CN109960573A - A kind of cross-domain calculating task dispatching method and system based on Intellisense - Google Patents
A kind of cross-domain calculating task dispatching method and system based on Intellisense Download PDFInfo
- Publication number
- CN109960573A CN109960573A CN201811643211.1A CN201811643211A CN109960573A CN 109960573 A CN109960573 A CN 109960573A CN 201811643211 A CN201811643211 A CN 201811643211A CN 109960573 A CN109960573 A CN 109960573A
- Authority
- CN
- China
- Prior art keywords
- domain
- decision
- time
- data
- task
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/48—Program initiating; Program switching, e.g. by interrupt
- G06F9/4806—Task transfer initiation or dispatching
- G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5011—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resources being hardware resources other than CPUs, Servers and Terminals
- G06F9/5016—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resources being hardware resources other than CPUs, Servers and Terminals the resource being the memory
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5066—Algorithms for mapping a plurality of inter-dependent sub-tasks onto a plurality of physical CPUs
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Debugging And Monitoring (AREA)
Abstract
The present invention proposes a kind of cross-domain calculating task dispatching method and system based on Intellisense, comprising: step 1 trains decision-tree model based on label data;Step 2, the execution time based on relative time complexity estimation calculating task;Step 3, the change in resources trend indicator that each domain is predicted based on resource historical record and ARIMA algorithm;Step 4, the resource real-time status index that each domain is obtained using resource status interface;Step 5, the transit time that each domain is moved to based on available bandwidth estimated data;Step 6 is based on decision-tree model and overall target decision task optimal execution domain.The present invention, in cross-domain calculating task scheduling scenario, avoids task resource and seizes phenomenon, solve the problems, such as that scheduling decision accuracy is low creatively by trend prediction algorithm and decision Tree algorithms integrated application;By streaming machine learning techniques, the performance issue of trend prediction algorithm and decision Tree algorithms is overcome, cross-domain calculating task scheduling overall time is greatly shortened.
Description
Technical field
The invention belongs to task schedule fields, and in particular to cluster grade task scheduling scenario, especially one kind is towards cross-domain
Calculate the calculating task dispatching technique of environment.
Background technique
Cross-domain calculating environment is made of multiple domains being mutually isolated, and each domain includes one or more complete storages and meter
Cluster is calculated, specific calculating task can be independently executed.Domain where participating in the key data calculated is known as data field.In cross-domain meter
It calculates in environment, it is not optimal scheduling strategy that calculating task, which is always submitted to data field execution,.When data field surplus resources
When insufficient, task will enter waiting list, cause the task start time uncontrollable.When data field surplus resources anxiety, task
Calculated performance will be affected, and cause task execution time elongated.It is excessively high when calculating task is committed to other low-load domains
Data cross-domain moving costs also results in the task start time and substantially postpones.Therefore, it is necessary to a kind of task schedules of overall importance
Technology, on the basis of comprehensively considering the influence factors such as each domain resource situation and data moving costs, intelligent decision task is most
Excellent execution domain.
Domain, which is able to carry out specific calculating task, must satisfy two premises: 1) domain must satisfy calculating task to CPU
With the demand of the resources such as memory;2) domain must store the data for participating in calculating, and be needed when necessary by data from other domain migrations
To this domain.The size of migrating data directly affects the length of cross-domain transit time, and file data can be by summarizing each fragment text
The size of part obtains its data volume, and database data can estimate its data by calculating tables of data width and recording the product of number
Amount.
Summary of the invention
Present invention aim to address the resource utilization in each domain under cross-domain calculating environment is unbalanced and domain is domestic-investment
The problem of source deficiency leads to task execution failure or executes overlong time is maintaining the resource utilization in each domain opposite to realize
On the basis of equilibrium shorten calculating task the overall execution time regulation goal, propose it is a kind of by Intellisense it is cross-domain based on
Method for scheduling task and system are calculated, for determining the optimal execution domain of specific calculating task.
In order to achieve the above objectives, the technical scheme of the present invention is realized as follows:
A kind of cross-domain calculating task dispatching method based on Intellisense, comprising:
Step 1 trains decision-tree model based on label data;
Step 2, the execution time based on relative time complexity estimation calculating task;
Step 3, the change in resources trend indicator that each domain is predicted based on resource historical record and ARIMA algorithm;
Step 4, the resource real-time status index that each domain is obtained using resource status interface;
Step 5, the transit time that each domain is moved to based on available bandwidth estimated data;
Step 6 is based on decision-tree model and overall target decision task optimal execution domain.
