secondary admission control policy, each application will check to see if it can complete before a given deadline. Unlike the core scheduler, this "schedulability" may be done in linear time, if the deadline is exceeded, the workload request will not be executed. During the scheduling process, users compete for VM resources under the supervision of CSP, and some user requests may not be satisfied due to resource limitations. If the deadline is still not met after a large amount of optimization work, the workload request is discarded.

Cloud operation

The CSP is responsible for configuring the VMs and assigning tasks, each virtual machine can only be occupied by one user until the CSP stops the virtual machine service. Suppose there is a ready task T_i ^a，

If the ready task T_i ^aIs dispatched to the server D_xIn the above, serving g-class virtual machines, two conditions must be satisfied:

1) the target VM is available and serves only user a;

2) all necessary data has been completed to be output.

Suppose task T_i ^aThe former task of (2) is

At server D_yOn execution, then the data output time is

The scheduling quality is determined by two factors: i) overall energy consumption; ii) the number of requests discarded due to violations of deadlines. Sometimes the scheduling quality is high, but eventually not feasible due to deadline violations, the scheduling policy should be adjusted to meet the deadline requirements and minimize the number of requests dropped. For example, let the CSP serve two users, the workload information is shown in Table I below.

TABLE 1 task graph and task delay

If the workload is treated as a separate atom, the figure shows that the scheduling method is to place the two applications on the most energy efficient server 5 (D)₅) When two type 1 virtual machines are allocated to it, the utilization ratio Util₅(t) is less than Opt₅. As a result, as shown in FIG. 5(a), the table length was 19 units. If it is not

(shown by the dashed red line), this schedule violates the deadlines of both users. Before the user request is discarded, the cost of calculating the schedule is:

to meet the deadlines of both users, the CSP exploits the data parallelism within G1 and G2. The greedy approach would produce the scheduling scheme generated by the schedule shown in FIG. 5(b), where all virtual machines are concentrated on D5, so both users have a reduced completion time, but D5 is overburdened. At the same time, Util_x(t) is greater than Opt_xThe scheduling energy consumption is as follows:

in detail, when t is 9, the type 1 virtual machine reserved for the user 1 is stopped, and the configuration of the type 1 virtual machine is reserved for the user 2. Drawing (A)

The scheduling energy consumption shown is:

another solution is to use other servers, such as server 6 (D6). The CSP adds one user to server D6 so that neither D5 nor D6 is loaded. Fig. 5 is the transfer of user 2 to the D6 server. D6 and D5 are not necessarily in the same server farm. The scheduling energy consumption is:

suppose that

β₅＝β₆＝10，Opt₅＝Opt₆＝0.7，

And

the value of (c) is variable. For the sake of simplicity, let

If only energy consumption is considered, cosp_aIs the best scheduling scheme. cosp_bAnd cosp_dAll have a value greater than cosp_aValue cosp_bAnd cosp_dIs dependent on P_staticAnd other parameters such as: beta is a₅/β₆. If the multi-user problem of the large task graph is considered, which can be realized in FIG. 5(c), then cosp_cIs most preferred, cosp_aBecause all deadlines are met, the energy consumption ratio cosp_aLow. How server energy efficiency affects scheduling is shown in fig. 6.

The main content of the research of the embodiment is as follows:

1) accelerating applications through additional virtual machine allocation and parallel execution typically increases CSP energy overhead.

2) There are various scheduling schemes when performing VM allocation and task migration, with scheduling schemes that do not violate deadlines and low energy costs being prioritized. The optimal allocation and migration scheme depends on the characteristics of the cloud platform and the workload situation.

Five, GMaP framework

In this section, the CSP's "guided migration and consolidation" (GMaP) scheduling optimization framework will be introduced. GMaP is based on directional search and is completely parallelized, and CSP runs GMaP by using distributable cloud resources.

