CN109815009A - Scheduling of resource and optimization method under a kind of CSP - Google Patents
Scheduling of resource and optimization method under a kind of CSP Download PDFInfo
- Publication number
- CN109815009A CN109815009A CN201811625775.2A CN201811625775A CN109815009A CN 109815009 A CN109815009 A CN 109815009A CN 201811625775 A CN201811625775 A CN 201811625775A CN 109815009 A CN109815009 A CN 109815009A
- Authority
- CN
- China
- Prior art keywords
- scheduling
- gmap
- server
- resource
- csp
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000005457 optimization Methods 0.000 title claims abstract description 29
- 238000000034 method Methods 0.000 title claims abstract description 22
- 238000005265 energy consumption Methods 0.000 claims abstract description 34
- 238000012545 processing Methods 0.000 claims abstract description 13
- 239000012141 concentrate Substances 0.000 claims abstract description 5
- 101800000863 Galanin message-associated peptide Proteins 0.000 claims abstract 3
- 102100028501 Galanin peptides Human genes 0.000 claims abstract 3
- 238000013508 migration Methods 0.000 claims description 22
- 230000005012 migration Effects 0.000 claims description 22
- 230000010354 integration Effects 0.000 claims description 6
- 230000001174 ascending effect Effects 0.000 claims description 3
- 230000002068 genetic effect Effects 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 5
- 238000004891 communication Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 238000002474 experimental method Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 230000002829 reductive effect Effects 0.000 description 4
- 241001044157 Nephele Species 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 241001596784 Pegasus Species 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 238000012856 packing Methods 0.000 description 2
- 238000012913 prioritisation Methods 0.000 description 2
- 238000013468 resource allocation Methods 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 125000002015 acyclic group Chemical group 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000003780 insertion Methods 0.000 description 1
- 230000037431 insertion Effects 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000006798 recombination Effects 0.000 description 1
- 238000005215 recombination Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 239000011800 void material Substances 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Power Sources (AREA)
- Information Transfer Between Computers (AREA)
Abstract
The invention discloses the scheduling of resource and optimization method under a kind of CSP, belong to field of cloud computer technology, firstly, the son scheduling of each application program is parallel generation, and are assembled into pre-scheduling concentration;It concentrates in pre-scheduling using each pre-scheduling as root processing, executes GMaP parallel, finally generate multiple scheduling results, select best scheduling result as solution;GMAP mainly solves energy problem but need to meet the off period;If all off periods all meet, GMaP only focuses on energy consumption, unless there is the new off period;Under GMaP frame, scheduling strategy can become excessively power consumption or time-consuming, but the resource request of GMaP and runing time can carry out two aspect adjustment based on the power combination of target cloud environment: the quantity of root can be adjusted according to random natural number;The size of each search tree can individually adjust.GMAP of the invention is very flexible in the control of search space size and Riming time of algorithm.
Description
Technical field
The invention belongs to field of cloud computer technology, specifically, scheduling of resource and optimization method under being related to a kind of CSP.
Background technique
Cloud computing due to on-demand Self-Service, ubiquitous network insertion, the resource pool unrelated with position and
The advantages such as risk transfer are considered as next-generation computing technique mode.Cloud computing calculates data and resource is stored from network transitions
To " cloud ", required resource can be accessed in user whenever and wherever possible from cloud.Meanwhile the large data of service is provided for cloud user
Center and server cluster-cloud service provider (CSP) are also come into being, it will provide cloud service operation cost minimize and
Efficiency of operation maximizes, and attracts increasingly multiple access cloud service center.
Virtualization technology is one of the important impetus of cloud computing development, it provides server, data center CPU and memory
Source etc. is reconfigured on cloud computing deployment and the basic unit virtual platform (VM) of management, and user calls VM to construct by CSP
Service and application platform, how CSP decision distributes physical resource uses to each user collaborative;It is well known that the major part of CSP
Operation cost derived from dynamic and static energy consumption, the energy consumption that large-scale Cloud Server is waited during idle time is up to the 50% of peak power.
For this purpose, selectivity closing free time Cloud Server or in all active servers balance resource utilization can to enable consumption most
Smallization.In order to pacify user to the worry of potential adverse consequences, the both sides that standardized using Service Level Agreement (SLA) agreement
Service quality, including off period, privacy and safety standard etc..CSP will be managed while must unconditionally meeting individual subscriber request
Large scale scale heterogeneous elastic cloud resource is managed to improve efficiency to the maximum extent.To realize this target, it is necessary to follow following three
Principle, i.e. 3A criterion:
1) cloud platform accurate modeling cloud platform modeling only considers physical factor, it is most important to model optimization, can not only
Computation complexity is reduced, but also is able to maintain enough model accuracies.
2) appropriate user job load modeling is currently, there are two types of work load model frame, " batch processing moulds in cloud computing
Formula scheduling " and " relying on mode dispatching ".Key difference is that the dependence of handled task is different between them.In coarseness ring
Under border, each user job load is mutually indepedent as an atomic task and other tasks, according to poisson arrival rate and response
Time can judge whether set of tasks constitutes a batch processing task, and batch processing task greatly reduces CSP task schedule
Complexity.Under fine granularity environment, the workload of each user is a more accurate task image, in figure task mutually according to
Rely, mutually supports, it is common to complete CSP task schedule.
3) acceptable complexity ensures that optimization process itself will not generate too long fortune for relying on mode dispatching, CSP
Row time and higher energy consumption.
