CN115150277B - Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center - Google Patents
Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center Download PDFInfo
- Publication number
- CN115150277B CN115150277B CN202210665311.4A CN202210665311A CN115150277B CN 115150277 B CN115150277 B CN 115150277B CN 202210665311 A CN202210665311 A CN 202210665311A CN 115150277 B CN115150277 B CN 115150277B
- Authority
- CN
- China
- Prior art keywords
- real
- time
- data center
- tasks
- cloud data
- 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.)
- Active
Links
- 230000007246 mechanism Effects 0.000 title claims abstract description 45
- 238000000034 method Methods 0.000 claims abstract description 69
- 230000008859 change Effects 0.000 claims abstract description 7
- 238000013139 quantization Methods 0.000 claims abstract description 5
- 230000008569 process Effects 0.000 claims description 36
- 239000013598 vector Substances 0.000 claims description 33
- 230000006399 behavior Effects 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 239000011159 matrix material Substances 0.000 claims description 8
- 230000009977 dual effect Effects 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 6
- 230000002068 genetic effect Effects 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- 238000013178 mathematical model Methods 0.000 claims description 3
- 238000013508 migration Methods 0.000 claims description 3
- 230000005012 migration Effects 0.000 claims description 3
- 238000005265 energy consumption Methods 0.000 abstract description 19
- 238000004134 energy conservation Methods 0.000 description 3
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 2
- 229910052799 carbon Inorganic materials 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 230000000875 corresponding effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001934 delay Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006386 neutralization reaction Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0803—Configuration setting
- H04L41/0823—Configuration setting characterised by the purposes of a change of settings, e.g. optimising configuration for enhancing reliability
- H04L41/0833—Configuration setting characterised by the purposes of a change of settings, e.g. optimising configuration for enhancing reliability for reduction of network energy consumption
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0893—Assignment of logical groups to network elements
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/01—Protocols
- H04L67/10—Protocols in which an application is distributed across nodes in the network
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Power Sources (AREA)
Abstract
The application provides an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in a cloud data center, which comprises the following steps: dividing all isomorphic servers in a cloud data center into two clusters, namely a basic cluster and a standby cluster, setting a task classifier to classify and track tasks, constructing an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism, describing the working principle of the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism, evaluating the performance index of the working principle, analyzing the quantization behavior of the cloud data center under the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism, searching for the pareto optimal solution, and balancing the overall power consumption of the cloud data center with the average waiting time of real-time tasks. According to the method and the system, the partial servers are dynamically started and stopped according to the system load change, so that idle energy consumption can be effectively reduced, energy consumption expenditure of the cloud data center is greatly reduced, and cloud service providers are helped to select the best compromise solution among different targets according to the preference of the cloud service providers.
Description
Technical Field
The application belongs to the technical field of cloud computing, and particularly relates to an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in a cloud data center.
Background
In recent years, companies such as Huacheng, aliba, ten, google, amazon and the like invest huge amounts of funds to construct a super-large-scale cloud data center, and China communication institute 'data center white paper 2020' shows that the number of frames of the global data center reaches 750 ten thousand in 2019, and the number of frames of the global data center is increased by 4.3% in a same way. At the same time, the energy consumption of 2018 global data center is estimated to rise to 205TWh, accounting for about 1% of the global power consumption. As is well known, data centers are energy intensive enterprises, and the greatly increased and ever-increasing power costs of ultra-large scale cloud data centers result in more power expenditures, causing various energy conservation and efficiency problems. Achieving energy conservation and low carbon operation of cloud data centers is a key to continuously reducing Power Utilization Efficiency (PUE) and achieving the goal of "carbon neutralization".
To cope with the sudden traffic surge and server overload problems, the existing cloud data centers are built according to annual peak load prediction, however, due to over-allocation of resources, the average server utilization of the global data center is only 12% -18% according to statistics, and in addition, when the server is idle, the idle power consumption is only 50% -60% of the peak power. Thus, one key strategy to effectively reduce data center power consumption is to dynamically adjust the turning on and off of servers, which may be turned off temporarily by wake on lan (WoL) techniques, and to start up the servers when the cloud data center workload increases again substantially. It is clear that there is a significant tradeoff between powering down the idle server to save power consumption and keeping the idle server on to minimize the average latency of tasks.