The basic idea of the GMaP algorithm is to start an unlimited deadline and energy-saving scheduling framework to optimize application programs violating the deadline and the total energy consumption cost. First, the child schedules for each application are generated in parallel and assembled into a pre-schedule set. And taking each pre-scheduling as a root in the pre-scheduling set for processing, executing GMaP in parallel, finally generating a plurality of scheduling results, and selecting the best scheduling result as a solution. GMAP mainly addresses energy issues but is required to meet deadlines. If all deadlines are met, GMaP only focuses on energy consumption unless a new deadline occurs. Under the GMaP framework, scheduling policies can become too power or time consuming, but resource requests and runtime of the GMaP can be adjusted in two ways based on the power combination of the target cloud environment:

1) the number of roots can be adjusted according to any natural number;

2) the size of each search tree can be adjusted individually.

Stage 1: generating child schedules

The sub-scheduling of user a is based on U^aG of (A)_aScheduling, assume that one VM is instantiated for each requested VM type. While all virtual machines map to a single virtual server. Regardless of the scheduling algorithm, the sub-scheduling algorithms can be generated in parallel in all applications.

And (2) stage: application features

At this stage, each application has different feature parameters. The first characteristic parameter is the sub-scheduling length of Ga, noted as

The second parameter is the deadline ambiguity parameter

The obfuscated application may better handle CSP energy costs.

And (3) stage: preliminary scheduling generation

The prescheduled set is generated by overlaying the child schedule onto the server. As shown in fig. 5a, the pseudo code is as follows:

and (4) stage: optimization

The core of GMaP is program optimization for directed search. This stage will migrate tasks from one virtual machine to another to meet deadlines and maximize energy efficiency. GMaP utilizes an evolutionary algorithm to achieve global optimality. Each iterative optimization goes through two steps, migration and integration. The pseudo code is as follows:

migrate () has the function of placing the current location at the source serviceDevice D_xType g virtual machine on to another destination server D_yThe same type of virtual machine. The source server and the target server may be different, but in most cases they should be located in the same server farm in order to avoid high communication delays.

Three important decisions are made for each migration attempt, which together determine the solution quality, namely:

1) should it be selected which user to migrate?

2) Should it be selected which task to migrate?

3) To which server should the task migrate?

The CSP selects tasks that cause application delays and migrates them to servers that do not incur energy-efficient overhead, while considering whether they affect other applications. The decision process will typically cross-check L_seed[↑]，PAR[↑]And SLACK [ stag [ °]To select high L with negative SLACK value_seedAnd PAR value, because the application violates the deadline requirement, generates a negative s (slack) value, but because of the higher PAR value, the target server will be selected based on utilization level and task dependency by performing the reduction in length in parallel. However, since the current priority meets the deadline, the server on which the predecessor and successor tasks reside will be selected and communication across servers is minimized. When most applications meet the deadline, GMaP will have more flexibility to handle applications with high S values to minimize energy consumption. For example, migrating a task to a less congested server, after migration, the plan in the source server and the plan in the target server need to be adjusted, where the current task is a task that is directly or indirectly a successor to the migrated task. Migration may result in over-configuring VMs due to virtual machine integration and reorganization.

Sixth, experimental results

In this section, the effectiveness of GMaP is proved through a massive workload experiment on a large-scale cloud platform. The data input in each experiment is different, and the scale of the cloud platform is different. Table two provides the upper and lower limits for some key parameters.

TABLE 2 model parameter Table

The final scheduling result is first compared to the "best deadline for fuzzy prescheduling (BDOPS)", i.e., the best scheduling way is to treat the workload request as an atomic entity and ignore all deadlines. The BDOPS achieves optimal energy efficiency, but contains a large number of deadline violations. The invention establishes SLA to ensure that 30-80% of users in BDOPS violate deadline, and the BDOPS is a reference for energy consumption overhead calculation.

Second, the solution is compared to a baseline that subtracts the energy efficiency from the GMaP to obtain the optimization, i.e., the baseline scheduling is the fuzzy energy consumption. The results are shown in Table 3.