Summary of the invention
It is an object of the invention to propose scheduling of resource and optimization method under a kind of CSP.This method provides for CSP
General scheduling and Optimization Framework, it is intended to meet the off period of all users, the frame while improving efficiency to the maximum extent
It is enough to carry under large-scale cloud computing platform and handles the load of multi-user's major work.
Its technical solution is as follows:
Scheduling of resource and optimization method under a kind of CSP (cloud service provider cloud service provider) are opened
One unlimited off period and energy-efficient Scheduling Framework, to the application program and overall energy consumption cost optimization for violating the off period.Firstly,
The son scheduling of each application program is parallel generation, and is assembled into pre-scheduling concentration.It concentrates in pre-scheduling by each pre-scheduling
As root processing, GMaP is executed parallel, finally generates multiple scheduling results, selects best scheduling result as solution.
GMAP (Guided Migrate and Pack guidance migration and integration) mainly solves energy problem but need to meet the off period.Such as
Fruit all off periods all meet, then GMaP only focuses on energy consumption, unless there is the new off period.Under GMaP frame, plan is dispatched
Excessively power consumption or time-consuming can slightly be become, but the resource request of GMaP and runing time can be based on the power combinations of target cloud environment
Carry out two aspect adjustment:
1) quantity of root can be adjusted according to random natural number;
2) size of each search tree can individually adjust.
Stage 1: son scheduling is generated
The son scheduling of user a is based on UaGaScheduling, it is assumed that (virtual machine is virtual for the VM of each request
Machine) one VM of Type Concretization.Virtual machines all simultaneously are mapped to single virtual server.No matter which kind of dispatching algorithm,
Sub- dispatching algorithm can in all application programs parallel generation, use improved greedy algorithm.
Stage 2: feature is applied
In this stage, each application program has different characteristic parameters.First characteristic parameter is GaSub- scheduling length,
It is denoted asSecond parameter is off phase fuzzy parameterFuzzy application program can more preferably be located in
Manage CSP cost of energy.
Finally, it is assumed that VM resource is infinitely great, subtracts PARa with the timing length of genetic algorithm, presses ascending order according to these parameters
Generate the application list of three sequences: Lseed[↑], PAR [↑] and SLACK [↑].
Stage 3: preliminary scheduling generates
By covering sub- scheduling to generation pre-scheduling collection on server.
Stage 4: optimization
The core of GMaP is the program optimization of beam search.This stage can by task from a virtual machine (vm) migration to another
Virtual machine is maximized with meeting off period and efficiency.GMaP realizes global optimum using evolution algorithm.Each iteration optimization all passes through
Cross two steps, migration and integration.
The function of Migrate () (migration function) is will to be currently located at source server DxOn type g virtual machine (vm) migration arrive
Another destination server DyOn same type virtual machine.Source and target server can be different, but most of
In the case of, to avoid high communication delay, they should be located in the same server zone.
Further, migration is attempted all make three important decisions every time, codetermines solution quality, it may be assumed that
1) which user's migration should be selected;
2) which task immigration should be selected;
3) which server is the task should move to.
The invention has the benefit that
GMAP of the invention is very flexible in the control of search space size and Riming time of algorithm.The experimental results showed that when
When disposing GMaP for CSP, global energy consumption improves 23% or more when providing service for 30-50 user, for 60-100
A user improves 16% or more when providing service.
Detailed description of the invention
Fig. 1 is environment cyclogram;
Fig. 2 is Application models;
Fig. 3 is server zone distribution map;
Fig. 4 is dynamic energy consumption with CPU (central processing unit) utilization rate change curve;
Fig. 5 is scheduling virtual machine figure;
Fig. 6 is COSP (total amount of energy consumption) value variation diagram;
Fig. 7 is the virtual machine figure for testing 6 each user's distribution.
Specific embodiment
Technical solution of the present invention is described in more detail with reference to the accompanying drawings and detailed description.
The present invention uses the dependence mode dispatching of workload, and proposes a kind of new cloud resource configuration mode-task tune
Degree model and energy optimization frame, this frame have the characteristics that following:
(1) workload is modeled as the Multi-Task Graph set with output dependence.Based on this work load model
Cloud system Optimization Framework include: Nephele, Pegasus, VGrADS.
(2) cloud platform is modeled as representing and there is different resource capacity, efficiency and the heterogeneous server for communicating the limit
Weighted graph.
(3) user requests virtual machine (such as Amazon EC2) in the charging protocol of pay-for-use, but does not have to consider resource
Configuration and Mission Scheduling.
(4) it is excellent to handle deadline, resource distribution, arrangement virtual machine, task schedule and energy consumption cost in an integral manner by CSP
The problems such as change.
(5) under the conditions of the cloud resource of optimization, dispatching algorithm fully achieves parallelization processing.
Fig. 1 task schedule is executed by CSP, and CSP selectively receives workload request by Admission control, so
An appropriate number of virtual machine is distributed for each workload request afterwards, this virtual machine is distributed into corresponding physical server, such as
It is necessary to then merge virtual machine.Under the premise of meeting SLA deadline, all requests are handled, while most
Reduce energy consumption cost to limits.Above-mentioned work load model not only has task image parallelization advantage, but also there are also global excellent
Change feature;It is only applicable to batch processing workload frame at present, which is offline, but recurrence when entering by new user
Triggering, it is easy to which ground is translated into adaptive on-line Algorithm.
