CN116708581A - High-reliability function scheduling method for server-free edge computing - Google Patents

Abstract

The invention provides a high-reliability function scheduling method for server-free edge calculation, which comprises the following steps: presetting a personality-driven application program QoS prediction method, and predicting QoS of a single application program under different user personality types based on the application program QoS prediction method; constructing a deterministic function scheduling algorithm developed in combination with the enhanced NSGA-II aiming at a specific problem; a random function scheduling strategy is provided; and performing parallel function scheduling optimization. The invention can well balance the service profit of the target network and the overall application QoS of the application of the Internet of things on the premise of meeting all design constraints.

Description

High-reliability function scheduling method for server-free edge computing

Technical Field

The invention belongs to the technical field of function scheduling in server edge computing, and particularly relates to a high-reliability function scheduling method for non-server edge computing.

Background

With the popularity of internet of things (IoT) applications, smart medicine and augmentation, virtual reality, etc., the demand for low-latency computing has exploded. Edge computing provides an excellent opportunity to meet computing needs by deploying network, control, computing, and storage infrastructure between end users and cloud servers. In this way, most of the data generated by the distributed internet of things facilities can be properly processed near the source from which they were generated. In essence, edge computing extends the mature cloud computing paradigm from a centralized architecture to network edges to improve service latency of internet of things applications. Recently, the concept of server-less computing has attracted widespread attention by edge computing communities.

In serverless edge computing, one topical theme is to handle delay-constrained internet of things application scheduling that includes a dependency function to maximize the profit of the service provider. In practice, this is a particularly challenging task, mainly because there is a personalized user demand in the internet of things environment regarding service payment and application delays. At the same time, energy and reliability issues present an increasingly serious challenge to server-less edge computing. Although the existing work has been studied in abundance, none has taken into account the important factors of energy, reliability, personalized user needs and random application execution. The invention discloses a personalized function scheduling method (REPFS) for guaranteeing reliability, which is used for balancing profit of a service provider and integral quality of service (QoS) of a random Internet of things application program in sustainable server-free edge calculation.

Disclosure of Invention

In order to solve the technical problems, the invention provides the high-reliability function scheduling method for the edge calculation of the no-server, which can well balance the service profit of the target network and the overall application QoS of the application of the Internet of things on the premise of meeting all design constraints.

In order to achieve the above object, the present invention provides a method for scheduling a function with high reliability for edge computation without a server, comprising:

step 1: presetting a personality-driven application program QoS prediction method, and predicting QoS of a single application program under different user personality types based on the application program QoS prediction method;

step 2: constructing a deterministic function scheduling model, and based on the deterministic function scheduling model, accurately estimating the QoS predicted in the step 1;

step 3: constructing a random function scheduling strategy, and scheduling the QoS after accurate estimation based on the random function scheduling strategy;

step 4: and performing parallel function scheduling optimization, wherein the parallel function scheduling optimization is used for adapting to a random function scheduling strategy of a modern multi-core scheduler platform to accelerate generation of a solution.

Optionally, predicting QoS of a single application under different user personality types based on the application QoS prediction method includes:

step 101: initializing a set of iteration markersWherein I is _par Representing the number of users participating in the questionnaire, which is a constant less than the user population I;

step 102: initializing preference factors for participating questionnaire users

Step 103: judging whether all iteration marks are zero, if yes, turning to step 111, otherwise, turning to step 104;

Step 104: calculation ofWherein j is more than or equal to 1 and I is more than or equal to 1 _par J is the index of the questionnaire participant, L represents the number of all possible delay-payment level combinations, θ _j,l And Q _j,l Respectively represent the questionnaire participants U at the first delayed payment level _j Is>And->Are all preference value parameters in the linear regression problem;

step 105: iteratively deriving preference factors for all participants in the questionnaire using a linear regression solver;

step 106: computing participant U _j Is of a deviation of (1)Good factorThe formula is:

whereinAnd->Respectively represent the participants U in the first delay-payment level combination _j Response delay and service payment fee;

step 107: comparing randomly generated preference factors eta _j And derived preference factorsAbsolute difference ρ between;

step 108: judging whether the absolute difference value rho is smaller than a preset value, if so, turning to step 109, otherwise, turning to step 110;

step 109: participant U _j Iteration flag of (a)Set to zero, i.e.)>

Step 110: matrix updating using Latin hypercube sampling technique

Step 111: let S _j ＝[AScore _j ，CScore _j ，EScore _j ，OScore _j ，NScore _j ]Representing participant U _j Is described, wherein the user personality vector is decomposed into five different features: humanized, dead and external Tropism, openness and neuromatter, respectively correspond to AScore _j 、CScore _j 、EScore _j 、OScore _j And NScore _j ；

Step 112: calculating a linear mapping from personality score to user preference, the formula:wherein-> The representation dimension is I _par X1 preference factor matrix, X _5×1 Is a mapping matrix with 5 multiplied by 1 dimension, and shows the influence of personality score on user preference;

step 113: for the user group with the total number of I, constructing any user U in a traversing mode _i Is the personality score vector S of (1) _i Wherein I (1.ltoreq.i.ltoreq.i) is an index of the user;

step 114: computing a matrixStored user U _i Preference factor of->The formula is: /> An optimal preference factor matrix with a total number of rows of I and a total number of columns of 5, wherein I is the total number of users, < >>Personality score matrix with total number of rows of I and total number of columns of 5>An influence matrix of personality scores with total number of rows of I and total number of columns of 5 on user preference is represented;

step 115: calculating input delay-payment combinationsDown prediction application +.>QoS Q of (C) _i,l The formula is: />Wherein the method comprises the steps ofFor preference value parameter in step 104 +.>Is>For preference value parameter in step 104 +.>Is>For application program->Preference factor of->Representing user U _i Response delay at the first class of service, +. >Representing user U _i Payment at the first service level;

step 116: the QoS of the application is output.

