CN102135914B - Cloud computing system load predicting method capable of automatically adjusting parameters - Google Patents

Abstract

The invention discloses a cloud computing system load predicting method capable of automatically adjusting parameters, which comprises the following steps: at the moment t, computing the actual load O(t) of a system at the moment t through system call; executing short-term prediction; computing alpha (t) and E(t) by utilizing the O(t) value and historical data; executing long-term prediction; computing alpha T(t) and ET(t) by utilizing the O(t) and the historical data; combining the short-term prediction and the long-term prediction; when t is less than T, outputting the O(t) and switching to the next step; otherwise, taking the maximum value or average value of E(t-1) and ET(t-T) as the output at the moment t; and updating the historical data, waiting the moment t+1 and switching to the first step. In the invention, the alpha (t) and the alpha T(t) are computed in real time through error functions, thereby enhancing the prediction accuracy of classic EWMA (Exponentially Weighted Moving Average); the requirement that a prediction value is slightly larger than an actual value can be met by expanding the alpha (t) and the alpha T(t) to an interval (-1, 1); and the responsiveness of the prediction to the load periodicity of a cloud computing platform is enhanced by introducing a long-term prediction module.

Description

A kind of cloud computing system load predicting method that can regulate parameter automatically

Technical field

The invention provides a kind of system load Forecasting Methodology, be specifically related to the load estimation of physical machine in a kind of cloud computing platform or virtual machine, belong to computer system and network field.

Background technology

In order to improve the resource utilization ratio of whole cloud computing platform, need dispatch the virtual machine that moves on each physical machine.But,, can only obtain existing load state constantly through checking the method for physical machine and virtual machine information.Therefore, should set up suitable model, prognoses system is at loading demand constantly in future, thereby realizes effective and reasonable scheduling.

In practical application, load estimation has two targets: the error of i. predicted value and actual value is as far as possible little; Ii. predicted value is greater than actual value.The latter is from following idea: for each virtual machine, would rather overabsorption system resource, and because of causing application, the system resource deficiency can't normally carry out avoiding.And in the model of reality, often need make balance at these 2.

Exponentially Weighted Moving Average (hereinafter to be referred as the EWMA) method that Roberts proposed in nineteen fifty-nine is applied to many computer realms (ROBERTS, S.W. (1959) .Control chart tests based on geometric moving average.Technometrics 1239-250) at present.This is a Short-term Forecasting Model, and its form is following:

E(t)＝αE(t-1)+(1-α)O(t)

Wherein E (t) is illustrated in the predicted value of t constantly, and O (t) is illustrated in the actual observation value of t constantly, and α is a fixing weight, value between 0 to 1.

The problem that the EWMA method exists comprises: when the load sequence is in rising trend, predicted value will can't satisfy target ii all the time between historical data and actual value; The value of α is fixed, and needs to set through experience, and can not adapt to the load characteristic of different phase; EWMA can only carry out short-term forecasting to load, can't portray the periodic property of load.

Summary of the invention

For ease of explanation, this paper agreement: t representes that T representes a duty cycle constantly; O (t) expression is the system load actual value of t constantly, is provided by system at each moment t; E (t) is illustrated in the short-term forecasting value of moment t to moment t+1 system load; E _T(t) be illustrated in the long-term forecasting value of moment t to moment t+T system load; α (t) expression is used to ask the parameter of E (t); α _T(t) expression is used to ask E _T(t) parameter; E (0), E (1) ... ..E (t-1), O (0), O (1) ... ..O (t-1) is referred to as historical data.

The objective of the invention is: a kind of parameter real-time update mechanism is provided, descends in order to solve the precision of prediction that preset parameter causes among the EWMA; Through parameter range being expanded on [1,1] interval, to reach the purpose that predicted value is slightly larger than actual value; Method is extended to long-term forecasting, the periodicity of reflected load; Long-term forecasting is combined with short-term forecasting.

Principle of the present invention is: at each moment t, obtain the actual loading O of system (t) of current time earlier through system call; Then when carrying out short-term forecasting, utilize O (t) value and historical data calculation of alpha (t) and E (t); When carrying out long-term forecasting, utilize O (t) and historical data calculation of alpha _T(t) and E _T(t); Get E (t-1) and E at last _T(t-T) higher value in is as the output valve of moment t.