One, related work
For CSP, the first step of cloud service optimization is resource distribution, and substantially this is with physical server and virtual machine
Mode distribute the process of computing resource.Simplest mode is cloud platform to be envisioned for the server of mutually disjoint isomorphism,
And workload is the individual requests mode of given arrival rate.In this case, resource distribution is carried out by the following method
It solves: modification bin packing algorithm or the adaptive workload prediction algorithm inspired by queuing theory.More accurate hardware modeling
Method is related with communication capacity.In the prior art, author proposes a kind of graphical model based on a variety of service cloud environments, and
The settlement server by way of MILP (mixed integer linear programming mixed integer linear program)
Configuration and message routing issue also introduce other variants, for example binding virtual machine request, virtual machine performance changeability take more
The sleep state of business device or price auction.
It is the mapping of application program or virtual machine to physical server after resource distribution.The target of this process be each
Server is maintained close to optimum utilization level to realize high energy efficiency.This problem is similar to the classical load in Internet service
Equilibrium problem, it can be solved simultaneously during the resource allocation for the load that works independently.For example, bin packing algorithm in the prior art
The number of servers that deployment can not only be reduced to the maximum extent can also prevent server unloaded or overload, existing
Virtual machine configuration and Layout Problem are established as an extent function in technology.Using other classical solutions, such as
MCMF (minimum cost maximum stream flow), when workload is unpredictable in advance, then dynamic workload migration and virtual at this time
Machine reallocation just becomes extremely important.The variation of electricity price over time and space is so that load keeps balance.Correlation technique includes
MILP, original dual, bargaining game and probabilistic forecasting.
In cloud computing system, the scheduler task figure of the workload based on dependence is different from batch processing task tune
Degree.For higher level face, this problem is similar to chip multiprocessing (the CMP chip of parallel/PC cluster community
The multiprocessing of multiprocessor chip) scheduling.But the technology that it is developed does not directly apply to cloud user, mainly
Be as public cloud it is opaque caused by;It due to multiple competition users and deposits, it is often more important that the bullet of cloud computing bottom hardware
Property feature be not suitable for CSP, improved Scheduling Framework is optimized from the angle of personal user, for example Nephele and Pegasus fill
Divide the dynamic characteristic for considering computing resource in cloud computing.Due to lacking to other competition users and entire cloud resource figure
Solution, they can not capture global CSP administrative mechanism, and such as admission control, virtual machine arrangement and merging, as a result only basis is single
The resource utilization and deadline of user is evaluated, rather than whole energy consumption.
Two, user job load modules
The present invention is using directed acyclic graph (DAG directed acyclic graphs directed acyclic graph) to user job
Load modeling, entire workload are expressed as N number of disjoint DAG set: { G1(V1,E1),G2(V2,E2),...,GN(VN,
EN), each directed acyclic graph Ga(1≤a≤N) indicates a workload request, GaIn each vertex Vi a(1≤i≤|Va
|) indicate a task.General, it is assumed that each workload request is the application program for belonging to individual consumer.Therefore,
Each application program is equivalent to a user job load requests.GaIn fromTo Tj aLine segment indicate Tj aOutput rely on
Weight on arrow line segmentIt indicates from previous taskIt is successfully delivered to Tj aRequired data volume.Example such as Fig. 2 institute
Show.
Task feature
Task is all handled by virtual machine, and virtual machine is divided into K seed type: { VM1,VM2......,VMK, every kind of virtual machine VMg
(1≤g≤K) is coupled with a two-spot set of integers, the quantity of CPU and memory source needed for this virtual machine of the integer set representations,
AsWith
Each taskIt is coupled as a binary set of integersFirst parameter in binary group set of integers
Expression taskIt can only be run on the type of virtual machine.Second parameterIt is taskIn typeLongest is held on virtual machine
Row time, the binary group set of integers are the data input units of optimization algorithm.Maximum execution time is regarded as influence by we to be owned
The common factors of dispatching algorithm.The team of Nephele project contemplates a kind of study mechanism, its Neng Shi cloud operator is from past
This is summed up in experience executes the time.
Resource request
Other than workload request, user must also obtain the computing resource of CSP.Certainly, CSP is according to scheduled charge
Contract is charged for the resource request of user.In current cloud system, resource request is mutually bundled with type of virtual machine.Often
A user can only specify the type of required virtual machine, but cannot specify the quantity of the virtual machine of required type, so user
It need not be concerned about the detailed problem of resource allocation.
Virtual machine request is expressed as: each using GaThere is relationship with binary array, binary array UaThere is K element
Composition (K is virtual machine quantity):IfIndicate VMgIt is asked by user a
It asks, otherwiseUaIt must assure that all tasks are mapped on virtual machine, that is to say, that whenWhen, it has to comply with
VMgBelong to { VM1,VM2,......,VMK, andThe application of each user is by workload request GaU is requested with virtual machinea
Definition, but scheduling is not carried out in user.Wherein reason has three: I) cloud platform is opaque to user, II) user do not have accordingly
Computing capability;III) CSP has overall scheduling permission to realize higher efficiency.
Off period
Although user not can request that the virtual machine of more same types, the workload of oneself can not be arranged, can be led to
Cross the off period control workload performance in specified SLA.User a (Ga) the workload off period be denoted asOne
As, when user a provides a smaller off period, CSP will be GaMore VM resources are distributed, so as to GaIn task can
To execute and be finished as early as possible parallel.However, the workload of user is constrained by Admission control, therefore, beyond cut-off
The application program of phase will be rejected scheduling or abandon during scheduling.