Optionally, based on the deterministic function scheduling model, accurately estimating the QoS predicted in step 1 includes:

step 201: initializing an iteration counter g, namely g+.0;

step 202: selecting chromosomes with less constraint violations to initialize populationsWherein U is the total number of users;

step 203: initializing operation marks of three tables, namely a function sequence, a function-to-container scheduling and a resource-to-container allocation, namely a flag of (ζ) ζ 1;

step 204: judging iteration counter g < g _max If yes, go to step B5 to enter the population evolution process, otherwise go to step 229, wherein g _max Is the upper bound of the iteration counter;

step 205: calling a Modulo function module (flag++, 3) to obtain the remainder of dividing an operation mark flag by 3, and if the result is 1, performing evolutionary operation on a function sequence table; otherwise, carrying out evolutionary operation on the allocation table from the function to the container and the allocation table from the resource to the container;

step 206: based on Gaussian distributionGenerating a set of cut-off samples->Wherein x is a cut-off sample index within the range of x being more than or equal to 1 and U, and the average value is +. >Variance is->

Step 207: indexing two chromosomesAnd->Set to the values of 1 and U, respectively, i.e. +.>

Step 208: judgingIf yes, go to step 215, otherwise go to step 209;

step 209: judging whether flag=1 is satisfied, if so, indexingAnd->Respectively assigned to the cut-off sample index x ₁ And x ₂ I.e. x ₁ ←u，x ₂ ζ; otherwise, index ++>Simultaneous assignment of value to cut-off sample index x ₁ And x ₂ I.e. x ₁ ←u，x ₂ ≡u, indicating that the same cut-off sample was selected;

step 210: using split functionsChromosome->Contained element tableAccording to the truncated sample->Divided into->And->Two parts, i.e.)>

Step 211: using split functionsChromosome->Contained element tableAccording to the truncated sample->Divided into->And->Two parts, i.e.)>

Step 212: using a join functionUpdate element->I.e. < ->

Step 213: using a join functionUpdate element->I.e. < ->

Step 214: updating chromosome index for crossover operations of other chromosomesTurning to step 208;

step 215: computing gaussian distribution of crossing samples of next generation populationThe formula isAnd->Wherein K is the total number of observation intervals, H is the total number of position samples, +.>Weight for the h position sample, +.>Is the observation value of the kth interval of the sample at the h position, h is the algebra of population evolution, { circumflex } >Is the mean value of the evolution of the population at the next iteration of the flag,is the variance;

step 216: if it isIf yes, the cross operation is completed, and the step 217 is executed, otherwise, the step 205 is executed;

step 217: starting to run the mutation operation, judgingIf yes, go to step 218 if true, otherwise go to step 224;

step 218: each chromosomeCalculating function order table +.>Resource-to-container allocation tableFunction to container schedule->Variation probability vector->And->The formula is: />Wherein->Is chromosome->Variation factor of the kth interval, +.>Is chromosome->Coefficient of variation, g _th And g _max Sequentially the threshold value and the maximum value of the iteration counter g, phi ₄ Is constant (I)>Is used for +.>Scaled to [0,1 ]]Is a normalized function of->Is the mutation probability;

step 219: random generation falls within [0,1 ]]Variation probability threshold sigma of (2) ₁ 、σ ₂ Sum sigma ₃ I.e. sigma ₁ ,σ ₂ ,σ ₃ ←Random(0,1)；

Step 220: for element tableEach position +.>If the mutation probability->Greater than threshold sigma ₁ Variant function->Applied to the position/>

Step 221: for element tableIs +.>If the probability of variationGreater than threshold sigma ₂ Then the variation function will be scheduled->Applied to position->

Step 222: for element tableEach position +.>If the mutation probability- >Greater than threshold sigma ₃ Then assign the variation function->Applied to position->

Step 223: once the current population evolution is completed, entering a competition selection stage;

step 224: deducing an application according to the method of step 1QoS of program and calculates profit gamma of service provider _profit ；

Step 225: arranging all chromosomes in a descending order by adopting a non-dominant ordering genetic algorithm;

step 226: calculation of chromosomesTotal constraint violation degree Φ _u The formula is: />Wherein the method comprises the steps ofTo the extent that only the energy constraint is violated, +.>To the extent that only reliability constraints are violated;

step 227: calling population generation functionsTo construct the next generation population +.>The ratio-changing method is utilized to enable the proportion of infeasible individuals remained in the population to be linearly reduced along with the iteration times of the population;

step 228: updating the iteration counter g, namely: g≡g+1, go to step 204;

step 229: after the population iteration of preset times, a non-dominant solution set is outputAnd then terminates.

Optionally, scheduling the QoS after accurate estimation based on the random function scheduling policy includes:

step 301: assuming that the total number of users is I, the execution adaptation variable of the I users is alpha= { alpha ₁ ,α ₂ ,…,α _I Then iterateGenerating an execution adaptation variable α= { α ₁ ,α ₂ ,…,α _I Deriving a deterministic function schedule that limits the deadline miss rate for all applications to within design requirements, a problem-specific enhancement NSGA-II for use in executing the adaptation variable α= { α ₁ ,α ₂ ,…,α _I Generating a deterministic schedule at a particular value;

step 302: performing adaptive variable alpha= { alpha for simulating and evaluating user population as I in terms of application deadline deletion rate by using Monte Carlo ₁ ,α ₂ ,…,α _I }；

Step 303: evaluating performance of the current function schedule using a time metric of deadline miss rate;

step 304: judging whether the estimated expiration date deletion rate is greater than a threshold value, if so, turning to step 305; otherwise, go to step 306;

step 305: updating the execution adaptation variable α= { α of I ₁ ,α ₂ ,…,α _I Step 301; wherein I is the total number of users;

step 306: the current function schedule and associated execution adaptation variables are thus a priority scheduling solution and the entire iterative process exits.

Optionally, the parallel function scheduling optimization includes:

step 401: creating an execution adaptation variable set α= { α ₁ ,α ₂ ,…,α _I -I is the total number of users;

step 402: will execute the adaptation variable alpha ₁ ,α ₂ ,…,α _I Assigned 0, i.e. alpha ₁ ,α ₂ ,…,α _I The number of users is 0,I;

step 403: setting the search step size to a fixed value in the interval [0,1 ];

Step 404: traversing alpha according to the search step _i E, alpha, I is the index of the user, which satisfies that 1 is less than or equal to I, and for the variable alpha _i Performing a round of monte carlo simulation to search for application QoS;

step 405: the function scheduling scheme is firstly satisfiedVariable alpha of the date loss rate constraint _i The value is assigned as the value of the random attribute executed by the application.