The cloud computing is realized based on a server cluster technology, the server clusters are pooled through a software means, unified management and scheduling are carried out, and cloud services with the characteristics of elasticity, flexible charging on demand, high availability, high safety and the like are provided for the outside. A cluster is a collection of computing nodes organized together to serve, as if there were only one server, on a client side, and a cluster management system runs thousands of tasks on multiple clusters, which can have tens of thousands of nodes per cluster. Therefore, in order to achieve the energy consumption-performance tradeoff, designing a flexible cluster scheduling mechanism becomes a key place for the elastic expansion of the cloud data center. Different tasks are differentially served in cloud computing, and different service rates are provided for the different tasks, so that the tasks of multiple classes are reasonably scheduled, and a great amount of luxury energy consumption caused by unreasonable task scheduling is reduced. In order to track the whole process of the multi-class tasks being served in the cloud data center, a multi-class customer queuing system is utilized to describe a very close theoretical model of the random behavior of different classes of tasks in the cloud data center, and is very suitable for describing the problem that the multi-class customers compete for the same resources. In the existing research, the poisson process (PoissonArrival Process) is widely used for describing the arrival flows of different tasks, different customers in the poisson process are mutually independent and do not consider the correlation characteristic, however, the information flows in the real cloud data center are generally correlated, such as the common phenomena of sudden flow surge and server overload. For this reason, it is very important to extend the poisson process to a more general markov arrival process (Marked MarkovianArrival Process, MMAP) with its natural association with a multi-class customer queuing system with relevance, the first proposal of the markov arrival process being to characterize a random point process marking different classes of arrival. To date, the literature for modeling analysis of task random behavior, energy conservation, and performance tradeoff strategies in cloud data centers using multi-class customer queuing systems based on markov arrival processes is very sparse. Further, from the optimization approach of the energy saving strategy, some feasible solutions may be optimal from different angles for cloud users and cloud service providers when there are multiple evaluation criteria, while pareto frontier points typically select a set of potential optimal solutions for decision makers to choose the best solution according to their preferences. Therefore, the pareto optimal method is very suitable for solving the multi-objective optimization problem of energy consumption and performance trade-off in the cloud data center. In summary, for the above research background, it is urgent and necessary to find an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in a cloud data center to develop a construction method and a solution idea of an analytical model so that the analytical model is richer and closer to reality.
Disclosure of Invention
Aiming at the defects in the prior art, the application provides an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in a cloud data center. The method comprises the steps of dividing all isomorphic servers in a cloud data center into two clusters, namely a basic cluster and a standby cluster, setting a task classifier to classify and track tasks, constructing an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism, describing the working principle of the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism, evaluating performance indexes below the working principle, analyzing quantization behaviors of the cloud data center under the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism, searching for a pareto optimal solution, and balancing the overall power consumption of the cloud data center with the average waiting time of real-time tasks. According to the method and the system, the partial servers are dynamically started and stopped according to the system load change, so that idle energy consumption can be effectively reduced, energy consumption expenditure of the cloud data center is greatly reduced, and cloud service providers are helped to select the best compromise solution among different targets according to the preference of the cloud service providers.
The application provides an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in a cloud data center, which comprises the following steps:
s1, all isomorphic servers N in one cloud data center total Dividing into basic clusters N 1 And standby cluster N 2 Two clusters, wherein the servers in the basic cluster are always in operation, while the servers in the standby cluster are dynamically turned on and off along with the load change and are controlled by a hysteresis mechanism of double thresholds, wherein the double thresholds comprise an on threshold T 1 And a shutdown threshold T 2 ；
S2, setting a task classifier to classify and track the tasks: setting a task classifier to divide tasks reaching a cloud data center into real-time tasks and non-real-time tasks, marking the tasks with different labels, and tracking the service process of each type of tasks in the system by using the labels;
s3, constructing an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism:
s31, when a task arrives at the cloud data center, as long as an idle server exists, the newly arrived task immediately occupies one of the idle servers; if all servers are busy and are real-time tasks, executing step S32, and if all servers are busy and are non-real-time tasks, executing step S33;
s32, if the real-time task queued in the buffer area does not reach the access threshold H, the newly arrived real-time task is added into the buffer area, otherwise, the real-time task leaves the cloud data center and is timely scheduled to other cloud data center nodes;
s33, when one non-real-time task arrives at the cloud data center, the arriving non-real-time task enters a buffer area to wait in a queuing manner without any limitation even if all servers are busy;
s34, when the service of one task is completed, if at least one real-time task exists in the buffer zone, once the server is idle, the first waiting real-time task is served; if the buffer area has only non-real-time tasks, the first waiting non-real-time task is served; for the same class of tasks, the service rule is first come first served;
s4, describing the working principle of the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism;
s5, evaluating performance indexes under the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism;
s6, analyzing quantization behaviors of the cloud data center under the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism: and searching the pareto optimal solution, namely the pareto front point, from the angles of cloud users and cloud service providers respectively, so that the overall power consumption of the cloud data center and the average waiting time of real-time tasks are balanced.