Three, cloud platform models
Cloud is by M server: { D1,D2,……,DMComposition, and it is modeled as the non-directed graph on M vertex, each vertex generation
One server of table, the weight B on each sidex,yRepresent (Dx,Dy) between message capacity, multiple adjacent servers constitute local
The server zone of connection, the server zone are in communication with each other by high-speed channel, the distance between server and bandwidth chahnel value
Bx,yIt indicates.Bx,xIt is by default ∞, i.e., executing on the same server for task will not generate any communication overhead.This
Outside, it is assumed that all there is a high-speed channel between any two server, either by direct-connected or multi-hop, multihop path will be by
It is abstracted as low value Bx,yConnection side.In Fig. 3, cloud platform is made of 9 servers, and constitutes two server zones, wherein one
A server zone has 6 servers, and it is clear that another server zone, which has 3 servers, and locality connection is also likely to be isomery
For the sake of, do not show that Servers-all connects.
Virtual machine configuration and resource utilization
In operation, each server DxWith the integer array with K elementPhase
It closes, whereinIndicate server DxOn operate to type g VMs (VMg) virtual machine quantity, QxIt is dynamic, because it can be because
VM's stops working and starts and change.QxIt is Q in the value of time tx(t), each server DxMoney comprising limited quantity
Source, i.e. respectively CPU quantityAnd amount of memoryObviously, DxVM configuration according to total resources, i.e.,MeetAndServer DxIt include quiescent dissipation in the power consumption of time tAnd dynamic power consumptionThe two factors all with DxResource utilization Util on time txIt is (t) related,
Utilx(t) only consider Qx(t) cpu busy percentage on the VM of quantity, without considering that virtual machine is operation or free time, because in sky
During spare time, CPU is also required to operate.Utilx(t) it is expressed as follows:
Work as Utilx(t) > 0 when,It is a constant, is otherwise 0.According to every watt of performance in the prior art, clothes
Business device has an optimum utilization, and the present invention defines server DxOptimal utilization rate be Optx.The Opt of server at presentx≈
0.7, work as Utilx(t)<OptxWhen, power consumption increase is faster.ParameterAnd βxRespectively represent server DxIn Utilx(t)<OptxWith
Utilx(t) ≥OptxWhen power consumption increase ratio, even identical utilization rate, different server efficiencies be would also vary from.Calculation formula it is as follows:
Fig. 4 is depicted under different cpu busy percentages, whenβx=10, and OptxWhen=0.7's
Change curve.
Assuming that the maximum scheduling length upper bound of all applications is Lmax, then the total amount (COSP) of energy consumption is the whole service time
In the summation of the power consumption of Servers-all:
Admission control
The purpose of access control is the user job loading problem for judging and solve to consume excessively resource.In cloud, some
User requests a large amount of VMs are reserved to cause resource occupied, brings difficulty to the demand and scheduling of resource of other users.In the present invention
In, the off period based on user, we are using a kind of second level Admission control come screening and filtering user.
Under second level Admission control, it will check each application judge its whether can give off period it
Preceding completion.Different from kernel scheduling program, this " schedulability " can carry out in linear session, the off period if more than,
Then the workload request will not be performed.In scheduling process, user competes VM resource under the supervision of CSP, due to resource
It is limited and certain user's request may be unable to satisfy.If still not meeting the off period after a large amount of Optimization Works, then this work
Making load requests will be dropped.
The operation of four, clouds
CSP is responsible for configuring VM and distribution task, and every virtual machine can only be by a user occupancy, until CSP stops the void
Quasi- machine service.Assuming that there is a ready task Ti a,If ready task Ti aIt is scheduled to server DxOn, serve g
Class virtual machine then must satisfy two conditions:
1) target VM is available, and only serves user a;
2) all necessary datas of output are completed.
Assuming that task Ti aPrevious task beIn server DyUpper execution, then data output the time be
Scheduling quality is determined by two factors: i) overall energy consumption;Ii) the number of requests abandoned because violating the off period.
Sometimes scheduling quality is very high, but finally infeasible due to violating the off period, at this moment should adjust scheduling strategy and be wanted with meeting the off period
It asks, and reduces the number of requests being dropped to the greatest extent.For example, setting CSP serves two users, workload information see the table below I.
1 task image of table and task delay
If workload is regarded as individual atoms, display scheduling method is to be put into two application programs most in figure
Energy-efficient 5 (D of server5) on, when distributing two 1 type Virtual machines for it, utilization rate Util5(t) it is less than Opt5.Knot
Shown in fruit such as Fig. 5 (a), table length is 19 units.If(shown in dotted line red line), then this is adjusted
Degree violates the off period of two users.Before user's request is dropped, the cost of scheduling is calculated are as follows:
For the off period for meeting two users, CSP utilizes the data parallelism in G1 and G2.Greedy method can generation time
The scheduling scheme generated shown in table Fig. 5 (b), all virtual machines all concentrate on D5 in the figure, so when two users complete
Between reduced, but over-burden by D5.Meanwhile Utilx(t) it is greater than Optx, this energy consumption of scheduling are as follows:
From the point of view of in detail, in t=9, stop the Class1 virtual machine retained for user 1, is 2 Retention Type of user, 1 virtual machine
Configuration.FigureShown in energy consumption of scheduling are as follows:
Another solution is using other servers, such as server 6 (D6).CSP increases by one to server D6
User, so that D5 and D6 are not in load.Fig. 5 is that user 2 is transferred to D6 server.D6 and D5 is not necessarily in the same clothes
Be engaged in device group.This energy consumption of scheduling are as follows:
Assuming thatβ5=β6=10, Opt5=Opt6=0.7,With
Value be variable.For the sake of simplicity, ifIf only considering energy consumption, cospaIt is best dispatching party
Case. cospbAnd cospdValue be all larger than cospa, value cospbAnd cospdBetween comparison depend on PstaticAnd other parameters,
Such as: β5/β6.If it is considered that multi-user's problem of big task image, may be implemented in Fig. 5 (c), then cospcIt is best, cospa
Because meeting all off periods, energy consumption ratio cospaIt is low.It is as shown in Figure 6 how server efficiency influences scheduling.