Compared with the prior art, the application has the following advantages and technical effects:

the application designs a personalized function scheduling technology with guaranteed reliability for sustainable server-free edge calculation through a personalized driven application QoS prediction method, a deterministic function scheduling strategy based on enhanced NSGA-II and a random parallel function scheduling strategy. The application can well balance the service profit of the target network and the overall application QoS of the application of the Internet of things on the premise of meeting all design constraints.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

FIG. 1 is a schematic flow chart of a method according to an embodiment of the application;

FIG. 2 is a schematic diagram showing a comparison of the number of functions contained in each DAG after sampling a total of 2119 DAG applications having different depth and width sizes in accordance with an embodiment of the present application;

FIG. 3 (a) is a schematic diagram of predicted application QoS for 200 users; FIG. 3 (b) is a diagram of predicted application QoS for 300 users; FIG. 3 (c) is a diagram of predicted application QoS for 400 users; FIG. 3 (d) is a diagram of predicted application QoS for 500 users; FIG. 3 (e) is a diagram of predicted application QoS for 600 users; FIG. 3 (f) is a diagram of predicted application QoS for 700 users;

FIG. 4 is a diagram illustrating a total estimation error of a QoS prediction method for a personal driver according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a run-time comparison of an embodiment of the present application based on an enhanced NSGA-II deterministic algorithm and six baseline algorithms, B-NSGA-II, c-DPEA, DLS-MOEA, MOALO, MOEA/D and MaOEA/IGD implementations;

FIG. 6 is a comparative schematic diagram of various parallelization level runtime of an embodiment of the present application.

Detailed Description

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.

It should be noted that the steps illustrated in the flowcharts of the figures may be performed in a computer system such as a set of computer executable instructions, and that although a logical order is illustrated in the flowcharts, in some cases the steps illustrated or described may be performed in an order other than that illustrated herein.

The REPFS scheduling proposed by the present invention, as shown in FIG. 1, comprises the following steps:

step 1: presetting a personality-driven application program QoS prediction method, and predicting QoS of a single application program under different user personality types based on the application program QoS prediction method;

step 2: constructing a deterministic function scheduling model, and based on the deterministic function scheduling model, accurately estimating the QoS predicted in the step 1;

step 3: constructing a random function scheduling strategy, and scheduling the QoS after accurate estimation based on the random function scheduling strategy;

step 4: and performing parallel function scheduling optimization, wherein the parallel function scheduling optimization is used for adapting to a random function scheduling strategy of a modern multi-core scheduler platform to accelerate generation of a solution.

Wherein. In order to solve the problem of uncertainty optimization, a deterministic function scheduling algorithm is difficult to develop, step 3 is a random function scheduling strategy, the strategy considers the uncertainty of the instruction cycle number and the communication data volume of each Internet of things application program, and step 4 is the parallel implementation of the scheduling algorithm, so that the generation of a solution can be accelerated.

The method for predicting the QoS of the application program driven by the personality in the step 1 specifically comprises the following steps:

step 101: initializing a set of iteration markersWherein I is _par Representing the number of users participating in the questionnaire, which is a constant less than the user population I;

step 102: initializing preference factors for participating questionnaire users

Step 103: judging whether all iteration marks are zero, if yes, turning to step 111, otherwise, turning to step 104;

step 104: calculation ofWherein j is more than or equal to 1 and I is more than or equal to 1 _par J is the index of the questionnaire participant, L represents the number of all possible delay-payment level combinations, θ _j,l And Q _j,l Respectively represent the questionnaire participants U at the first delayed payment level _j Is>And->Are all preference value parameters in the linear regression problem;

step 105: iteratively deriving preference factors for all participants in the questionnaire using a linear regression solver;

step 106: computing participant U _j Preference factor of (a)The formula is:

wherein->And->Respectively represent the participants U in the first delay-payment level combination _j Response delay and service payment fee;

step 107: comparing randomly generated preference factors eta _j And derived preference factorsAbsolute difference ρ between;

step 108: judging whether the absolute difference value rho is smaller than a preset value, if so, turning to step 109, otherwise, turning to step 110;

Step 109: participant U _j Iteration flag of (a)Set to zero, i.e.)>

Step 110: matrix updating using Latin hypercube sampling technique

Step 111: let S _j ＝[AScore _j ，CScore _j ，EScore _j ，OScore _j ，NScore _j ]Representing participant U _j Is described, wherein the user personality vector is decomposed into five different features: humanization, responsibility, exogenousness, openness and neuro-properties, corresponding to AScore _j 、CScore _j 、EScore _j 、OScore _j And NScore _j ；

Step 112: calculating a linear mapping from personality score to user preference, the formula:wherein-> The representation dimension is I _par X1 preference factor matrix, X _5×1 Is a mapping matrix with 5 multiplied by 1 dimension, and shows the influence of personality score on user preference;

step 113: for the user group with the total number of I, constructing any user U in a traversing mode _i Is the personality score vector S of (1) _i Wherein I (1.ltoreq.i.ltoreq.i) is an index of the user;

step 114: computing a matrixStored user U _i Preference factor of->The formula is: /> An optimal preference factor matrix with a total number of rows of I and a total number of columns of 5, wherein I is the total number of users, < >>Personality score matrix with total number of rows of I and total number of columns of 5>An influence matrix of personality scores with total number of rows of I and total number of columns of 5 on user preference is represented;

step 115: calculating input delay-payment combinations Down prediction application +.>QoS Q of (C) _i，l The formula is: />Wherein the method comprises the steps ofFor preference value parameter in step 104 +.>Is>For preference value parameter in step 104 +.>Is>For application program->Preference factor of->Representing user U _i Response delay at the first class of service, +.>Representing user U _i Payment at the first service level;

step 116: the QoS of the application is output.