Further, the step S4 specifically includes the following steps:
s41, describing the arrival process of the multi-class task as a Markov arrival process (MMAP) with marks, and respectively obeying the parameters of service time of the real-time task and service time of the non-real-time task as mu 1 and μ2 Is used for constructing an MMAP [ K ]]/M[K]/N 1 +N 2 A non-preemptive priority queue;
s42, constructing a five-dimensional continuous time Markov chain by considering random behaviors of multiple tasks in the cloud data center:
{(i t ,w t ,r t ,j t ,v t ),t≥0} (1)
wherein ,it Indicating the number of tasks in the system at time t, and i t 0, called system waterLeveling; w (w) t Representing the state of the standby cluster at time t, anIf w t If 0 indicates that the standby cluster is off, if w t =1 then indicates that the standby cluster is active; r is (r) t Indicating the real-time task number in the buffer at time t, and +.>j t Representing the number of real-time tasks being serviced at time t, and +.>The arrival of tasks is performed by a random process { v t Guided by t.gtoreq.0 } the random process { v } t T.gtoreq.0 } is an irreducible continuous time Markov chain, the underlying markov chain, called the MMAP arrival process, has a state space of {0, 1..m }, v t Represents the underlying Markov chain { v } at time t t State of t.gtoreq.0 }, and +.>
S43, adopting a pseudo-extinction process, a matrix geometric solution method and a Gaussian-Saidel method to obtain steady-state distribution n of the system:
Π＝(π 0 ,π 1 ,π 2 ,...) (2)
wherein ,πi A probability vector representing a system level i, and satisfies
π i ＝(π(i,0),π(i,1)) (3)
Wherein pi (i, 0), pi (i, 1) respectively represent a probability vector when the standby cluster is not turned on and a probability vector when the standby cluster is turned on, and the following are satisfied:
π(i,0)＝(π(i,0,0),π(i,0,1),...,π(i,0,max{0,min{i-N 1 ,H}})) (4)
π(i,1)＝(π(i,1,0),π(i,1,1),...,π(i,1,max{0,min{i-N 1 -N 2 ,H}})) (5)
wherein the method comprises the steps ofPi (i, 0, r), pi (i, 1, r) respectively represent a probability vector of a previous three-dimensional when the standby cluster is not turned on and a probability vector of a previous three-dimensional when the standby cluster is turned on, and r= {0, 1.. t -N 1 H }, satisfy:
π(i,0,r)＝(π(i,0,r,0),π(i,0,r,1),...,π(i,0,r,min{i,N 1 })) (6)
π(i,1,r)＝(π(i,1,r,0),π(i,1,r,1),...,π(i,1,r,min{i,N 1 +N 2 })) (7)
where pi (i, w, r, j) represents the probability vector of the first four dimensions, and w= {0,1}, j= {0,1,.. 1 +N 2 -meeting the following:
π(i,w,r,j)＝(π(i,w,r,j,0),π(i,w,r,j,1),...,π(i,w,r,j,m)) (8)。
further, the step S5 specifically includes the following steps:
s51, from the perspective of cloud users, solving QoS indexes related to service quality, wherein the QoS indexes related to the service quality comprise average waiting time of real-time tasks, average waiting time of non-real-time tasks and loss rate of real-time tasks; average waiting time W of the real-time task and the non-real-time task real ,W non-real The method comprises the following steps of:
wherein ,λ1 Representing the arrival rate of the real-time task; lambda (lambda) 2 Representing the arrival rate of the non-real-time task;
loss rate of the real-time taskThe method comprises the following steps:
wherein ,representation (N) 1 +1)×(N 1 A unit vector of +1); d (D) 1 Representing a real-time task arrival matrix; />Representation (N) 1 +1) (m+1) ×1 column vectors with all elements 1; />Representation (N) 1 +N 2 +1)×(N 1 +N 2 A unit vector of +1);representation (N) 1 +N 2 +1) (m+1) ×1 column vectors with all elements 1; />Represents the kronecker product;
s52, from the perspective of cloud service providers, solving TCO indexes related to overall utilization, wherein the TCO indexes related to overall utilization comprise the activation rate, the system utilization rate and the overall system power consumption of the standby cluster, and the activation rate P of the standby cluster act The method comprises the following steps:
wherein ,ei Column vectors representing 1 for all elements of ix1;
the system utilization U s The method comprises the following steps:
the overall power consumption E of the system is as follows:
wherein ,representing the execution power consumption of the server when processing the real-time task; />Representing the execution power consumption of the server when processing non-real-time tasks; e (E) idle Representing idle power consumption of the server; e (E) off Representing the power consumption of the server to shut down.
Preferably, the step S6 specifically includes the following steps:
s61, defining the overall power consumption of the cloud data center and the average waiting time of the real-time tasks as key parameter variables N 1 、N 2 、T 1 、T 2 The functions of H, i.e. E (N) 1 ,N 2 ,T 1 ,T 2 ,H)、W real (N 1 ,N 2 ,T 1 ,T 2 H) represented by a mathematical model of multi-objective optimization:
s62, obtaining the pareto optimal solution, namely the pareto leading edge point, by utilizing a non-dominant ordering genetic algorithm with an improved elite strategy, so as to provide a group of optimal solutions.
Preferably, in step S1, if the number of tasks in the buffer exceeds the threshold T 1 The servers in the standby cluster will all be activated and remain operational until the total number of tasks in the cloud data center falls to the shutdown threshold T 2 And migrating the task being serviced in the standby cluster to the basic cluster, and shutting down the standby cluster after all migration.
Preferably, said step S1 takes into account ensuring that once ready for useThe cluster is activated and all tasks in the buffer can be serviced immediately, setting an on threshold T 1 <N 2 The method comprises the steps of carrying out a first treatment on the surface of the Setting a shutdown threshold T in consideration of timely shutdown of the standby cluster when the workload of the cloud data center is low 2 ＝N 1 -1。
Preferably, in the step S2, a higher non-preemptive priority is given to the real-time task than to the non-real-time task, and an access threshold H (H<T 1 )。
Compared with the prior art, the application has the technical effects that:
1. aiming at the problem of rapid increase of energy consumption caused by rapid increase of industrial scale of the global data center, the application designs an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in the cloud data center, and considers that partial servers are dynamically started and stopped according to system load change of the servers of the cloud data center so as to effectively reduce idle energy consumption, thereby greatly reducing energy consumption expenditure of the cloud data center; in order to balance the service quality of multi-class tasks and the energy consumption level of a cloud data center, an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism is provided, and a non-preemptive priority queuing model based on a markov arrival process with a mark is constructed aiming at the energy-saving strategy.