Technical scheme provided by the invention is following:

Scheme 1: a kind of cloud computing system load predicting method that can regulate parameter automatically, it is characterized in that, comprise the steps (flow process is referring to Fig. 1):

A. at moment t, calculate the actual loading O of system (t) of t constantly through system call;

B. carry out short-term forecasting: utilize O (t) value and historical data calculation of alpha (t) and E (t);

C. carry out long-term forecasting: utilize O (t) and historical data calculation of alpha _T(t) and E _T(t);

D. comprehensive short-term forecasting and long-term forecasting: when t＜T, output O (t) changes step e; Otherwise, get E (t-1), E _T(t-T) maximal value of the two or mean value are as the output of moment t;

E. upgrade historical data, wait for t+1 constantly, change steps A.

Scheme 2: a kind of preferred version for scheme 1 is characterized in that the system call in the steps A is: call the application programming interfaces that virtual machine manager provides, perhaps virutal machine memory page or leaf, the perhaps self-defining method of invoke user are checked in sampling.

Scheme 3: a kind of preferred version for scheme 2 is characterized in that the method for obtaining the actual loading O of system (t) is: for each virtual machine, a working set is chosen in random sampling, obtains the load state of whole virtual machine through the load of statistical work collection.

Scheme 4: a kind of preferred version for scheme 1 is characterized in that the implementation method of step B is:

B1. under the initial situation, promptly during t=0, get E (t)=O (t), change step C over to; Otherwise, change step B2 over to;

B2. the square error of last short-term forecasting is f [α (t)]=[E (t-1)-O (t)] ², with classical EWMA formula E (t)=α (t) E (t-1)+[1-α (t)] O (t) substitution, launch, utilize the historical data of preserving, instrument error polynomial f [α (t)]={ [E (t-3)-O (t-2)] * α (t) ²+ [O (t-2)-O (t-1)] * α (t)+[O (t-1)-O (t)] } ², this is a polynomial function about α (t);

B3. to error polynomial f [α (t)] differentiate, obtain derived function f ' [α (t)], solve its all on [1,1] interval and separate; If error polynomial does not exist or f ' [α (t)] separates in [1,1] last nothing, get α (t)=-1, change step B5 over to; Otherwise because f [α (t)] number of times is no more than four times, so the skill of f ' [α (t)] on [1,1] be no more than 3, and disaggregation is designated as S, changes step B4 over to;

B4. { 1, the functional value of all elements in the 1} compares their size, and choosing minimum in them a pairing argument value is α (t) to S ∪ to ask error polynomial f [α (t)]; If the functional value of a plurality of minimums is arranged, get that minimum independent variable as α (t);

B5. calculate E (t)=α (t) E (t-1)+[1-α (t)] O (t).

Scheme 5: a kind of preferred version for scheme 4 is characterized in that the implementation method of step C is:

C1. under the initial situation, promptly during t＜T, E _T(t)=and O (t), change step D over to; Otherwise, change step C2 over to;

C2. the square error of last long-term forecasting is f _T[α _T(t)]=[E _T(t-T)-O (t)] ², with the periodicity popularizing form E of classical EWMA formula _T(t)=α _T(t) E _T(t-T)+[1-α _T(t)] O (t) substitution launches, and utilizes the historical data of preserving, the instrument error polynomial f _T[α _T(t)]={ [E _T(t-3T)-and O (t-2T)] * α _T(t) ²+ [O (t-2T)-O (t-T)] * α _T(t)+[O (t-T)-O (t)] } ², this be one about α _T(t) polynomial function;

C3. to above-mentioned error polynomial f _T[α _T(t)] differentiate obtains derived function f _T' [α _T(t)], solving its all on [1,1] interval separates; If error polynomial does not exist or f _T' [α _T(t)] go up nothing in [1,1] and separate, get α _T(t)=-1, change step C5 over to; Otherwise, because f _T[α _T(t)] number of times is no more than four times, so f _T' [α _T(t)] skill on [1,1] is no more than 3, and disaggregation is designated as S _T, change step C4 over to;

C4. ask error polynomial f _T[α _T(t)] to S _T{ 1, the functional value of all elements in the 1} compares their size to ∪, and choosing minimum in them a pairing argument value is α _T(t); If the functional value of a plurality of minimums is arranged, get that minimum independent variable as α _T(t);

C5. calculate E _T(t)=α _T(t) E _T(t-T)+[1-α _T(t)] O (t).