Further, decision-tree model described in step 1 show that steps are as follows by decision Tree algorithms training:
1.1, initial labels data are constructed, and are divided into training set and test set;
1.2, training set is input in decision tree training algorithm and is arranged training parameter, obtain decision-tree model;
1.3, decision-tree model and test set are input in decision tree assessment algorithm, show that decision-tree model assessment refers to
Mark;1.4, when decision-tree model evaluation index is unsatisfactory for requiring:
A) adjusting training parameter repeats step 1.2 and 1.3, until index is met the requirements;Alternatively,
B) adjustment label rule, repeats step 1.1,1.2 and 1.3, until index is met the requirements.
Further, evaluation method described in step 2 includes:
2.1, a kind of benchmark algorithm is chosen, and fits the time complexity curve of the benchmark algorithm;
2.2, the time complexity according to task to be evaluated relative to benchmark algorithm calculates the expected time T that goes out on missions.
Further, algorithm described in step 3 specifically includes:
3.1 obtain the resource historical data of this domain the past period;
3.2 calculate the resources data of this domain following a period of time using ARIMA algorithm;
3.3 obtain the expected time T of current time t0 and current task;
3.4 intercept the data in the section [t0, t0+T] from resources data, calculate variation tendency index;
3.5 each domains repeat 3.1~3.4 steps, respectively calculate the variation tendency index in this domain.
Further, resource real-time status index described in step 4 specifically includes following 5 indexs:
4.1 cluster CPU idleness, for describing the overall service condition of cluster CPU;
4.2 cluster core cpu sums, for describing the core total number of cluster CPU;
4.3 cluster free memories, the summation of the memory remaining space size for describing each node of cluster;
4.4 cluster disk remaining spaces, the summation of the disk remaining space size for describing each node of cluster;
4.5 cross-domain network availability bandwidths, for describing the service condition of the network bandwidth between two clusters.
A kind of another aspect of the present invention, it is also proposed that cross-domain calculating task scheduling system based on Intellisense, comprising:
Model training module, based on label data training decision-tree model;
Task execution time evaluator, the execution time based on relative time complexity estimation calculating task;
Change in resources trend prediction device predicts the change in resources trend in each domain based on resource historical record and ARIMA algorithm
Index;
Resource real-time indicators collector obtains the resource real-time status index in each domain using resource status interface;
Data Migration time evaluator, the transit time in each domain is moved to based on available bandwidth estimated data;
Task optimal execution domain decision-making device is based on decision-tree model and overall target decision task optimal execution domain.
Further, model training module includes:
Initial data unit for constructing initial labels data, and is divided into training set and test set;
Model unit obtains decision tree for being input in decision tree training algorithm and being arranged training parameter for training set
Model;
Index unit obtains decision tree mould for decision-tree model and test set to be input in decision tree assessment algorithm
Type evaluation index;
Adjustment unit reuses mould for the adjusting training parameter when decision-tree model evaluation index is unsatisfactory for requiring
Type unit and index unit, until index is met the requirements;Alternatively, adjustment label rule, reuse initial data unit,
Model unit and index unit, until index is met the requirements.
Further, task execution time evaluator includes:
Fitting unit chooses a kind of benchmark algorithm, and fits the time complexity curve of the benchmark algorithm;
Computing unit, the time complexity according to task to be evaluated relative to benchmark algorithm, calculating, which is gone out on missions, to be expected to execute
Time T.
Further, change in resources trend prediction implement body includes:
Data capture unit obtains the resource historical data of this domain the past period;
Data Computation Unit calculates the resources data of this domain following a period of time using ARIMA algorithm;
Time acquisition unit obtains the expected time T of current time t0 and current task;
Indicator calculating unit intercepts the data in the section [t0, t0+T] from resources data, calculates variation tendency and refers to
Mark;
Each domain uses said units, respectively calculates the variation tendency index in this domain.