Present case research is to the effect that:
1) CSP energy expense usually will increase come accelerating application by additional virtual machine distribution and parallel execution.
2) execute VM distribution and when task immigration there are a variety of scheduling schemes, without violate off period and low energy consumption at
This scheduling scheme is preferentially considered.Optimal allocation and migration scheme depend on the characteristic and workload situation of cloud platform.
Five .GMaP frames
In this section, " guidance migrates and integration " (GMaP) optimizing scheduling frame of CSP will be introduced.GMaP is based on beam search
And it is fully parallelized, CSP runs GMaP using governable cloud resource.
The basic thought of GMaP algorithm is to open a unlimited off period and energy-efficient Scheduling Framework, to the violation off period
Application program and overall energy consumption cost optimization.Firstly, the son scheduling of each application program is parallel generation, and it is assembled into presetting
Degree is concentrated.It concentrates in pre-scheduling using each pre-scheduling as root processing, executes GMaP parallel, finally generate multiple scheduling results,
Select best scheduling result as solution.GMAP mainly solves energy problem but need to meet the off period.If all
Off period all meets, then GMaP only focuses on energy consumption, unless there is the new off period.Under GMaP frame, scheduling strategy can become
Excessively power consumption or time-consuming, but the resource request of GMaP and runing time can carry out two based on the power combination of target cloud environment
Aspect adjustment:
1) quantity of root can be adjusted according to random natural number;
2) size of each search tree can individually adjust.
Stage 1: son scheduling is generated
The son scheduling of user a is based on UaGaScheduling, it is assumed that for one VM of VM Type Concretization of each request.Together
The virtual machine of Shi Suoyou is mapped to single virtual server.Regardless of which kind of dispatching algorithm, sub- dispatching algorithm can be all
Application program in parallel generation.
Stage 2: feature is applied
In this stage, each application program has different characteristic parameters.First characteristic parameter is the sub- scheduling length of Ga,
It is denoted asSecond parameter is off phase fuzzy parameterFuzzy application program can more preferably be located in
Manage CSP cost of energy.
Finally, it is assumed that VM resource is infinitely great, subtracts PARa with the timing length of genetic algorithm, presses ascending order according to these parameters
Generate the application list of three sequences: Lseed[↑], PAR [↑] and SLACK [↑].
Stage 3: preliminary scheduling generates
By covering sub- scheduling to generation pre-scheduling collection on server.As shown in Figure 5 a, pseudocode is as follows:
Stage 4: optimization
The core of GMaP is the program optimization of beam search.This stage can by task from a virtual machine (vm) migration to another
Virtual machine is maximized with meeting off period and efficiency.GMaP realizes global optimum using evolution algorithm.Each iteration optimization all passes through
Cross two steps, migration and integration.Pseudocode is as follows:
The function of Migrate () is will to be currently located at source server DxOn type g virtual machine (vm) migration to another purpose
Server DyOn same type virtual machine.Source and target server can be different, but in most cases, are
High communication delay is avoided, they should be located in the same server zone.
Migration is attempted all make three important decisions every time, codetermines solution quality, it may be assumed that
1) which user's migration should be selected?
2) which task immigration should be selected?
Which server should 3) task move to?
CSP selects the postponing application program of the task, is moved to the server that will not generate high energy efficiency expense
On, while to consider whether to have an impact other applications.Decision process would generally cross-check Lseed[↑], PAR [↑]
With SLACK [↑] to select that there is negative SLACK value, high LseedApplication with PAR value is migrated, because application program violates thus
Off period requirement produces negative S (slack) value, but since PAR value is higher, reduces length, mesh by executing parallel
Mark server will be selected according to utilization rate rank and task dependencies.But since current priority meets deadline,
Therefore the server that former subsequent tasks will be selected to be resident, and minimize the communication of cross-server.Journey is applied when most of
When sequence meets the off period, manipulation for greater flexibility is had the application program of high s-value to minimize energy consumption by GMaP.For example, by task
Less crowded server is moved to, after migration, needs to adjust the plan in source server and the plan in destination server,
Wherein current task is the direct or indirect subsequent tasks of migration task.Since migration will carry out virtual machine integration and recombination, because
This will lead to excessively configuration VM.
Six, experimental results
In this section, by magnanimity workload on large-scale cloud platform experiments have shown that the validity of GMaP.It tests every time defeated
The data entered are different, and cloud platform scale is not also identical.Table two provides the upper and lower bound of some key parameters.