Scheme 6: a kind of preferred version for scheme 4 is characterized in that the method for finding the solution collection S among the step B3 is:

1), utilize the red formula of card to try to achieve all three complex roots for cubic function; For quadratic function, utilize radical formula to try to achieve all two complex roots; For linear function, directly utilize slope and skew to try to achieve unique real number root;

2) all complex roots only keep its real part;

3) leave out all not roots in [1,1], remaining formation disaggregation S.

The present invention is through real-time calculation of alpha of error function (t) and α _T(t), improved the precision of prediction of classical EWMA; Through with α (t) and α _T(t) extend on [1,1] interval, make predicted value can satisfy the demand that is slightly larger than actual value; Through introducing the long-term forecasting module, improved the reaction capacity of prediction to cloud computing platform duty cycle property.

Description of drawings

Fig. 1 main frame flow process.

Fig. 2 finds the solution the diversity method process flow diagram for derived function among the embodiment.

Embodiment

Through instance the present invention is done further explanation below.

The actual loading sequence that will run into after equipment, method begins is 50,42,35,31,26,22,21, cycle T=2.

During t=0, obtain O (0)=50 through systems approach; Carrying out short-term forecasting constantly, because t=0, so E (0)=50; Carrying out long-term forecasting constantly, because t＜T, so E _T(0)=50; When comprehensive, because t＜T, so export 50; The historical data of this moment has:

t	0
		O(t)	50
E(t)	50

E T(t)

50

During t=1, obtain O (1)=42 through systems approach;

When carrying out short-term forecasting, the error polynomial of structure does not exist, and disaggregation S is an empty set, therefore gets α _T(1)=-1, with EWMA calculate E (1)=-1*E (0)+2*O (1)=34;

When carrying out long-term forecasting, because t＜T, so E _T(0)=42;

When comprehensive, because t＜T, so export 42;

The historical data of this moment has:

t	0	1
			O(t)	50	42
E(t)	50	34
			E T(t)	50	42

During t=2, obtain O (2)=35 through systems approach;

When carrying out short-term forecasting, the error polynomial of structure does not exist, and disaggregation S is an empty set, therefore gets α (2)=-1, with EWMA calculate E (2)=-1*E (1)+2*O (2)=36;

When carrying out long-term forecasting, the error polynomial of structure does not exist, disaggregation S _TDo not exist, therefore get α _T(2)=-1, calculate E with EWMA _T(2)=-1*E _T(0)+2*O (2)=20;

When comprehensive, output max{E (1), E _T(0) }=50;

The historical data of this moment has:

t	0	1	2
				O(t)	50	42	35
E(t)	50	34	36
				E T(t)	50	42	20

During t=3, obtain O (3)=31 through systems approach;

When carrying out short-term forecasting, the error polynomial of structure is f [α (t)]=64* α (t) ⁴+ 112* α (t) ³+ 113* α (t) ²+ 56* α (t)+16, the derived function that obtains after the differentiate is f ' [α (t)]=256* α (t) ³+ 336* α (t) ²+ 226* α (t)+56;

It is (see accompanying drawing 2, below all use this method) that derived function used herein is found the solution diversity method:

1. for cubic function, utilize the red formula of card to try to achieve all three complex roots; For quadratic function, utilize radical formula to try to achieve all two complex roots; For linear function, directly utilize slope and skew to try to achieve unique real number root;

2. all complex roots only keep its real part;

3. leave out all not roots in [1,1], remaining formation disaggregation S;

Solve S={-0.4375},, get α (3)=-0.4375 through relatively, with EWMA calculate E (3)=-0.4375*E (2)+1.4375*O (3)=28.8;

When carrying out long-term forecasting, the error polynomial of structure does not exist, disaggregation S _TDo not exist, therefore get α _T(3)=-1, calculate E with EWMA _T(3)=-1*E _T(1)+2*O (3)=20;

When comprehensive, output max{E (2), E _T(1) }=42;

The historical data of this moment has:

t	0	1	2	3
					O(t)	50	42	35	31
E(t)	50	34	36	28.8
					E T(t)	50	42	20	20

During t=4, obtain O (4)=26 through systems approach; When carrying out short-term forecasting, the error polynomial of structure is f [α (t)]=α (t) ⁴-8* α (t) ³+ 6* α (t) ²+ 40* α (t)+25, the derived function that obtains after the differentiate is f ' [α (t)]=4* α (t) ³-24* α (t) ²+ 12* α (t)+40;