The deterministic function scheduling of the enhanced NSGA-II based on the specific problem in the step 2 specifically comprises the following steps:

step 201: initializing an iteration counter g, namely g+.0;

step 202: selecting chromosomes with less constraint violations to initialize populationsWherein U is the total number of users;

step 203: initializing operation marks of three tables, namely a function sequence, a function-to-container scheduling and a resource-to-container allocation, namely a flag of (ζ) ζ 1;

step 204: judging iteration counter g < g _max If yes, go to step B5 to enter the population evolution process, otherwise go to step 229, wherein g _max Is the upper bound of the iteration counter;

step 205: calling a Modulo function module (flag++, 3) to obtain the remainder of dividing an operation mark flag by 3, and if the result is 1, performing evolutionary operation on a function sequence table; otherwise, carrying out evolutionary operation on the allocation table from the function to the container and the allocation table from the resource to the container;

Step 206: based on Gaussian distributionGenerating a set of cut-off samples->Wherein x is a cut-off sample index within the range of x being more than or equal to 1 and U, and the average value is +.>Variance is->

Step 207: indexing two chromosomesAnd->Set to the values of 1 and U, respectively, i.e. +.>

Step 208: judgingWhether or not it is true, if it isStanding, turning to step 215, otherwise turning to step 209;

step 209: judging whether flag=1 is satisfied, if so, indexingAnd z are assigned to the cut-off sample index x, respectively ₁ And x ₂ I.e. x ₁ ←u，x ₂ ζ; otherwise, index ++>Simultaneous assignment of value to cut-off sample index x ₁ And x ₂ I.e. x ₁ ←u，x ₂ ≡u, indicating that the same cut-off sample was selected;

step 210: using split functionsChromosome->Contained element tableAccording to the truncated sample->Divided into->And->Two parts, i.e.)>

Step 211: using split functionsChromosome->Contained element tableAccording to the truncated sample->Divided into->And->Two parts, i.e.)>

Step 212: using a join functionUpdate element->I.e. < ->

Step 213: using a join functionUpdate element->I.e. < ->

Step 214: updating chromosome index for crossover operations of other chromosomesTurning to step 208;

step 215: computing gaussian distribution of crossing samples of next generation population The formula isAnd->Wherein K is the total number of observation intervals, H is the total number of position samples, +.>Weight for the h position sample, +.>Is the observation value of the kth interval of the sample at the h position, g is the algebra of population evolution,/and>is the mean value of the evolution of the population at the next iteration of the flag,is the variance;

step 216: if it isIf yes, the cross operation is completed, and the step 217 is executed, otherwise, the step 205 is executed;

step 217: starting to run the mutation operation, judgingIf yes, go to step 218 if true, otherwise go to step 224;

step 218: each chromosomeCalculating function order table +.>Resource-to-container allocation tableFunction to container schedule->Variation probability vector->And->The formula is: />Wherein->Is chromosome->Variation factor of the kth interval, +.>Is chromosome->Coefficient of variation, g _th And g _max Sequentially the threshold value and the maximum value of the iteration counter g, phi ₄ Is constant (I)>Is used for +.>Scaled to [0,1 ]]Is a normalized function of->Is the mutation probability;

step 219: random generation falls within [0,1 ]]Variation probability threshold sigma of (2) ₁ 、σ ₂ Sum sigma ₃ I.e. sigma ₁ ,σ ₂ ,σ ₃ ←Random(0,1)；

Step 220: for element tableEach position +.>If the mutation probability->Greater than threshold sigma ₁ Variant function->Applied to position- >

Step 221: for element tableIs +.>If the probability of variationGreater than threshold sigma ₂ Then the variation function will be scheduled->Applied to position->

Step 222: for element tableEach position +.>If the mutation probability->Greater than threshold sigma ₃ Then assign the variation function->Applied to position->

Step 223: once the current population evolution is completed, entering a competition selection stage;

step 224: according to the method of step 1, qoS of the application is deduced, and profit gamma of the service provider is calculated _profit ；

Step 225: arranging all chromosomes in a descending order by adopting a non-dominant ordering genetic algorithm;

step 226: calculation of chromosomesTotal constraint violation degree Φ _u The formula is: />Wherein the method comprises the steps ofTo the extent that only the energy constraint is violated, +.>To the extent that only reliability constraints are violated;

step 227: calling population generation functionsTo construct the next generation population +.>The ratio-changing method is utilized to enable the proportion of infeasible individuals remained in the population to be linearly reduced along with the iteration times of the population;

step 228: updating the iteration counter g, namely: g≡g+1, go to step 204;

step 229: after the population iteration of preset times, a non-dominant solution set is outputAnd then terminates.

Step 3 provides a random function scheduling strategy, which specifically comprises the following steps:

Step 301: assuming that the total number of users is I, the execution adaptation variable of the I users is alpha= { alpha ₁ ，α ₂ ，…，α _I Iterative generation of the execution adaptation variable α= { α ₁ ，α ₂ ，…，α _I Deriving a deterministic function schedule that limits the deadline miss rate for all applications to within design requirements, a problem-specific enhancement NSGA-II for use in executing the adaptation variable α= { α ₁ ，α ₂ ，…，α _I Generating a deterministic schedule at a particular value;

step 302: performing adaptive variable alpha= { alpha for simulating and evaluating user population as I in terms of application deadline deletion rate by using Monte Carlo ₁ ，α ₂ ，…，α _I }；

Step 303: evaluating performance of the current function schedule using a time metric of deadline miss rate;

step 304: judging whether the estimated expiration date deletion rate is greater than a threshold value, if so, turning to step 305; otherwise, go to step 306;

step 305: updating the execution adaptation variable α= { α of I ₁ ，α ₂ ，…，α _I Step 301; wherein I is the total number of users;

step 306: the current function schedule and associated execution adaptation variables are thus a priority scheduling solution and the entire iterative process exits.