2. According to the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism in the cloud data center, qoS indexes such as average waiting time of real-time tasks and non-real-time tasks, loss rate of real-time tasks and the like, and TCO indexes such as activation probability of standby clusters, system utilization rate and overall system power consumption are obtained by adopting a simulated vanishing process, a matrix geometric solution method and a Gaussian-seidel method, and the average waiting time of real-time tasks and the overall power consumption of the cloud data center are weighed, the pareto front edge point is obtained by utilizing an improved non-dominant ordering genetic algorithm with elite strategy, and a set of optimal solutions are provided for cloud service providers, so that the cloud service providers are helped to select a solution representing optimal compromise among different targets according to preferences of the cloud service providers.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the detailed description of non-limiting embodiments made with reference to the following drawings.
FIG. 1 is a flow chart of a power saving strategy based on a dual threshold hysteresis cluster scheduling mechanism in a cloud data center of the present application;
FIG. 2 is a schematic scheduling diagram of a power saving strategy based on a dual threshold hysteresis cluster scheduling mechanism in a cloud data center of the present application;
fig. 3 is a pareto front point of the average latency of the real-time task versus the overall power consumption of the system of the present application.
Detailed Description
The application is described in further detail below with reference to the drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the application and are not limiting of the application. It should be noted that, for convenience of description, only the portions related to the present application are shown in the drawings.
It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other. The application will be described in detail below with reference to the drawings in connection with embodiments.
Aiming at the problems of energy waste such as large idle energy consumption caused by idle computing nodes and large luxury energy consumption caused by unreasonable task scheduling in the operation process of a cloud data center, an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism is provided.
Fig. 1 and 2 illustrate a power saving strategy based on a dual threshold hysteresis cluster scheduling mechanism in a cloud data center of the present application, the method comprising the steps of:
s1, all isomorphic servers N in one cloud data center total Dividing into basic clusters N 1 And standby cluster N 2 Two clusters, wherein the servers in the basic cluster are always in operation state, while the servers in the standby cluster are dynamically turned on and off along with the load change and are controlled by a hysteresis mechanism of double thresholds, wherein the double thresholds comprise an on threshold T 1 And a shutdown threshold T 2 The method comprises the steps of carrying out a first treatment on the surface of the If any of the buffersThe number of transactions exceeds the turn-on threshold T 1 The servers in the standby cluster will all be activated and remain operational until the total number of tasks in the cloud data center falls to the shutdown threshold T 2 And migrating the task being serviced in the standby cluster to the basic cluster, and shutting down the standby cluster after all migration.
The turn-on threshold T is set to ensure that all tasks in the buffer can be serviced immediately once the standby cluster is activated 1 <N 2 The method comprises the steps of carrying out a first treatment on the surface of the Setting a shutdown threshold T in consideration of timely shutdown of the standby cluster when the workload of the cloud data center is low 2 ＝N 1 -1。
S2, setting a task classifier to classify and track the tasks: setting a task classifier to divide tasks reaching a cloud data center into real-time tasks and non-real-time tasks, marking the tasks with different labels, and tracking the service process of each type of tasks in the system by using the labels; giving a higher non-preemptive priority to a real-time task than to a non-real-time task while setting an access threshold H (H<T 1 )。
Real-time tasks are very sensitive to delay and delay jitter, and are typically performed in some applications where the expiration date is critical, such as autopilot, avionics, robotics, etc. Real-time tasks require quick service completion, while non-real-time tasks in certain application scenarios can tolerate longer delays, such as email systems, wild animal tracking networks, deep space communication networks, etc., requiring differentiated services to guarantee their service level agreements (Service Level Agreement, SLAs) based on the characteristics of each class of task. By differentially serving different classes of tasks, the two classes of tasks are scheduled reasonably at different service rates, improving their quality of service, and reducing luxury energy consumption.
S3, constructing an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism:
s31, when a task arrives at the cloud data center, as long as an idle server exists, the newly arrived task immediately occupies one of the idle servers; if all servers are busy and are real-time tasks, step S32 is performed, and if all servers are busy and are non-real-time tasks, step S33 is performed.
S32, if the real-time tasks queued in the buffer area do not reach the access threshold H, the newly arrived real-time tasks are added into the buffer area, otherwise, the real-time tasks leave the cloud data center and are timely scheduled to other cloud data center nodes.
And S33, when one non-real-time task arrives at the cloud data center, the arriving non-real-time task enters a buffer zone to wait in a queuing manner without any limitation even if all servers are busy.
S34, when the service of one task is completed, if at least one real-time task exists in the buffer zone, once the server is idle, the first waiting real-time task is served; if the buffer area has only non-real-time tasks, the first waiting non-real-time task is served; for the same class of tasks, the service rules are first come first served.
S4, describing the working principle of an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism.