Further, resource real-time indicators collector includes:
Cluster CPU idleness unit, for describing the overall service condition of cluster CPU;
The total counting unit of cluster core cpu, for describing the core total number of cluster CPU;
Cluster free memory unit, the summation of the memory remaining space size for describing each node of cluster;
Cluster disk remaining space unit, the summation of the disk remaining space size for describing each node of cluster;
Cross-domain network availability bandwidth unit, for describing the service condition of the network bandwidth between two clusters.
Compared with prior art, novelty of the invention is embodied in:
1) it is kept away creatively by trend prediction algorithm and decision Tree algorithms integrated application in cross-domain calculating task scheduling scenario
Task resource is exempted from and has seized phenomenon, has solved the problems, such as that scheduling decision accuracy is low;
2) by streaming machine learning techniques, the performance issue of trend prediction algorithm and decision Tree algorithms is overcome, substantially
Shorten cross-domain calculating task scheduling overall time.
Value dimension of the invention exists: 1) meeting the resource requirement of calculating task operation, it is ensured that each task is at runtime
There are enough resources;2) resource that may occur between calculating task is avoided to seize phenomenon;3) calculating task is dispatched to from master
The place for wanting data nearest executes;4) resource utilization is promoted, the loading condition between each domain is balanced;5) meet service application
The transparent demand using cross-domain resource.
The resource that the present invention can avoid to occur between calculating task seizes phenomenon.By introducing ARIMA algorithm, so that
It predicts that the change in resources trend in each domain is possibly realized, is seized so as to avoid current task and periodic task that resource occurs
Phenomenon.ARIMA is weighted and averaged the index value in period in past to obtain current index value, when week is presented in period in past index
When phase property, current period index also will appear similar quality.ARIMA algorithm is predicted according to the resource historical data in each domain
The following corresponding variation tendency.The periodicity of historical data is stronger, and the estimated execution period of predetermined period task is more accurate,
The probability for avoiding resource from seizing phenomenon is higher.
The invention can ensure that task optimal execution domain initial decision accuracy is greater than 80%, and it is allowed to run not with system
It is disconnected to improve.By introducing decision Tree algorithms, solves this kind of complicated scheduling problem of cross-domain calculating.The mistake to label for training set
Journey is substantially the cross-domain moving costs of adjustment, the Current resource service condition in each domain and Future variation tendency this tripartite
The weight relationship of face index, its significance lies in that: customizable initial labels training set, to customize initial decision model.It is fixed
The initial decision model of inhibition and generation improves initial decision accuracy to the full extent.As dispatching algorithm constantly executes, positive and negative use
Example is continuously replenished in training set, the continuous iteration of decision model, and decision accuracy will also be continuously improved.
The present invention can provide efficient scheduling performances.By introducing streaming machine learning techniques, by ARIMA algorithm and decision
Tree algorithm transform streaming machine learning task as, second grade response is realized, so that scheduling overall performance greatly improved.
Detailed description of the invention
Fig. 1 is the step schematic diagram of the embodiment of the present invention;
Fig. 2 is the resources curve graph of the embodiment of the present invention;
Fig. 3 is the scheduling implementation procedure schematic diagram of the embodiment of the present invention;
Fig. 4 is the streaming machine learning implementation procedure schematic diagram of the embodiment of the present invention.
Specific embodiment
It should be noted that in the absence of conflict, the feature in embodiment and embodiment in the present invention can phase
Mutually combination.
Technical solution of the present invention is described in further detail with reference to the accompanying drawing:
The present invention becomes in the Current resource service condition and Future for comprehensively considering data cross-domain moving costs, each domain
On the basis of change trend three aspect factor, the intelligent decision function in task optimal execution domain is realized, is mainly comprised the steps of (such as
Shown in Fig. 1):
Step 1 trains decision-tree model based on label data;
Step 2, the execution time based on relative time complexity estimation calculating task;
Step 3, the change in resources trend indicator that each domain is predicted based on resource historical record and ARIMA algorithm;
Step 4, the resource real-time status index that each domain is obtained using resource status interface;
Step 5, the transit time that each domain is moved to based on available bandwidth estimated data;
Step 6 is based on decision-tree model and overall target decision task optimal execution domain.