2 model parameter table of table
Scheduling result more final first and " the best off period (BDOPS) of fuzzy pre-scheduling ", i.e., best dispatching party
Formula is exactly that workload request as an atomic entities and is ignored all off periods.BDOPS realizes best efficiency, but wraps
Containing a large amount of off period violation situations.The present invention formulates SLA and the user of 30%-80% in BDOPS is made to violate off period, BDOPS
It is the reference of energy consumption overhead computational.
Secondly, comparing solution and baseline, which subtracts the optimization that efficiency obtains from GMaP, i.e. benchmark scheduling is exactly
Fuzzy energy consumption.The results are shown in Table 3.
Table 3 is on a large scale and ultra-large user job loads input results table
As 30-50 user of input on a large scale, BDOPS is compared, energy consumption averagely improves -14.12%, in other words, energy
Consume expense average out to 14.12%.This expense is inevitable, because it is distributed by additional virtual machine, for accelerating the off period to disobey
The application program of rule.CSP distributes appropriate VM to each user according to user demand, and Fig. 7 illustrates the virtual machine point of experiment 6
Match.Compared with baseline, the improvement average out to 23.61% of energy cost, this is very promising result.
As 60-100 user of ultra-large input, if realizing the solution quality of same levels, search should be expanded and calculated
Method space inputs the increase of number of users to match.However be being consistent property, the present invention has modified the size of search space such as
Shown in table 2.Therefore, energy cost expense averagely increases 49.72%, and the average energy consumption improvement relative to baseline is reduced to
9.35%.According to these inferred from input data, 50 roots of each search tree and the search space of 5000 nodes are sufficient to accommodate 30-50
A user job load, but be inadequate for 60-100 user job load.
Pass through the Util (t) of 10 servers all in observation experiment 6.Server is ranked up according to efficiency grade,
Confirm server 0 be it is most energy-efficient, it can be seen that the fundamental difference between BDOPS and scheduling scheme.For BDOPS, own
Application program is all placed in most energy-efficient server 0-4, and server 5-9 is not used.Although its Energy Efficiency Ratio is best,
In 29 workloads received, 12 violate the off period.
For scheduling scheme, since GMaP is assigned with other virtual machines server 0-4 utilization rate is increased, make efficiency compared with
The new virtual machine of the low online trustship of server 5-9, Util (t) value is relatively high in t=0, but these high level are kept very in short-term
Between just have dropped.
Search space analysis
Similar to other evolution algorithms, GMaP long-play can preferably be solved with expansion algorithm search space and generation
Scheme, but need to support bigger search space by increasing the calculating time, this, which will lead to, increases energy consumption and runs GMAP's
Time, cost and assessment expansion algorithm search space changing in terms of solution quality from deployment GMaP to GMaP itself
Into providing some schemes on how to balance income.Firstly, table 3 has the search space of extension, experiment 15- is reruned
10 of 50 roots and each search tree in 18, i.e. PSS4A node, the results are shown in Table 4.
4 expanded search spatial result table of table
When search space is double, GMaP more preferably than 3 result of table, almost by average energy consumption optimized amount from table 3
9.35% doubles 16.85% into table 4.20 detailed cases for having received user are checked in fixed cloud platform below
Research, has 50 roots in PSS, search tree size is increased to 105 nodes from 100 node indexs.As the result is shown such as table 5
It is shown.
Influence of 5 algorithm of table to scheduling prioritization scheme
With the growth of search tree, GMaP solution be may be implemented most preferably.Importantly, there is income reduction in GMaP
The phenomenon that, the size of search tree and PSS are highly dependent on actual operating condition.GMaP can only be realized in terms of search space at present
Flexibility.
The aggressiveness of off period will affect the quality of solution in SLA.It is real that the present invention will carry out a fixed cloud platform
The plan of testing, the platform include two server zones, share 40 workload requests of 15 servers and one group of fixation.Each
The off period of application program is a part of its seed projected length:If μ >=1, BDOPS
All off periods will be met, and immediately become baseline programme.As μ < 1, GMaP needs to solve the problems, such as off period violation.Due to μ
The off period of all application programs is controlled, therefore the slightly lower general who has surrendered of μ greatly increases the aggressiveness of whole off period.If μ is too
It is small, then it can abandon many requests.In order to be consistent with hypothesis, 6, table researchs are without abandoning the case where requesting.
Influence of 6 off period of table to scheduling prioritization scheme
As μ=1, it can be seen that GMAP keeps the energy consumption of entire optimization process minimum, improves 20.32% than BDOPS.Work as μ
When being decreased to less than 1, GMAP is limited to this rigid requirement of off period, therefore scheduling scheme is superior not as good as BDOPS.With μ value
Reduce, baseline programme will execute many virtual machine distribution unfavorable to energy consumption and task immigration, GMAP have successfully restored nearly 40%
Energy loss.
Seven, conclusions
The global running optimizatin problem of present invention cloud computing from the viewpoint of cloud service provider (CSP), target are
General scheduling and Optimization Framework are provided for CSP, it is intended to meet cutting for all users while improving efficiency to the maximum extent
The only phase, which, which is enough to carry under large-scale cloud computing platform, handles the load of multi-user's major work.
Cloud computing system uses two kinds of work load model: dependence is requested and is had in independent batch processing
Task image.Workload from multiple users is modeled as disjoint task set of graphs by the present invention.For cloud platform mould
Type, it is fully able to reflection server resource capacity and efficiency isomerism, including server communication bottleneck is also considered.By simultaneously
Row executes and the global energy consumption minimized fine granularity processing loaded to hardware resource and user job is the application towards the off period
Program acceleration provides chance it also requires making more efforts in admission control, resource distribution, virtual machine layout and task schedule.