S41, describing the arrival process of the multi-class task as a Markov arrival process (MMAP) with marks, and respectively obeying the parameters of service time of the real-time task and service time of the non-real-time task as mu 1 and μ2 Is used for constructing an MMAP [ K ]]/M[K]/N 1 +N 2 Non-preemptive priority queues.
S42, constructing a five-dimensional continuous time Markov chain by considering random behaviors of multiple tasks in the cloud data center:
{(i t ,w t ,r t ,j t ,v t ),t≥0} (1)
wherein ,it Indicating the number of tasks in the system at time t, and i t 0, called system level; w (w) t Representing the state of the standby cluster at time t, anIf w t If 0 indicates that the standby cluster is off, if w t =1 then indicates that the standby cluster is active; r is (r) t Representing real-time tasks in a buffer at time tCount and->j t Representing the number of real-time tasks being serviced at time t, and +.>The arrival of tasks is performed by a random process { v t Guided by t.gtoreq.0 } the random process { v } t T.gtoreq.0 } is an irreducible continuous time Markov chain, the underlying markov chain, called the MMAP arrival process, has a state space of {0, 1..m }, v t Represents the underlying Markov chain { v } at time t t State of t.gtoreq.0 }, and +.>
S43, adopting a pseudo-extinction process, a matrix geometric solution method and a Gaussian-Saidel method to obtain steady-state distribution n of the system:
Π＝(π 0 ,π 1 ,π 2 ,...) (2)
wherein ,πi A probability vector representing a system level i, and satisfies
π i ＝(π(i,0),π(i,1)) (3)
Wherein pi (i, 0), pi (i, 1) respectively represents a probability vector when the standby cluster is not turned on and a probability vector when the standby cluster is turned on, and the following conditions are satisfied:
π(i,0)＝(π(i,0,0),π(i,0,1),...,π(i,0,max{0,min{i-N 1 ,H}})) (4)
π(i,1)＝(π(i,1,0),π(i,1,1),...,π(i,1,max{0,min{i-N 1 -N 2 ,H}})) (5)
wherein pi (i, 0, r), pi (i, 1, r) respectively represent a probability vector of a previous three-dimensional when the standby cluster is not turned on and a probability vector of a previous three-dimensional when the standby cluster is turned on, and r= {0, 1.. t -N 1 H }, satisfy:
π(i,0,r)＝(π(i,0,r,0),π(i,0,r,1),...,π(i,0,r,min{i,N 1 })) (6)
π(i,1,r)＝(π(i,1,r,0),π(i,1,r,1),...,π(i,1,r,min{i,N 1 +N 2 })) (7)
where pi (i, w, r, j) represents the probability vector of the first four dimensions, and w= {0,1}, j= {0,1,.. 1 +N 2 -meeting the following:
π(i,w,r,j)＝(π(i,w,r,j,0),π(i,w,r,j,1),...,π(i,w,r,j,m)) (8)。
s5, evaluating performance indexes under an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism.
S51, from the perspective of cloud users, solving QoS indexes related to service quality, wherein the QoS indexes related to the service quality comprise average waiting time of real-time tasks, average waiting time of non-real-time tasks and loss rate of the real-time tasks; average waiting time W of real-time task and non-real-time task real ,W non-real The method comprises the following steps of:
wherein ,λ1 Representing the arrival rate of the real-time task; lambda (lambda) 2 Representing the arrival rate of non-real-time tasks.
Loss rate of real-time tasksThe method comprises the following steps:
wherein ,representation (N) 1 +1)×(N 1 A unit vector of +1); d (D) 1 Representing a real-time task arrival matrix; />Representation (N) 1 +1) (m+1) ×1 column vectors with all elements 1; />Representation (N) 1 +N 2 +1)×(N 1 +N 2 A unit vector of +1); />Representation (N) 1 +N 2 +1) (m+1) ×1 column vectors with all elements 1; />Representing the kronecker product.
S52, from the perspective of cloud service providers, solving TCO indexes related to overall utilization, wherein the TCO indexes related to overall utilization comprise the activation rate of the standby cluster, the utilization rate of the system and the overall power consumption of the system, and the activation rate P of the standby cluster act The method comprises the following steps:
wherein ,ei Representing a column vector with 1 for all elements of i x 1.
System utilization U s The method comprises the following steps:
the overall power consumption E of the system is as follows:
wherein ,representing processing real-time tasksExecuting power consumption of the time server; />Representing the execution power consumption of the server when processing non-real-time tasks; e (E) idle Representing idle power consumption of the server; e (E) off Representing the power consumption of the server to shut down.
S6, analyzing quantization behaviors of the cloud data center under an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism: and searching the pareto optimal solution, namely the pareto front point, from the angles of cloud users and cloud service providers respectively, so that the overall power consumption of the cloud data center and the average waiting time of real-time tasks are balanced.