Step 4, researching parallel function scheduling optimization, specifically comprising the following steps:

step 401: creating an execution adaptation variable set α= { α ₁ ，α ₂ ，…，α _I -I is the total number of users;

Step 402: will execute the adaptation variable alpha ₁ ，α ₂ ，…，α _I Assigned 0, i.e. alpha ₁ ，α ₂ ，…，α _I The number of users is 0,I;

step 403: setting the search step size to a fixed value in the interval [0,1 ];

step 404: traversing alpha according to the search step _i E, alpha, I is the index of the user, which satisfies that 1 is less than or equal to I, and for the variable alpha _i Performing a round of monte carlo simulation to search for application QoS;

step 405: variable alpha for first meeting constraint of loss rate of expiration date by function scheduling scheme _i The value is assigned as the value of the random attribute executed by the application.

In the architecture model used in the invention, a common server-free edge network model is considered, and the topological structure is an undirected connected graphThe architecture model includes three main modules: an energy harvesting module, an energy buffering module, and an energy consuming module. The mth edge server in the model is composed of S _m Expressed, its maximum computing power is expressed as ψ _m And assume edge server S _m Support a set of containers->Assigned to the container->Is defined by->Indicating (I)>Not pre-specified, but one of the optimization variables in the present invention. Edge server S _m Can be through virtual link->With partner S _n (1.ltoreq.n.ltoreq.M, m.noteq.n) for communication. With virtual link l _m,n The associated bandwidth is defined by a constant b _m,n And (3) representing. b _m,n ＝b _n,m This holds true and the data communication time generated by two containers on the same edge server is negligible.

In the application model used in the invention, real-time Internet of things applicationWill execute on the target serverless edge computing network. Every application->Are all by unique user U _i Submissions, typically modeled as directed acyclic graphs +.>Because the number of instruction cycles applied fluctuates with different inputs, a group of execution adaptation variables alpha=α1, alpha 2, …, alpha I are introduced, and the value range is [0,1 ]]To capture the uncertainty of the number of instruction cycles.

In the energy model used in the invention, the edge server S _m The power collected at time t is expressed asThe renewable energy source energy within the specific time range delta t is +.>The energy harvesting and buffering module may both provide energy supply for application execution. />Buffering energy ordered in buffering module, edge server S _m At time interval [ t, t+Δt ]]Internal integral energy supply usage-> And (5) estimating. Edge server S _m And also depends on the number of functions allocated to these containers. When the function f _i,p In allocation with computing resources->Is>When executing on, the energy requirement generated by the function execution is expressed as Zeta is static power _m For effective switching capacitance, +.>Is the supply voltage. Let u _m,n For link l _m,n Unit energy dissipation of upper communication data transmission, then the slave function f _i,p The energy overhead of transmitting data to all immediate successor is expressed as: />/>Introducing a binary variable lambda _m,i,q To indicate the function f _i,q Whether or not to be dispatched to the edge server S _m . If yes, Λ _m,i,q Setting 1 or setting 0; introducing a binary variable +.>Only when f _i,p In the container->The 1 is taken when executing. Finally, the edge server S is obtained _m The overall energy requirements of (a) are:

in the reliability model used in the invention, two common fault mechanisms are concerned: one is transient faults at the edge server side in the function execution process, and the other is bit errors at the virtual link side in the data communication process. Taking into account inter-function data dependencies, function f _i,p In the containerThe execution reliability is expressed as +.> Besides soft errors, errors can also occur on the virtual link side during data communication, setting +.>Representing link l _m,n Bit error rate at the location, then deployed at edge servicesDevice S _m Function f of above _i,p With its single direct precursorThe communication reliability formula between the two is +.>Since all communication data should be +_ from the parent set>Is successfully transferred to function f _i,p Thus the communication reliability is defined by +.> And (5) calculating. Container->Upper function f _i,p The reliability calculation and the communication reliability in the equations can be performed by combining the equation sets, and the final reliability model is:

in the QoS model used in the invention, an elastic payment model is designed.Representing application +.>Service payment of->And use->The relationship between the expressed response delays is:

is a time measure that measures the degree of violation of the expected response delay. To characterize the QoS of an application, the latent variable +.>Defined as->Wherein eta _u ∈[0,1]For preference factor, ++>And->For two parameters, the application program is calculated at this time +.>QoSQ of (2) _i The method comprises the following steps:

/>

the invention aims to jointly optimize profit of a service provider and overall QoS of an application of the Internet of things, so that the whole network can adapt to a function scheduling solution to meet different design requirements on service profit and overall service level. The service profit function of the service provider is thatThe overall QoS function of the application of the Internet of things is that

The objective function of the present invention can be expressed as:

maximization: gamma ray _profit &Q _fair

Constraint conditions:

wherein, the liquid crystal display device comprises a liquid crystal display device,for application program->Is satisfied, < +.>Is to suggest that the application is satisfied->Data dependent constraint of any two functions, < +. >Is promised to meet the server S _m Is used for the energy limitation of the (a),indicating that edge server S can be violated _m Maximum computational power of (a).

Examples:

the network topology of the hardware platform is from a total of 3233 distributed base stations from Shanghai telecom randomly selected 600 to build the test network structure. For each of these 600 base stations, it is assumed that one edge server has been placed at the same location. Table 1 lists the main parameters of the heterogeneous edge servers.

TABLE 1

The average failure rate corresponding to the maximum computing power of the edge server falls to [10 ] ^-8 ,10 ^-6 ]The sensitivity of the failure rate to the expansion of the computational power of the container is selected from the group consisting of [1,10 ]]The collected power of the edge server is divided into sections by Calculation of>Obeys a standard gaussian distribution-> Is a random variable. Assume that the bandwidth between two base stations falls within [500,5000 ]]Within KB/s and shared equally by all virtual links on both base stations, the average failure rate of the virtual links is assumed to be [10 ] ^-7 ,10 ^-5 ]Within the interval.

Application information is tested on the hardware platform for a function-oriented dataset tracked using an alembia cluster. The dataset recorded an archibar cluster track of over 300 ten thousand applications, including 20365 applications with unique structural information.