Unified scheduling and Optimization Framework in the present invention we have proposed GMAP as CSP, is solved these problems with general manner.
GMAP is also very flexible in the control of search space size and Riming time of algorithm.The experimental results showed that when disposing GMaP for CSP
When, global energy consumption improves 23% or more when providing service for 30-50 user, raising when providing service for 60-100 user
16% or more.
The foregoing is only a preferred embodiment of the present invention, the scope of protection of the present invention is not limited to this, it is any ripe
Know those skilled in the art within the technical scope of the present disclosure, the letter for the technical solution that can be become apparent to
Altered or equivalence replacement are fallen within the protection scope of the present invention.
Claims (2)
1. scheduling of resource and optimization method under a kind of CSP, which is characterized in that firstly, the son scheduling of each application program is simultaneously
What row generated, and it is assembled into pre-scheduling concentration;It concentrates in pre-scheduling using each pre-scheduling as root processing, executes GMaP parallel,
Multiple scheduling results are finally generated, select best scheduling result as solution;GMAP mainly solves energy problem but needs
Meet the off period;If all off periods all meet, GMaP only focuses on energy consumption, unless there is the new off period;In GMaP
Under frame, scheduling strategy can become excessively power consumption or time-consuming, but the resource request of GMaP and runing time can be based on target cloud ring
The power combination in border carries out two aspect adjustment:
1) quantity of root can be adjusted according to random natural number;
2) size of each search tree can individually adjust;
Stage 1: son scheduling is generated
The son scheduling of user a is based on UaGaScheduling, it is assumed that for one VM of VM Type Concretization of each request;Institute simultaneously
Some virtual machines are mapped to single virtual server;Regardless of which kind of dispatching algorithm, sub- dispatching algorithm can be in all applications
Parallel generation in program;
Stage 2: feature is applied
In this stage, each application program has different characteristic parameters;First characteristic parameter is the sub- scheduling length of Ga, is denoted asSecond parameter is off phase fuzzy parameterFuzzy application program can preferably handle CSP energy
Measure cost;
Finally, it is assumed that VM resource is infinitely great, subtracts PARa with the timing length of genetic algorithm, is generated according to these parameters by ascending order
The application list of three sequences: Lseed[↑], PAR [↑] and SLACK [↑];
Stage 3: preliminary scheduling generates
By covering sub- scheduling to generation pre-scheduling collection on server;
Stage 4: optimization
The core of GMaP is the program optimization of beam search;This stage can be virtual from a virtual machine (vm) migration to another by task
Machine is maximized with meeting off period and efficiency;GMaP realizes global optimum using evolution algorithm;Each iteration optimization all passes through two
Step, migration and integration;
The function of migrate () is will to be currently located at source server DxOn type g virtual machine (vm) migration to another purpose service
Device DyOn same type virtual machine;Source and target server is located in the same server zone.
2. scheduling of resource and optimization method under CSP according to claim 1, which is characterized in that migration is attempted all every time
Three important decisions can be made, solution quality is codetermined, it may be assumed that
1) which user's migration should be selected;
2) which task immigration should be selected;
3) which server is the task should move to.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811625775.2A CN109815009B (en) | 2018-12-28 | 2018-12-28 | Resource scheduling and optimizing method under CSP |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811625775.2A CN109815009B (en) | 2018-12-28 | 2018-12-28 | Resource scheduling and optimizing method under CSP |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109815009A true CN109815009A (en) | 2019-05-28 |
CN109815009B CN109815009B (en) | 2022-01-25 |
Family
ID=66602724
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811625775.2A Expired - Fee Related CN109815009B (en) | 2018-12-28 | 2018-12-28 | Resource scheduling and optimizing method under CSP |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109815009B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111124665A (en) * | 2019-11-22 | 2020-05-08 | 奇瑞汽车股份有限公司 | Method and device for distributing computing resources |
CN112363811A (en) * | 2020-11-16 | 2021-02-12 | 中国电子科技集团公司电子科学研究院 | Artificial intelligence computing resource scheduling method and computer readable storage medium |
CN114296868A (en) * | 2021-12-17 | 2022-04-08 | 中电信数智科技有限公司 | Virtual machine automatic migration decision method based on user experience in multi-cloud environment |
CN114827142A (en) * | 2022-04-11 | 2022-07-29 | 浙江大学 | Scheduling method for ensuring real-time performance of containerized edge service request |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1737752A (en) * | 2004-08-18 | 2006-02-22 | 华为技术有限公司 | Managing method for shared data |
US20110093500A1 (en) * | 2009-01-21 | 2011-04-21 | Google Inc. | Query Optimization |
CN102546771A (en) * | 2011-12-27 | 2012-07-04 | 西安博构电子信息科技有限公司 | Cloud mining network public opinion monitoring system based on characteristic model |
CN105447565A (en) * | 2015-11-19 | 2016-03-30 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | On-chip network mapping method based on discrete bat algorithm |