S61, defining the overall power consumption of the cloud data center and the average waiting time of the real-time tasks as key parameter variables N 1 、N 2 、T 1 、T 2 The functions of H, i.e. E (N) 1 ,N 2 ,T 1 ,T 2 ,H)、W real (N 1 ,N 2 ,T 1 ,T 2 H) represented by a mathematical model of multi-objective optimization:
s62, obtaining the pareto optimal solution, namely the pareto leading edge point, by utilizing a non-dominant ordering genetic algorithm with an improved elite strategy, so as to provide a group of optimal solutions. In a specific embodiment, the pareto front point of the overall power consumption of the system-average waiting time of the real-time task shown in fig. 3 is obtained, and from the perspective of cloud users, the schemes can improve the service quality of the real-time task and simultaneously save energy consumption to the greatest extent; from the perspective of the cloud service provider, the cloud service provider can choose the desired energy savings effect according to a guaranteed energy savings agreement (Guaranteed Energy SavingsAgreement, GESA) and infer whether the corresponding real-time task latency meets the requirements of its service level agreement.
Aiming at the problem of rapid increase of energy consumption caused by rapid increase of industrial scale of the global data center, the application designs an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in the cloud data center, and considers that partial servers are dynamically started and stopped according to system load change of the servers of the cloud data center so as to effectively reduce idle energy consumption, thereby greatly reducing energy consumption expenditure of the cloud data center; in order to balance the service quality of multi-class tasks and the energy consumption level of a cloud data center, an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism is provided, and a non-preemptive priority queuing model based on a Markov arrival process with a mark is constructed aiming at the energy-saving strategy; qoS indexes such as average waiting time of real-time tasks and non-real-time tasks in a steady state, loss rate of real-time tasks and the like, TCO indexes such as activation probability of standby clusters, system utilization rate, overall system power consumption and the like are obtained by adopting a pseudo-living process, a matrix geometric solution method and a Gaussian-Saider method, and the average waiting time of the real-time tasks and the overall power consumption of a cloud data center are balanced, and the pareto front edge point is obtained by utilizing an improved non-dominant ordering genetic algorithm with elite strategy, so that a set of optimal solutions are provided for cloud service providers, and the cloud service providers are helped to select solutions representing optimal trade-offs among different targets according to preferences of the cloud service providers.
Finally, what should be said is: the above embodiments are merely for illustrating the technical aspects of the present application, and it should be understood by those skilled in the art that although the present application has been described in detail with reference to the above embodiments: modifications and equivalents may be made thereto without departing from the spirit and scope of the application, which is intended to be encompassed by the claims.
Claims (7)
1. An energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism in a cloud data center is characterized by comprising the following steps of:
s1, all isomorphic servers N in one cloud data center total Dividing into basic clusters N 1 And standby cluster N 2 Two clusters, wherein servers in the base cluster are always in operation, and servers in the standby cluster are in serviceThe server is dynamically turned on and off along with the load change and is controlled by a hysteresis mechanism of double thresholds, wherein the double thresholds comprise an on threshold T 1 And a shutdown threshold T 2 ；
S2, setting a task classifier to classify and track the tasks: setting a task classifier to divide tasks reaching a cloud data center into real-time tasks and non-real-time tasks, marking the tasks with different labels, and tracking the service process of each type of tasks in the system by using the labels;
s3, constructing an energy-saving strategy based on a dual-threshold hysteresis cluster scheduling mechanism:
s31, when a task arrives at the cloud data center, as long as an idle server exists, the newly arrived task immediately occupies one of the idle servers; if all servers are busy and are real-time tasks, executing step S32, and if all servers are busy and are non-real-time tasks, executing step S33;
s32, if the real-time task queued in the buffer area does not reach the access threshold H, the newly arrived real-time task is added into the buffer area, otherwise, the real-time task leaves the cloud data center and is timely scheduled to other cloud data center nodes;
s33, when one non-real-time task arrives at the cloud data center, the arriving non-real-time task enters a buffer area to wait in a queuing manner without any limitation even if all servers are busy;
s34, when the service of one task is completed, if at least one real-time task exists in the buffer zone, once the server is idle, the first waiting real-time task is served; if the buffer area has only non-real-time tasks, the first waiting non-real-time task is served; for the same class of tasks, the service rule is first come first served;
s4, describing the working principle of the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism;
s5, evaluating performance indexes under the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism;
s6, analyzing quantization behaviors of the cloud data center under the energy-saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism: and searching the pareto optimal solution, namely the pareto front point, from the angles of cloud users and cloud service providers respectively, so that the overall power consumption of the cloud data center and the average waiting time of real-time tasks are balanced.