FIG. 2 samples a total of 2119 DAG applications of different depth and width sizes, with minimum and maximum execution cycle amounts scaled to [4×10 ] for each function in the DAG ³ ,5×10 ⁵ ]And [6×10 ] ⁶ ,7×10 ⁸ ]Is a constant value, and is a constant value. The traffic from the function to any directly following function thereof is scaled to [500,1500 ]]Megabyte intervals. The reliability goal of an application is from [0.7,0.9999 ]]Randomly selected from the intervals of (a). The required completion time and expiration date of the application are set as:

/>

and->Is a coefficient related to the number of execution functions. Set->And->The average computing and switching power of the estimated target network is 2000GHz and 2000KB/s, respectively. The upper limit of the application expiration date loss ratio is set to 10%. It is assumed that each DAG application is submitted by an individual user with a unique personality. By conducting a questionnaire, the personality scores of 700 participants and the application QoS scores at 36 paid QoS levels were collected. Leaving 1419 persons as a validation set to evaluate personality-based application QoS predictionsGeneralization ability of the model.

The above was set as an experimental environment.

Comparing the present invention with six other representative reference algorithms:

algorithm 1: B-NSGA-II is a basic NSGA-II version with traditional crossover and mutation manipulations. As previously mentioned, crossover and mutation probabilities are specified as constants, i.e. they are not allowed to change throughout the evolution process.

Algorithm 2: the c-DPEA is a novel evolutionary algorithm, and a double population is adopted to balance solution diversity and algorithm convergence. In c-DPEA, two complementary populations are generated and simultaneously participate in the evolution process.

Algorithm 3: DLS-MOEA is a multi-objective evolutionary algorithm with dual local searches. One significant improvement is to perform a double local search in the target and decision space. An archive-based non-inferior solution generator was developed based on novel dual local search mechanisms to force population evolution towards pareto boundaries.

Algorithm 4: MOALO is a multi-objective ant lion optimizer that incorporates popular roulette methods to direct promising ants (i.e., solutions) toward the ideal search area of the multi-objective optimization problem.

Algorithm 5: MOEA/D is a decomposition-based multi-objective evolutionary algorithm. In MOEA/D the initial problem is divided into multiple scalar optimization sub-problems that can be solved simultaneously to reduce computational complexity.

Algorithm 6: maOEA/IGD is a competitive index based on a multi-objective evolutionary algorithm. It uses inverse algebraic distance (IGD) as an index to select an advantageous solution to maintain better algorithm convergence and population diversity.

The executable program or pseudocode of the above reference algorithm and the appropriate parameters are all from its original study. The algorithm source code was rewritten in Java language for fair comparison and implemented on a machine configured with 64GB memory and an Intel XeonW-10855M6 core processor.

Fig. 3 (a) to 3 (f) depict six confusion matrices under different numbers of participants for the personality-driven application QoS prediction method of the present invention. Fig. 3 (a) shows a confusion matrix for predicting application QoS for the remaining 1919 users using data for 200 participants. Diagonal elements in the confusion matrix represent the accuracy of QoS predictions, while other elements represent the likelihood of outputting erroneous prediction results. When the actual application QoS is 5, the probability of outputting a correct prediction result is only 51.1%. However, when using 700 participants' data in the training process, the application QoS prediction method has a 94.3% likelihood of generating the correct prediction result for an actual application QoS of 5, as shown in fig. 3 (f). When the actual application QoS takes other values in fig. 3 (f), the prediction method of the present invention can maintain the estimation accuracy between 91.9% and 95.5%; in the rest, fig. 3 (b) is a predicted application QoS for 300 users, fig. 3 (c) is a predicted application QoS for 400 users, fig. 3 (d) is a predicted application QoS for 500 users, and fig. 3 (e) is a predicted application QoS for 600 users.

Fig. 4 investigates the total estimation error of the personality-driven application QoS prediction method of the present invention. 700 participants' data are sufficient to train the QoS prediction model of the present invention, and the overall prediction error can be reduced to 7.5%. Table 2 is used to evaluate deterministic function scheduling algorithms, listing the deterministic algorithm of the present invention and superscripts of six benchmark algorithms, where each data point is the average result of 100 individual algorithm runs. A larger supersvolume is preferred here because it suggests that the corresponding algorithm will produce a non-dominant solution that is closer to the pareto boundary. The present invention is superior to all benchmark algorithms, regardless of network structure variations, as shown in table 2. The enhanced NSGA-II method of the invention has average superscales of 8.3%, 8.2%, 15.0%, 14.8%, 7.3% and 10.4% greater than the baseline algorithms B-NSGA-II, c-DPEA, DLS-MOEA, MOALO, MOEA/D and MaOEA/IGD, respectively.

TABLE 2

The present example also compares the IGD values achieved based on the deterministic algorithm of enhanced NSGA-II with the six baseline algorithms B-NSGA-II, c-DPEA, DLS-MOEA, MOALO, MOEA/D and MaOEA/IGD. The metric of IGD is typically used to measure the average distance between a feasible solution and a set of points evenly distributed on the target problem pareto boundary. Unlike supersvolume, a smaller reverse generation distance value is desirable because it suggests that the corresponding algorithm is more likely to approach a set of points with a viable solution.

Table 3 is used to evaluate deterministic function scheduling algorithms, giving the average anti-generation distance and the improvement of the algorithm of the present invention by running each algorithm 100 times. The inverse generation distance is typically used to measure the average distance between the feasible solution and the set of points evenly distributed on the target problem pareto boundary. Unlike supersvolume, the smaller the anti-generation distance, the better the performance of the solution. As shown in table 3, the enhanced NSGAII-based deterministic algorithm outperforms all benchmark algorithms in terms of the anti-generation distance.

FIG. 5 shows the run times of the implementation based on the enhanced NSGA-II deterministic algorithm and the six baseline algorithms B-NSGA-II, c-DPEA, DLS-MOEA, MOALO, MOEA/D and MaOEA/IGD. Each data point is an average of 100 simulation experiments for different network structures. The algorithm of this experiment consumes less run time than the baseline algorithms MaOEA/IGD and MOEA/D. But since the algorithm of the present invention incorporates new crossover and mutation operators, these operators would incur higher time overhead to derive the appropriate crossover and mutation probabilities, more runtime (up to 49.3%) would be spent to arrive at the ideal proposed solution, compared to the other baseline algorithms B-NSGA-II, c-DPEA, DLS-MOEA and MOALO.