TABLE 3 Large and very Large user workload input results sheet

When 30-50 users are input in a large scale, the energy consumption is improved by-14.12% on average compared with BDOPS, in other words, the energy consumption overhead is 14.12% on average. This overhead is unavoidable because it is allocated by additional virtual machines for speeding up the application that violates the deadline. The CSP assigns an appropriate amount of VMs to each user based on the user's needs, and FIG. 7 illustrates the virtual machine assignment for experiment 6. The improvement in energy costs averaged 23.61% compared to baseline, a very promising result.

When entering 60-100 users at very large scale, the search algorithm space should be expanded to match the increase in the number of users entered if the same level of solution quality is achieved. However, to maintain consistency, the present invention modifies the size of the search space as shown in table 2. Thus, the energy cost increased by 49.72% on average, and the average energy consumption improvement relative to baseline decreased to 9.35%. From these data it was concluded that a search space of 50 roots and 5000 nodes per search tree was sufficient to accommodate 30-50 user workloads, but not sufficient for 60-100 user workloads.

By observing util (t) of all 10 servers in experiment 6. Ranking the servers according to energy efficiency level confirms that server 0 is the most energy efficient, from which the fundamental difference between BDOPS and scheduling schemes can be seen. For BDOPS, all applications are placed in the most energy efficient servers 0-4, while servers 5-9 are unused. While its energy efficiency ratio is optimal, of the 29 workloads accepted, 12 violated the deadline.

For the scheduling scheme, as GMaP allocates other virtual machines, so that the utilization rate of servers 0-4 increases, and the servers 5-9 with lower energy efficiency are online to host new virtual machines, the value of util (t) is higher when t is equal to 0, but the high values are reduced after being kept for a short time.

Search space analysis

Similar to other evolutionary algorithms, long-term GMaP operation may expand the algorithm search space and produce better solutions, but need to support larger search spaces by increasing the computation time, which will result in increased energy consumption and time to run the GMaP, the cost from deploying the GMaP to the GMaP itself, and evaluating the expanded algorithm search space for improvements in solution quality provide some solutions on how to balance the gains. First, Table 3 has an expanded search space, and runs experiments 15-18 anew, i.e., 50 roots in PSS and 10 for each search tree⁴The results are shown in Table 4.

Table 4 extended search space results table

When the search space is doubled, GMaP is better than the table 3 results, almost doubling the average energy consumption optimization from 9.35% in table 3 to 16.85% in table 4. Detailed case studies of 20 accepted users are examined below on a fixed cloud platform, 50 roots in PSS, exponentially increasing the search tree size from 100 to 105 nodes. The results are shown in Table 5.

TABLE 5 Effect of the Algorithm on scheduling optimization

As search trees grow, the GMaP solution may achieve the best. More importantly, GMaP has a reduced yield phenomenon, and the sizes of the search tree and the PSS are highly dependent on the actual operating environment. Currently GMaP can only achieve flexibility in search space.

The aggressiveness of the deadline in the SLA may affect the quality of the solution. The invention will perform a fixed cloud platform experimental plan, the platform comprising two server farms, 15 servers in total and a fixed set of 40 workload requests. The deadline for each application is a fraction of its seed plan length:

if μ ≧ 1, BDOPS will satisfy all deadlines and become the baseline plan immediately. When μ < 1, GMaP needs to address deadline violations. Since μ controls the deadline of all applications, a small drop in μ will greatly increase the aggressiveness of the overall deadline. If μ is too small, many requests will be dropped. To be consistent with the assumptions, Table 6 only studies the case where no requests are dropped.

TABLE 6 Effect of deadline on scheduling optimization

When μ ═ 1, GMAP can be seen to minimize energy consumption for the entire optimization process, an improvement of 20.32% over BDOPS. When μ drops below 1, GMAP is limited by the hard requirement of deadline, so the scheduling scheme is less advantageous than BDOPS. As the value of μ decreases, the baseline plan will perform many allocation of virtual machines and task migrations that are detrimental to energy consumption, and GMAP successfully recovers nearly 40% of the energy consumption penalty.