2. The energy saving strategy based on the dual threshold hysteresis cluster scheduling mechanism in the cloud data center according to claim 1, wherein the step S4 specifically comprises the steps of:
s41, describing the arrival process of the multi-class task as a Markov arrival process (MMAP) with marks, and respectively obeying the parameters of service time of the real-time task and service time of the non-real-time task as mu 1 and μ2 Is used for constructing an MMAP [ K ]]/M[K]/N 1 +N 2 A non-preemptive priority queue;
s42, constructing a five-dimensional continuous time Markov chain by considering random behaviors of multiple tasks in the cloud data center:
{(i t ,w t ,r t ,j t ,v t ),t≥0} (1)
wherein ,it Indicating the number of tasks in the system at time t, and i t 0, called system level; w (w) t Representing the state of the standby cluster at time t, anIf w t If 0 indicates that the standby cluster is off, if w t =1 then indicates that the standby cluster is active; r is (r) t Indicating the real-time task number in the buffer at time t, and +.>j t Representing the number of real-time tasks being serviced at time t, and +.>The arrival of tasks is performed by a random process { v t Guided by t.gtoreq.0 } the random process { v } t T.gtoreq.0 } is an irreducible continuous time Markov chain, called MMAP-toThe state space of the bottom markov chain of the reach process is {0,1,..m }, v t Represents the underlying Markov chain { v } at time t t State of t.gtoreq.0 }, and
s43, adopting a pseudo-extinction process, a matrix geometric solution method and a Gaussian-Saidel method to obtain steady-state distribution n of the system:
Π＝(π 0 ,π 1 ,π 2 ,...) (2)
wherein ,πi A probability vector representing a system level i, and satisfies
π i ＝(π(i,0),π(i,1)) (3)
Wherein pi (i, 0), pi (i, 1) respectively represent a probability vector when the standby cluster is not turned on and a probability vector when the standby cluster is turned on, and the following are satisfied:
π(i,0)＝(π(i,0,0),π(i,0,1),...,π(i,0,max{0,min{i-N 1 ,H}})) (4)
π(i,1)＝(π(i,1,0),π(i,1,1),...,π(i,1,max{0,min{i-N 1 -N 2 ,H}})) (5)
wherein pi (i, 0, r), pi (i, 1, r) respectively represent a probability vector of a previous three-dimensional when the standby cluster is not turned on and a probability vector of a previous three-dimensional when the standby cluster is turned on, and r= {0, 1.. t -N 1 H }, satisfy:
π(i,0,r)＝(π(i,0,r,0),π(i,0,r,1),...,π(i,0,r,min{i,N 1 })) (6)
π(i,1,r)＝(π(i,1,r,0),π(i,1,r,1),...,π(i,1,r,min{i,N 1 +N 2 })) (7)
where pi (i, w, r, j) represents the probability vector of the first four dimensions, and w= {0,1}, j= {0,1,.. 1 +N 2 -meeting the following:
π(i,w,r,j)＝(π(i,w,r,j,0),π(i,w,r,j,1),...,π(i,w,r,j,m)) (8)。
3. the energy saving strategy based on the dual threshold hysteresis cluster scheduling mechanism in the cloud data center according to claim 1, wherein the step S5 specifically comprises the steps of:
s51, from the perspective of cloud users, solving QoS indexes related to service quality, wherein the QoS indexes related to the service quality comprise average waiting time of real-time tasks, average waiting time of non-real-time tasks and loss rate of real-time tasks; average waiting time W of the real-time task and the non-real-time task real ,W non-real The method comprises the following steps of:
wherein ,λ1 Representing the arrival rate of the real-time task; lambda (lambda) 2 Representing the arrival rate of the non-real-time task;
loss rate of the real-time taskThe method comprises the following steps:
wherein ,representation (N) 1 +1)×(N 1 A unit vector of +1); d (D) 1 Representing a real-time task arrival matrix; />Representation (N) 1 +1) (m+1) ×1 column vectors with all elements 1; />Representation (N) 1 +N 2 +1)×(N 1 +N 2 A unit vector of +1);representation (N) 1 +N 2 +1) (m+1) ×1 column vectors with all elements 1; />Represents the kronecker product;
s52, from the perspective of cloud service providers, solving TCO indexes related to overall utilization, wherein the TCO indexes related to overall utilization comprise the activation rate, the system utilization rate and the overall system power consumption of the standby cluster, and the activation rate P of the standby cluster act The method comprises the following steps:
wherein ,ei Column vectors representing 1 for all elements of ix1;
the system utilization U s The method comprises the following steps:
the overall power consumption E of the system is as follows:
wherein ,representing the execution power consumption of the server when processing the real-time task; />Representing the execution power consumption of the server when processing non-real-time tasks; e (E) idle Representing idle power consumption of the server; e (E) off Representing the power consumption of the server to shut down.
4. The energy saving strategy based on the dual threshold hysteresis cluster scheduling mechanism in the cloud data center according to claim 1, wherein the step S6 specifically comprises the steps of:
s61, defining the overall power consumption of the cloud data center and the average waiting time of the real-time tasks as key parameter variables N 1 、N 2 、T 1 、T 2 The functions of H, i.e. E (N) 1 ,N 2 ,T 1 ,T 2 ,H)、W real (N 1 ,N 2 ,T 1 ,T 2 H) represented by a mathematical model of multi-objective optimization:
s62, obtaining the pareto optimal solution, namely the pareto leading edge point, by utilizing a non-dominant ordering genetic algorithm with an improved elite strategy, so as to provide a group of optimal solutions.
5. The energy saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism in the cloud data center according to claim 1, wherein in the step S1, if the number of tasks in the buffer exceeds the turn-on threshold T 1 The servers in the standby cluster will all be activated and remain operational until the total number of tasks in the cloud data center falls to the shutdown threshold T 2 And migrating the task being serviced in the standby cluster to the basic cluster, and shutting down the standby cluster after all migration.