TABLE 3 Table 3

/>

Tables 4 and 5 are used to evaluate the random parallel function scheduling scheme, comparing the certainty of the present invention with the random function scheduling strategy. The set of comparison experiments uses three fixed execution adaptation variable settings α= { α for deterministic function scheduling algorithms ₁ ,α ₂ ,…,α _I α=0, α=0.5, α=1. In contrast, the random function scheduling scheme of the present invention allows each element α _i E alpha selects any value between 0 and 1. As shown in table 4, the random scheme of the present invention is superior to deterministic algorithms in terms of both service profit and overall application QoS, regardless of network structure variations. The average service profit and overall application QoS are increased by 78.6% and 49.6%, respectively.

TABLE 4 Table 4

/>

TABLE 5

FIG. 6 depicts runtime for various parallelization levels. The random parallel function scheduling scheme of the invention has great advantage in reducing time overhead. When parallelization is not used (i.e., level # 1), the random function scheduling scheme generates the highest time overhead to derive the ideal function schedule. On the other hand, as the level of parallelization increases, the resulting time overhead may decrease significantly. For example, the time overhead for parallelization level #6 is about six times shorter than the time overhead for no parallelization policy (i.e., level # 1).

From the above experimental data, it can be clearly seen that the present application has good performance in terms of reducing the running time and approaching a feasible solution.

The present application is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present application are intended to be included in the scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims (5)

1. The high-reliability function scheduling method for the server-free edge calculation is characterized by comprising the following steps of:

step 1: presetting a personality-driven application program QoS prediction method, and predicting QoS of a single application program under different user personality types based on the application program QoS prediction method;

step 2: constructing a deterministic function scheduling model, and based on the deterministic function scheduling model, accurately estimating the QoS predicted in the step 1;

step 3: constructing a random function scheduling strategy, and scheduling the QoS after accurate estimation based on the random function scheduling strategy;

Step 4: and performing parallel function scheduling optimization, wherein the parallel function scheduling optimization is used for adapting to a random function scheduling strategy of a modern multi-core scheduler platform to accelerate generation of a solution.

2. The server-edge-oriented computing high reliability function scheduling method of claim 1, wherein predicting QoS of a single application under different user personality types based on the application QoS prediction method comprises:

step 101: initializing a set of iteration markersWherein I is _par Representing the number of users participating in the questionnaire, which is a constant less than the user population I;

step 102: initializing preference factors for participating questionnaire users

Step 103: judging whether all iteration marks are zero, if yes, turning to step 111, otherwise, turning to step 104;

step 104: calculation ofWherein j is more than or equal to 1 and I is more than or equal to 1 _par J is the index of the questionnaire participant, L represents the number of all possible delay-payment level combinations, θ _j，l And Q _j，l Respectively represent the questionnaire participants U at the first delayed payment level _j And application QoS, θ ₁ And theta ₂ Are all preference value parameters in the linear regression problem;

step 105: iteratively deriving preference factors for all participants in the questionnaire using a linear regression solver;

step 106: computing participant U _j Preference factor of (a)The formula is:

whereinAnd->Respectively represent the participants U in the first delay-payment level combination _j Response delay and service payment fee;

step 107: comparing randomly generated preference factors eta _j And derived preference factorsAbsolute difference ρ between;

step 108: judging whether the absolute difference value rho is smaller than a preset value, if so, turning to step 109, otherwise, turning to step 110;

step 109: participant U _j Iteration flag of (a)Set to zero, i.e.)>

Step 110: matrix updating using Latin hypercube sampling technique

Step 111: let S _j ＝[AScore _j ，CScore _j ，EScore _j ，OScore _j ，NScore _j ]Representing participant U _j Is described, wherein the user personality vector is decomposed into five different features: humanization, responsibility, exogenousness, openness and neuro-properties, corresponding to AScore _j 、CScore _j 、EScore _j 、OScore _j And NScore _j ；

Step 112: calculating a linear mapping from personality score to user preference, the formula:wherein-> The representation dimension is I _par X1 preference factor matrix, X _5×1 Is a mapping matrix with 5 multiplied by 1 dimension, and shows the influence of personality score on user preference;

step 113: for the user group with the total number of I, traversing the squareConstruction of any one user U _i Is the personality score vector S of (1) _i Wherein I (1.ltoreq.i.ltoreq.i) is an index of the user;

Step 114: computing a matrixStored user U _i Preference factor of->The formula is: /> An optimal preference factor matrix with a total number of rows of I and a total number of columns of 5, wherein I is the total number of users, < >>Personality score matrix with total number of rows of I and total number of columns of 5>An influence matrix of personality scores with total number of rows of I and total number of columns of 5 on user preference is represented;

step 115: calculating input delay-payment combinationsDown prediction application +.>QoS Q of (C) _i，l The formula is: />Wherein->For preference value parameter θ in step 104 ₁ Is>For preference value parameter θ in step 104 ₂ Is>For application programsPreference factor of->Representing user U _i Response delay at the first class of service, +.>Representing user U _i Payment at the first service level;

step 116: the QoS of the application is output.