Wherein, step 1 is independently of other steps, once the decision-tree model met the requirements, step are generated by step 1
1 is no longer needed for executing.Hereafter, scheduling only carries out step 2 to step 6 every time, can obtain the unique of task optimal execution domain
Mark.
1, based on label data training decision-tree model
Decision-tree model is used to generate correct label, the i.e. unique identification in task optimal execution domain to decision data.Certainly
Plan tree-model show that steps are as follows by decision Tree algorithms training:
1) initial labels data are constructed, and are divided into training set and test set;
2) training set is input in decision tree training algorithm and is arranged training parameter, obtain decision-tree model;
3) decision-tree model and test set are input in decision tree assessment algorithm, obtain decision-tree model evaluation index;
4) when decision-tree model evaluation index is unsatisfactory for requiring:
A) adjusting training parameter repeats step 2 and 3, until index is met the requirements.Alternatively,
B) adjustment label rule, repeats steps 1 and 2 and 3, until index is met the requirements.
Label data includes two parts of label and feature, and label represents the unique identification in task optimal execution domain, feature
The every factor for influencing the result of decision is represented, including data cross-domain moving costs, the Current resource service condition in each domain and not
Carry out change in resources trend three aspect factor.The structure of label data is as follows:
(D, F)
Wherein, D represents label, and F represents feature and can further indicate that are as follows:
(Cd0, Rd0, Md0) ..., (Cdi, Rdi, Mdi) ..., (Cdn, Rdn, Mdn)
Wherein, C represents change in resources trend indicator, and R represents resource real-time status index, when M represents data cross-domain migration
Between.Di represents the i-th domain, and i ∈ [0, n], d0 represents data field.Therefore, Cd0Represent the change in resources trend indicator of data field, Rd0
The resource real-time status index of data field, Md0Represent time and M of the Data Migration to data fieldd0=0.Similarly, CdiRepresent
The change in resources trend indicator in the domain i, RdiRepresent the resource real-time status index in the i-th domain, MdiData Migration is represented to the i-th domain
Time.
Initial labels data acquisition system comprising n data can indicate are as follows:
(Di, Fi) | i ∈ [0, n]
0.7n data therein is randomly selected out as training set, remaining 0.3n data is as test set.
Trained decision-tree model f is used for the unique identification D according to feature F calculating task optimal execution domain, and f is indicated such as
Under:
D=f (F)
2, the execution time based on relative time complexity estimation calculating task
The execution time of calculating task is difficult directly to calculate, therefore proposes here a kind of based on relative time complexity
Indirect evaluation method.Firstly, choosing a kind of benchmark algorithm, such as Terasort, and the time for fitting the benchmark algorithm is multiple
Miscellaneous line of writing music.Then, the time complexity according to task to be evaluated relative to benchmark algorithm, calculating are gone out on missions the expected time
T, calculation formula are as follows:
T=Cslog (s) r+B
Wherein, behalf participates in calculating the size of data, and r represents the relative time complexity of task, and C represents benchmark algorithm
Coefficient constant, B represent set time consumption constant.S and r is passed to as variable by external callers.Relative time complexity r can
To be determined and under different data amount sample point according to the relationship of the execution time and the execution time of benchmark algorithm of task.
3, the change in resources trend indicator in each domain is predicted based on resource historical record and ARIMA algorithm
If the variation of resource history curve shows apparent periodicity, show that the domain has periodically long task.
After obtaining resource historical data by historical record access interface, the resource by ARIMA algorithm predictable domain future out becomes
Change curve.When harmonic compoment variation is presented in resource history curve, resources curve can equally show the similar period
Property, therefore the execution period of future period task can be predicted, so that current task and periodic task be avoided to occur
Resource seizes phenomenon.The periodicity of historical data is stronger, and the estimated execution period of periodic task is predictably more accurate, avoids
The probability that resource seizes phenomenon generation is higher.
Specific algorithm is as follows:
1) the resource historical data of this domain the past period (for example, one week in the past) is obtained;
2) the resources data of this domain following a period of time (for example, one day following) are calculated using ARIMA algorithm;
3) the expected time T of current time t0 and current task are obtained;
4) data that the section [t0, t0+T] is intercepted from resources data, calculate variation tendency index;
5) each domain repeats 1~4 step, respectively calculates the variation tendency index in this domain.