6. The energy saving strategy based on a dual threshold hysteresis cluster scheduling mechanism in a cloud data center according to claim 1, wherein the step S1 is considered to ensure that all tasks in the buffer are guaranteed once the standby cluster is activatedCan be immediately serviced, set up a turn-on threshold T 1 <N 2 The method comprises the steps of carrying out a first treatment on the surface of the Setting a shutdown threshold T in consideration of timely shutdown of the standby cluster when the workload of the cloud data center is low 2 ＝N 1 -1。
7. The energy saving strategy based on the dual-threshold hysteresis cluster scheduling mechanism in the cloud data center according to claim 1, wherein the step S2 gives the real-time tasks a higher non-preemptive priority than the non-real-time tasks, while setting the access threshold H (H<T 1 )。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210665311.4A CN115150277B (en) | 2022-06-13 | 2022-06-13 | Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210665311.4A CN115150277B (en) | 2022-06-13 | 2022-06-13 | Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115150277A CN115150277A (en) | 2022-10-04 |
| CN115150277B true CN115150277B (en) | 2023-09-15 |
Family
ID=83407736
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210665311.4A Active CN115150277B (en) | 2022-06-13 | 2022-06-13 | Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115150277B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108829230A (en) * | 2018-06-26 | 2018-11-16 | 西南交通大学 | The design method of data center's energy saving model based on queueing theory |
| EP3488342A1 (en) * | 2016-07-20 | 2019-05-29 | Worldline | Multi-criteria adaptive scheduling for a market-oriented hybrid cloud infrastructure |
| CN112926785A (en) * | 2021-03-10 | 2021-06-08 | 国网吉林省电力有限公司信息通信公司 | Cloud data center load optimization scheduling method supporting electric power big data |
| CN113515351A (en) * | 2021-09-07 | 2021-10-19 | 华南理工大学 | Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization |
-
2022
- 2022-06-13 CN CN202210665311.4A patent/CN115150277B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3488342A1 (en) * | 2016-07-20 | 2019-05-29 | Worldline | Multi-criteria adaptive scheduling for a market-oriented hybrid cloud infrastructure |
| CN108829230A (en) * | 2018-06-26 | 2018-11-16 | 西南交通大学 | The design method of data center's energy saving model based on queueing theory |
| CN112926785A (en) * | 2021-03-10 | 2021-06-08 | 国网吉林省电力有限公司信息通信公司 | Cloud data center load optimization scheduling method supporting electric power big data |
| CN113515351A (en) * | 2021-09-07 | 2021-10-19 | 华南理工大学 | Resource scheduling implementation method based on energy consumption and QoS (quality of service) cooperative optimization |
Non-Patent Citations (1)
| Title |
|---|
| 基于M/M/c排队模型的云计算中心能耗管理策略;张书豪;张延华;王倩雯;王朱伟;李萌;;计算机测量与控制(第08期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115150277A (en) | 2022-10-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2023184939A1 (en) | Deep-reinforcement-learning-based adaptive efficient resource allocation method for cloud data center | |
| Xiao et al. | Multi-objective parallel task offloading and content caching in D2D-aided MEC networks | |
| CN112835698A (en) | Heterogeneous cluster-based dynamic load balancing method for request classification processing | |
| Hussin et al. | Priority-based scheduling for large-scale distribute systems with energy awareness | |
| Li et al. | Cost-aware automatic scaling and workload-aware replica management for edge-cloud environment | |
| Wang et al. | Energy-efficient collaborative optimization for VM scheduling in cloud computing | |
| Cheng et al. | Proscale: Proactive autoscaling for microservice with time-varying workload at the edge | |
| Feng et al. | Towards heat-recirculation-aware virtual machine placement in data centers | |
| Li et al. | Dynamic virtual machine consolidation algorithm based on balancing energy consumption and quality of service | |
| Bai et al. | Performance analysis of an energy-saving strategy in cloud data centers based on a MMAP [K]/M [K]/N1+ N2 non-preemptive priority queue | |
| Magotra et al. | Adaptive computational solutions to energy efficiency in cloud computing environment using VM consolidation | |
| Liu et al. | An energy-saving task scheduling model via greedy strategy under cloud environment | |
| Hussin et al. | Efficient energy management using adaptive reinforcement learning-based scheduling in large-scale distributed systems | |
| Wang et al. | Task arrival based energy efficient optimization in smart-IoT data center | |
| Tian et al. | Modeling and analyzing power management policies in server farms using stochastic petri nets | |
| CN115150277B (en) | Energy-saving strategy based on dual-threshold hysteresis cluster scheduling mechanism in cloud data center | |
| Daoud et al. | Cloud-IoT resource management based on artificial intelligence for energy reduction | |
| Liu et al. | Energy‐aware virtual machine consolidation based on evolutionary game theory | |
| CN110308991B (en) | Data center energy-saving optimization method and system based on random tasks | |
| Yuan et al. | A DRL-Based Container Placement Scheme with Auxiliary Tasks. | |
| CN109783304A (en) | A kind of energy-saving scheduling method and corresponding intrument of data center | |
| Dewi et al. | Toward Task Scheduling Approaches to Reduce Energy Consumption in Cloud Computing Environment | |
| Lv et al. | HVMM: A Holistic Virtual Machine Management Strategy for Cloud Data Centers | |
| CN112422651A (en) | Cloud resource scheduling performance bottleneck prediction method based on reinforcement learning | |
| Yang et al. | Cross-layer self-similar coflow scheduling for machine learning clusters |
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 |