3. The server-edge-oriented computing high reliability function scheduling method of claim 1, wherein accurately estimating the QoS predicted in step 1 based on the deterministic function scheduling model comprises:

step 201: initializing an iteration counter g, namely g+.0;

step 202: selecting chromosomes with less constraint violations to initialize populationsWherein U is the total number of users;

Step 203: initializing operation marks of three tables, namely a function sequence, a function-to-container scheduling and a resource-to-container allocation, namely a flag of (ζ) ζ 1;

step 204: judging iteration counter g < g _max Whether or not to establishIf so, go to step B5 to enter the population evolution process, otherwise go to step 229, where g _max Is the upper bound of the iteration counter;

step 205: calling a Modulo function module (flag++, 3) to obtain the remainder of dividing an operation mark flag by 3, and if the result is 1, performing evolutionary operation on a function sequence table; otherwise, carrying out evolutionary operation on the allocation table from the function to the container and the allocation table from the resource to the container;

step 206: based on Gaussian distributionGenerating a set of cut-off samples->Wherein x is a cut-off sample index within the range of x being more than or equal to 1 and U, and the average value is +.>Variance is->

Step 207: indexing two chromosomesAnd z is set to the values of 1 and U, respectively, i.e. +.>

Step 208: judgingIf yes, go to step 215, otherwise go to step 209;

step 209: judging whether flag=1 is satisfied, if so, indexingAnd->Respectively assigned to the cut-off sample index x ₁ And x ₂ I.e. x ₁ ←u，x ₂ ζ; otherwise, index ++>Simultaneous assignment of value to cut-off sample index x ₁ And x ₂ I.e. x ₁ ←u，x ₂ ≡u, indicating that the same cut-off sample was selected;

step 210: using split functionsChromosome->Contains element list->According to the truncated sample->Divided into->And->Two parts, i.e.)>

Step 211: using split functionsChromosome->Contains element list->According to the truncated sample->Divided into->And->Two parts, i.e.)>

Step 212: using a join functionUpdate element->I.e.

Step 213: using a join functionUpdate element->I.e.

Step 214: updating chromosome index for crossover operations of other chromosomesTurning to step 208;

step 215: computing gaussian distribution of crossing samples of next generation populationThe formula isWherein K is the total number of observation intervals, H is the total number of position samples, +.>Weight for the h position sample, +.>Is the observation value of the kth interval of the sample at the h position, g is the algebra of population evolution,/and>for the mean value of the next iteration population evolution at flag, < >>Is the variance;

step 216: if it isIf yes, the cross operation is completed, and the step 217 is executed, otherwise, the step 205 is executed;

step 217:starting to run the mutation operation, judgingIf yes, go to step 218 if true, otherwise go to step 224;

step 218: for each chromosome Calculating function order table +.>Resource-to-container allocation tableFunction to container schedule->Variation probability vector->And->The formula is: />Wherein->Is chromosome->Variation factor of the kth interval, +.>Is chromosome->Coefficient of variation, g _th And g _max Sequentially the threshold value and the maximum value of the iteration counter g, phi ₄ Is constant (I)>Is used for +.>Scaled to [0,1 ]]Is a normalized function of->Is the mutation probability;

step 219: random generation falls within [0,1 ]]Variation probability threshold sigma of (2) ₁ 、σ ₂ Sum sigma ₃ I.e. sigma ₁ ，σ ₂ ，σ ₃ ←Random(0，1)；

Step 220: for element tableEach position +.>If the mutation probability->Greater than threshold sigma ₁ Variant function->Applied to position->

Step 221: for element tableIs +.>If the mutation probability->Greater than threshold sigma ₂ Then the variation function will be scheduled->Applied to position->

Step 222: for element tableEach position +.>If the mutation probability->Greater than threshold sigma ₃ Then assign the variation function->Applied to position->

Step 223: once the current population evolution is completed, entering a competition selection stage;

step 224: deriving QoS for the application and calculating profit gamma for the service provider according to the method described in step 1 _profit ；

Step 225: arranging all chromosomes in a descending order by adopting a non-dominant ordering genetic algorithm;

Step 226: calculation of chromosomesTotal constraint violation degree Φ _u The formula is: />Wherein->To the extent that only the energy constraint is violated, +.>To the extent that only reliability constraints are violated;

step 227: calling population generation functionsTo construct the next generation population +.>The ratio-changing method is utilized to enable the proportion of infeasible individuals remained in the population to be linearly reduced along with the iteration times of the population;

step 228: updating the iteration counter g, namely: g≡g+1, go to step 204;

step 229: after the population iteration of preset times, a non-dominant solution set is outputAnd then terminates.

4. The server-edge-oriented computing high reliability function scheduling method of claim 1, wherein scheduling the QoS after accurate estimation based on the random function scheduling policy comprises:

step 301: assuming that the total number of users is I, the execution adaptation variable of the I users is alpha= { alpha ₁ ，α ₂ ，…，α _I Iterative generation of the execution adaptation variable α= { α ₁ ，α ₂ ，…，α _I Deriving a deterministic function schedule that limits the deadline miss rate for all applications to within design requirements, a problem-specific enhancement NSGA-II for use in executing the adaptation variable α= { α ₁ ，α ₂ ，…，α _I Generating a deterministic schedule at a particular value;

step 302: performing adaptive variable alpha= { alpha for simulating and evaluating user population as I in terms of application deadline deletion rate by using Monte Carlo ₁ ，α ₂ ，…，α _I }；

Step 303: evaluating performance of the current function schedule using a time metric of deadline miss rate;

step 304: judging whether the estimated expiration date deletion rate is greater than a threshold value, if so, turning to step 305; otherwise, go to step 306;

step 305: updating the execution adaptation variable α= { α of I ₁ ，α ₂ ，…，α _I Step 301; wherein I is the total number of users;

step 306: the current function schedule and associated execution adaptation variables are thus a priority scheduling solution and the entire iterative process exits.

5. The server-less edge computation oriented high reliability function scheduling method of claim 1, wherein the parallel function scheduling optimization comprises:

step 401: creating an execution adaptation variable set α= { α ₁ ，α ₂ ，…，α _I -I is the total number of users;

step 402: will execute the adaptation variable alpha ₁ ，α ₂ ，…，α _I Assigned 0, i.e. alpha ₁ ，α ₂ ，…，α _I The number of users is 0,I;

step 403: setting the search step size to a fixed value in the interval [0,1 ];

step 404: traversing alpha according to the search step _i E alpha, i is the index of the userSatisfies 1.ltoreq.i.ltoreq.I for the variable alpha _i Performing a round of monte carlo simulation to search for application QoS;

step 405: variable alpha for first meeting constraint of loss rate of expiration date by function scheduling scheme _i The value is assigned as the value of the random attribute executed by the application.