CN105740051A (en) * | 2016-01-27 | 2016-07-06 | 北京工业大学 | Cloud computing resource scheduling realization method based on improved genetic algorithm |
US20170116107A1 (en) * | 2011-05-31 | 2017-04-27 | International Business Machines Corporation | Testing a browser-based application |
CN107301137A (en) * | 2017-07-04 | 2017-10-27 | 福建中金在线信息科技有限公司 | RSET interface realizing methods and device and electronic equipment and computer-readable recording medium |
-
2018
- 2018-12-28 CN CN201811625775.2A patent/CN109815009B/en not_active Expired - Fee Related
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1737752A (en) * | 2004-08-18 | 2006-02-22 | 华为技术有限公司 | Managing method for shared data |
US20110093500A1 (en) * | 2009-01-21 | 2011-04-21 | Google Inc. | Query Optimization |
US20170116107A1 (en) * | 2011-05-31 | 2017-04-27 | International Business Machines Corporation | Testing a browser-based application |
CN102546771A (en) * | 2011-12-27 | 2012-07-04 | 西安博构电子信息科技有限公司 | Cloud mining network public opinion monitoring system based on characteristic model |
CN105447565A (en) * | 2015-11-19 | 2016-03-30 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | On-chip network mapping method based on discrete bat algorithm |
CN105740051A (en) * | 2016-01-27 | 2016-07-06 | 北京工业大学 | Cloud computing resource scheduling realization method based on improved genetic algorithm |
CN107301137A (en) * | 2017-07-04 | 2017-10-27 | 福建中金在线信息科技有限公司 | RSET interface realizing methods and device and electronic equipment and computer-readable recording medium |
Non-Patent Citations (3)
Title |
---|
G.DINESH等: "QoS Ranking Prediction for Cloud Brokerage Services", 《INTERNATIONAL JOURNAL OF COMPUTER SCIENCES AND ENGINEERING》 * |
张栋梁等: "云计算中负载均衡优化模型及算法研究", 《SOFTWARE》 * |
林伟伟等: "基于CSP 的能耗高效云计算资源调度模型与算法", 《通信学报》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111124665A (en) * | 2019-11-22 | 2020-05-08 | 奇瑞汽车股份有限公司 | Method and device for distributing computing resources |
CN111124665B (en) * | 2019-11-22 | 2023-07-28 | 奇瑞汽车股份有限公司 | Method and device for distributing computing resources |
CN112363811A (en) * | 2020-11-16 | 2021-02-12 | 中国电子科技集团公司电子科学研究院 | Artificial intelligence computing resource scheduling method and computer readable storage medium |
CN114296868A (en) * | 2021-12-17 | 2022-04-08 | 中电信数智科技有限公司 | Virtual machine automatic migration decision method based on user experience in multi-cloud environment |
CN114296868B (en) * | 2021-12-17 | 2022-10-04 | 中电信数智科技有限公司 | Virtual machine automatic migration decision method based on user experience in multi-cloud environment |
CN114827142A (en) * | 2022-04-11 | 2022-07-29 | 浙江大学 | Scheduling method for ensuring real-time performance of containerized edge service request |
CN114827142B (en) * | 2022-04-11 | 2023-02-28 | 浙江大学 | Scheduling method for ensuring real-time performance of containerized edge service request |
Also Published As
Publication number | Publication date |
---|---|
CN109815009B (en) | 2022-01-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Gao et al. | An energy and deadline aware resource provisioning, scheduling and optimization framework for cloud systems | |
Masdari et al. | A survey of PSO-based scheduling algorithms in cloud computing | |
Liu et al. | Job scheduling model for cloud computing based on multi-objective genetic algorithm | |
Zuo et al. | A multi-objective optimization scheduling method based on the ant colony algorithm in cloud computing | |
CN104657221B (en) | The more queue flood peak staggered regulation models and method of task based access control classification in a kind of cloud computing | |
Chaurasia et al. | Comprehensive survey on energy-aware server consolidation techniques in cloud computing | |
CN109815009A (en) | Scheduling of resource and optimization method under a kind of CSP | |
Asghari et al. | Online scheduling of dependent tasks of cloud’s workflows to enhance resource utilization and reduce the makespan using multiple reinforcement learning-based agents | |
Patel et al. | Priority based job scheduling techniques in cloud computing: a systematic review | |
Thaman et al. | Green cloud environment by using robust planning algorithm | |
Sonkar et al. | A review on resource allocation and VM scheduling techniques and a model for efficient resource management in cloud computing environment | |
Maiti et al. | Internet of Things applications placement to minimize latency in multi-tier fog computing framework | |
Xiang et al. | Computing power allocation and traffic scheduling for edge service provisioning | |
CN111309472A (en) | Online virtual resource allocation method based on virtual machine pre-deployment | |
Narwal et al. | A novel approach for Credit-Based Resource Aware Load Balancing algorithm (CB-RALB-SA) for scheduling jobs in cloud computing | |
CN106802822A (en) | A kind of cloud data center cognitive resources dispatching method based on moth algorithm | |
Ahmed et al. | An Enhanced Workflow Scheduling Algorithm for Cloud Computing Environment. | |
Ramezani et al. | Task scheduling in cloud environments: A survey of population‐based evolutionary algorithms | |
Rajeshwari et al. | Efficient task scheduling and fair load distribution among federated clouds | |
Aoun et al. | Towards a fairer benefit distribution in grid environments | |
Rahman et al. | Group based resource management and pricing model in cloud computing | |
Sharma et al. | Adaptive Particle Swarm Optimization for Energy Minimization in Cloud: A Success History Based Approach | |
Praveenchandar et al. | An enhanced load balancing approach for dynamic resource allocation in cloud environments | |
Panwar et al. | Analysis of various task scheduling algorithms in cloud environment | |
Xiao et al. | A novel QoS-based co-allocation model in computational grid |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20220125 |