Solve S={-1},, get α (4)=-1 through relatively, with EWMA calculate E (4)=-1*E (3)+2*O (4)=23.2;

When carrying out long-term forecasting, the error polynomial of structure does not exist, disaggregation S _TDo not exist, therefore get α _T(4)=-1, calculate E with EWMA _T(4)=-1*E _T(2)+2*O (4)=32;

When comprehensive, output max{E (3), E _T(2) }=28.8; The historical data of this moment has:

t	0	1	2	3	4
						O(t)	50	42	35	31	26
E(t)	50	34	36	28.8	23.2
						E T(t)	50	42	20	20	32

During t=5, obtain O (5)=22 through systems approach;

When carrying out short-term forecasting, the error polynomial of structure is f [α (t)]=25* α (t) ⁴+ 50* α (t) ³+ 65* α (t) ²+ 40* α (t)+16, the derived function that obtains after the differentiate is f ' [α (t)]=100* α (t) ³+ 150* α (t) ²+ 130* α (t)+40;

Solve S={-0.5},, get α (5)=-0.5 through relatively, with EWMA calculate E (5)=-0.5*E (4)+1.5*O (5)=21.4;

When carrying out long-term forecasting, the error polynomial of structure does not exist, disaggregation S _TDo not exist, therefore get α _T(5)=-1, calculate E with EWMA _T(5)=-1*E _T(3)+2*O (5)=24;

When comprehensive, output max{E (4), E _T(3) }=23.2; The historical data of this moment has:

t	0	1	2	3	4	5
							O(t)	50	42	35	31	26	22
E(t)	50	34	36	28.8	23.2	21.4
							E T(t)	50	42	20	20	32	24

During t=6, obtain O (6)=21 through systems approach;

When carrying out short-term forecasting, the error polynomial of structure is f [α (t)]=7.9* α (t) ⁴+ 22.5* α (t) ³+ 21.6* α (t) ²+ 8* α (t)+1, the derived function that obtains after the differentiate is f ' [α (t)]=31.6* α (t) ³+ 67.5* α (t) ²+ 43.2* α (t)+8;

Solve S={-0.3236 ,-0.7111} through relatively, gets α (6)=-0.3236, with EWMA calculate E (6)=-0.3236*E (5)+1.3236*O (6)=20.9;

When carrying out long-term forecasting, the instrument error polynomial f _T[α _T(t)]=225* α _T(t) ⁴+ 270* α _T(t) ³+ 231* α _T(t) ²+ 90* α _T(t)+25, the derived function that obtains after the differentiate is f _T' [α _T(t)]=900* α _T(t) ³+ 810* α _T(t) ²+ 462* α _T(t)+90;

Disaggregation S _T={ therefore 0.3} gets α _T(6)=-0.3, calculate E with EWMA _T(6)=-1*E _T(4)+2*O (6)=17.7;

When comprehensive, output max{E (5), E _T(4) }=32; The historical data of this moment has:

t	0	1	2	3	4	5	6
								O(t)	50	42	35	31	26	22	21
E(t)	50	34	36	28.8	23.2	21.4	20.9
								E T(t)	50	42	20	20	32	24	17.7

Need to prove:

-α (t) and α _T(t) span, what adopt in this method is [1,1], but also available other interval as (∞ ,+∞) replacements;

-find the solution the derived function method at zero point, except directly applying mechanically the formula, can also be with the method for enumerating: with 0.001 or other values be step-length, enumerate α (t) or α _T(t) possible interval finds zero point;

-have under the situation of a plurality of minimum point, what select in this method is to get that minimum independent variable, but also can be with median or maximal value replacement.

Claims (6)

1. the cloud computing system load predicting method that can regulate parameter automatically is characterized in that, comprises the steps:

A. at moment t, calculate the actual loading O of system (t) of t constantly through system call;

B. carry out short-term forecasting: utilize O (t) value and historical data calculation of alpha (t) and E (t);

C. carry out long-term forecasting: utilize O (t) and historical data calculation of alpha _T(t) and E _T(t);

D. comprehensive short-term forecasting and long-term forecasting: work as t<during T, output O (t) changes step e; Otherwise, get E (t-1), E _T(t-T) maximal value of the two or mean value are as the output of moment t;

E. upgrade historical data, wait for t+1 constantly, change steps A;

Wherein, T representes a duty cycle; E (t) is illustrated in the short-term forecasting value of moment t to moment t+1 system load; E _T(t) be illustrated in the long-term forecasting value of moment t to moment t+T system load; α (t) expression is used to ask the parameter of E (t); α _T(t) expression is used to ask E _T(t) parameter.