Seventh, conclusion

The invention considers the global operation optimization problem of cloud computing from the perspective of a Cloud Service Provider (CSP), aims to provide a universal scheduling and optimization framework for the CSP, aims to improve the energy efficiency to the maximum extent and simultaneously meet the deadline of all users, and the framework is enough to bear the large-scale workload of processing multiple users under a large-scale cloud computing platform.

The cloud computing system employs two types of workload models: independent batch requests and task graphs with dependencies. The present invention models workloads from multiple users as disjoint sets of task graphs. For the cloud platform model, the resource capacity and energy efficiency heterogeneity of the server can be completely reflected, and the communication bottleneck of the server is also taken into consideration. Fine-grained processing of hardware resources and user workload by parallel execution and global energy consumption minimization provides opportunities for deadline-oriented application acceleration but also requires more effort in admission control, resource allocation, virtual machine placement, and task scheduling. In the present invention we propose GMAP as a unified scheduling and optimization framework for CSP to solve these problems in a comprehensive way. GMAP is also flexible in controlling the size of the search space and the running time of the algorithm. Experimental results show that when GMaP is deployed for CSP, the global energy consumption is improved by more than 23% when the CSP provides services for 30-50 users, and is improved by more than 16% when the CSP provides services for 60-100 users.

The above description is only a preferred embodiment of the present invention, and the scope of the present invention is not limited thereto, and any simple modifications or equivalent substitutions of the technical solutions that can be obviously obtained by those skilled in the art within the technical scope of the present invention are within the scope of the present invention.

Claims

1. A CSP resource scheduling and optimization method is characterized in that firstly, the sub-scheduling of each application program is generated in parallel and assembled into a pre-scheduling set; taking each pre-scheduling as a root in a pre-scheduling set for processing, executing GMaP in parallel, finally generating a plurality of scheduling results, and selecting the best scheduling result as a solution; GMAP mainly addresses energy issues but needs to meet deadlines; if all deadlines are met, GMaP only focuses on energy consumption unless a new deadline occurs; under the GMaP framework, scheduling policies can become too power or time consuming, but resource requests and runtime of the GMaP can be adjusted in two ways based on the power combination of the target cloud environment:

1) the number of the roots can be adjusted according to any natural number;

2) the size of each search tree can be adjusted independently;

stage 1: generating child schedules

The sub-scheduling of user a is based on U^aG of (A)_aScheduling, assuming one VM is instantiated for each requested VM type; simultaneously mapping all the virtual machines to a single virtual server; regardless of which scheduling algorithm, the sub-scheduling algorithms can be generated in parallel in all the application programs;

and (2) stage: application features

In this stage, each application has different characteristic parameters; the first characteristic parameter is the sub-scheduling length of Ga, noted as

The second parameter is the deadline ambiguity parameter

The fuzzy application can better handle CSP energy costs;

finally, assuming infinite VM resources, the PARa is subtracted from the timing length of the genetic algorithm, and three ordered lists of applications are generated in ascending order according to these parameters: l is_seed[↑]，PAR[↑]And SLACK [ stag [ °]；

And (3) stage: preliminary scheduling generation

Generating a prescheduling set by overlaying the child schedule on a server;

and (4) stage: optimization

The core of GMaP is program optimization of directional search; this stage will migrate tasks from one virtual machine to another to meet deadlines and energy efficiency maximization; GMaP realizes global optimization by using an evolutionary algorithm; each iteration optimization is carried out through two steps, migration and integration;

the function of migrate () is to locate the current source server D_xType g virtual machine on to another destination server D_yThe same type of virtual machine; the source server and the target server are located in the same server farm.

2. The CSP resource scheduling and optimization method according to claim 1, wherein each migration attempt makes three important decisions, jointly determining the solution quality, namely:

1) which user migration should be selected;

2) which task migration should be selected;

3) to which server the task should be migrated.