The variation tendency index in each domain becomes this after summarizing and waits for a part of decision data.Assuming that the resource in some domain
Although prediction curve is as shown in Fig. 2, the domain current residual resource is more sufficient, it is contemplated that resource can quilt in the following T time
Periodic task occupies, therefore the domain may not be the optimal execution domain of current task.
4, the resource real-time status index in each domain is obtained using resource status interface
Resource real-time status index is used to assess the resource real-time status in domain, and the resource real-time status index in each domain summarizes
Become this afterwards and waits for a part of decision data.Resource real-time status index specifically includes following 5 indexs:
1) cluster CPU idleness.For describing the overall service condition of cluster CPU;
2) cluster core cpu sum.For describing the core total number of cluster CPU;
3) cluster free memory.For describing the summation of the memory remaining space size of each node of cluster;
4) cluster disk remaining space.For describing the summation of the disk remaining space size of each node of cluster;
5) cross-domain network availability bandwidth.For describing the service condition of the network bandwidth between two clusters.
5, based on available bandwidth estimate Data Migration to each domain transit time
Available bandwidth is insufficient between two domains, data cross-domain transit time is directly affected, to influence task optimal execution
The final decision result in domain.The transit time in Data Migration to each domain becomes this after summarizing and waits for a part of decision data.
Mainly the transmission time by data between domain and file disk IO time are constituted data cross-domain transit time M.Calculation formula is as follows:
Wherein, the size of behalf migrating data, n represent available bandwidth, and C represents network performance wave constant, and I, which is represented, to be divided
Cloth file system write performance constant, O represent distributed file system reading performance constant.S and n is passed as variable by caller
Enter.
6, decision-tree model and overall target decision task optimal execution domain are based on
Before acquisition after the indices of step, according to the decision-tree model of step 1 training, decision is gone out on missions optimal hold
The unique identification in row domain.
To decision data by the change in resources trend indicator (step 3) in each domain, the resource real-time status index in each domain
(step 4) and data move to the transit time (step 5) composition in each domain.Indices are recombinated as unit of domain,
Ultimately form feature F, it may be assumed that
(Cd0, Rd0, Md0) ..., (Cdi, Rdi, Mdi) ..., (Cdn, Rdn, Mdn)
Then according to the unique identification D in decision-tree model f calculating task optimal execution domain, it may be assumed that
D=f (F).
System of the present invention is cooperated by following 5 components and is completed:
1) task execution time evaluator: it is responsible for the execution time of assessment calculating task;
2) Data Migration time evaluator: it is responsible for assessment data cross-domain transit time;
3) resource real-time indicators collector: it is responsible for the resource real-time status index in each domain of acquisition;
4) change in resources trend prediction device: the change in resources trend for being responsible for each domain of quick predict following a period of time refers to
Mark;
5) task optimal execution domain decision-making device: it is responsible for the optimal execution domain of high-speed decision calculating task.
It is following (as shown in Figure 3) to dispatch implementation procedure:
1) task execution time evaluator, input data amount and relative time complexity are called, task execution time is obtained
T;
2) the change in resources trend prediction device in each domain is called, incoming task expected time T obtains each domain in T
Change in resources trend indicator C in timedi;
3) the resource real-time indicators collector for calling each domain, obtains the resource real-time indicators R in each domaindi;
4) Data Migration time evaluator is called, the resource real-time indicators R in each domain is inputteddiAnd data volume, obtain data
Move to the transit time M in each domaindi;
5) task optimal execution domain decision-making device is called, the change in resources trend indicator C according to each domain is inputteddi, each domain
Resource real-time indicators RdiThe transit time M in each domain is moved to datadiAfter integration to decision data (Cd0, Rd0,
Md0) ..., (Cdi, Rdi, Mdi) ..., (Cdn, Rdn, Mdn), obtain the unique identification D in task optimal execution domain.