2. the method for claim 1 is characterized in that, the system call in the steps A is: call the application programming interfaces that virtual machine manager provides, perhaps virutal machine memory page or leaf, the perhaps self-defining method of invoke user are checked in sampling.

3. method as claimed in claim 2 is characterized in that, the method for obtaining the actual loading O of system (t) is: for each virtual machine, a working set is chosen in random sampling, obtains the load state of whole virtual machine through the load of statistical work collection.

4. the method for claim 1 is characterized in that, the implementation method of step B is:

B1. under the initial situation, promptly during t=0, get E (t)=O (t), change step C over to; Otherwise, change step B2 over to;

B2. the square error of last short-term forecasting is f [α (t)]=[E (t-1)-O (t)] ², with classical EWMA formula E (t)=α (t) E (t-1)+[1-α (t)] O (t) substitution, launch, utilize the historical data of preserving, instrument error polynomial f [α (t)]={ [E (t-3)-O (t-2)] * α (t) ²+ [O (t-2)-O (t-1)] * α (t)+[O (t-1)-O (t)] } ²

B3. to error polynomial f [α (t)] differentiate, obtain derived function f ' [α (t)], solve its all on [1,1] interval and separate; If error polynomial does not exist or f ' [α (t)] separates in [1,1] last nothing, get α (t)=-1, change step B5 over to; Otherwise because f [α (t)] number of times is no more than four times, so the skill of f ' [α (t)] on [1,1] be no more than 3, and disaggregation is designated as S, changes step B4 over to;

B4. { 1, the functional value of all elements in the 1} compares their size, and choosing minimum in them a pairing argument value is α (t) to S ∪ to ask error polynomial f [α (t)]; If the functional value of a plurality of minimums is arranged, get that minimum independent variable as α (t);

B5. calculate E (t)=α (t) E (t-1)+[1-α (t)] O (t).

5. method as claimed in claim 4 is characterized in that, the implementation method of step C is:

C1. under the initial situation, i.e. t<during T, E _T(t)=and O (t), change step D over to; Otherwise, change step C2 over to;

C2. the square error of last long-term forecasting is f _T[α _T(t)]=[E _T(t-T)-O (t)] ², with the periodicity popularizing form E of classical EWMA formula _T(t)=α _T(t) E _T(t-T)+[1-α _T(t)] O (t) substitution launches, and utilizes the historical data of preserving, the instrument error polynomial f _T[α _T(t)]={ [E _T(t-3T)-and O (t-2T)] * α _T(t) ²+ [O (t-2T)-O (t-T)] * α _T(t)+[O (t-T)-O (t)] } ²

C3. to above-mentioned error polynomial f _T[α _T(t)] differentiate obtains derived function f _T' [α _T(t)], solving its all on [1,1] interval separates; If error polynomial does not exist or f _T' [α _T(t)] go up nothing in [1,1] and separate, get α _T(t)=-1, change step C5 over to; Otherwise, because f _T[α _T(t)] number of times is no more than four times, so f _T' [α _T(t)] skill on [1,1] is no more than 3, and disaggregation is designated as S _T, change step C4 over to;

C4. ask error polynomial f _T[α _T(t)] to S _T{ 1, the functional value of all elements in the 1} compares their size to ∪, and choosing minimum in them a pairing argument value is α _T(t); If the functional value of a plurality of minimums is arranged, get that minimum independent variable as α _T(t);

C5. calculate E _T(t)=α _T(t) E _T(t-T)+[1-α _T(t)] O (t).

6. method as claimed in claim 4 is characterized in that, the method for finding the solution collection S among the step B3 is:

1), utilize the red formula of card to try to achieve all three complex roots for cubic function; For quadratic function, utilize radical formula to try to achieve all two complex roots; For linear function, directly utilize slope and skew to try to achieve unique real number root;

2) all complex roots only keep its real part;

3) leave out all not roots in [1,1], remaining formation disaggregation S.

Images