Here a specific embodiment of the invention is illustrated with a specific example.(it is assumed to be the domain A with the environment in two domains
With the domain B) for, the cross-domain calculating task dispatching technique implementation process based on Intellisense is as follows:
1) task execution time evaluator is called, task expected time T is obtained;
2) the change in resources trend prediction device for calling the domain A obtains change in resources of the domain A in task expected time T
Trend indicator Ca;
3) the change in resources trend prediction device for calling the domain B obtains change in resources of the domain B in task expected time T
Trend indicator Cb;
4) the resource real-time indicators collector for calling the domain A, obtains the resource real-time status index R in the domain Aa;
5) the resource real-time indicators collector for calling the domain B, obtains the resource real-time status index R in the domain Bb;
6) Data Migration time evaluator is called, data cross-domain transit time M is obtainedaAnd Mb;
7) summarize and recombinate all indexs, obtain to decision data (Ca, Ra, Ma), (Cb, Rb, Mb), then input optimal hold
Row domain decision-making device, it is final to obtain task optimal execution domain unique identification D.
Wherein, change in resources trend prediction device and task optimal execution domain decision-making device all use streaming machine learning techniques into
It has gone performance optimization, has realized second grade response.Either ARIMA algorithm or decision Tree algorithms are held as off-line calculation task
The capable response time is minute rank, seriously affects scheduling overall performance.It transform ARIMA algorithm and decision Tree algorithms as streaming
Machine learning task simultaneously executes on backstage, not only saves the starting time of algorithm, and realize to the quick of single request
Response.
Below by taking the decision-making device of task optimal execution domain as an example, illustrate the implementation procedure (as shown in Figure 4) of streaming machine learning:
1) task optimal execution domain decision making algorithm (streaming machine learning task) start and load decision-tree model, then into
Enter suspended state;
2) request of task optimal execution domain decision interface caller and it is forwarded to request queue;
3) task optimal execution domain decision making algorithm is activated, and acquisition request is simultaneously calculated according to decision-tree model, and will
As a result it is sent to response queue, then proceedes to suspended state;
4) task optimal execution domain decision interface obtains the result of decision from response queue and returns to caller, completes this
Decision process, and step 2 is constantly repeated to step 4;
When sending terminates order, task optimal execution domain decision making algorithm is waken up, and terminates streaming machine after discharging resource
Learning tasks.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of cross-domain calculating task dispatching method based on Intellisense characterized by comprising
Step 1 trains decision-tree model based on label data;
Step 2, the execution time based on relative time complexity estimation calculating task;
Step 3, the change in resources trend indicator that each domain is predicted based on resource historical record and ARIMA algorithm;
Step 4, the resource real-time status index that each domain is obtained using resource status interface;
Step 5, the transit time that each domain is moved to based on available bandwidth estimated data;
Step 6 is based on decision-tree model and overall target decision task optimal execution domain.
2. a kind of cross-domain calculating task dispatching method based on Intellisense according to claim 1, which is characterized in that step
Rapid 1 decision-tree model show that steps are as follows by decision Tree algorithms training:
1.1, initial labels data are constructed, and are divided into training set and test set;
1.2, training set is input in decision tree training algorithm and is arranged training parameter, obtain decision-tree model;
1.3, decision-tree model and test set are input in decision tree assessment algorithm, obtain decision-tree model evaluation index;
1.4, when decision-tree model evaluation index is unsatisfactory for requiring:
A) adjusting training parameter repeats step 1.2 and 1.3, until index is met the requirements;Alternatively,
B) adjustment label rule, repeats step 1.1,1.2 and 1.3, until index is met the requirements.
3. a kind of cross-domain calculating task dispatching method based on Intellisense according to claim 1, which is characterized in that step
Rapid 2 evaluation method includes:
2.1, a kind of benchmark algorithm is chosen, and fits the time complexity curve of the benchmark algorithm;
2.2, the time complexity according to task to be evaluated relative to benchmark algorithm calculates the expected time T that goes out on missions.
4. a kind of cross-domain calculating task dispatching method based on Intellisense according to claim 1, which is characterized in that step
Rapid 3 algorithm specifically includes:
3.1 obtain the resource historical data of this domain the past period;
3.2 calculate the resources data of this domain following a period of time using ARIMA algorithm;
3.3 obtain the expected time T of current time t0 and current task;
3.4 intercept the data in the section [t0, t0+T] from resources data, calculate variation tendency index;
3.5 each domains repeat 3.1~3.4 steps, respectively calculate the variation tendency index in this domain.
5. a kind of cross-domain calculating task dispatching method based on Intellisense according to claim 1, which is characterized in that step
The rapid 4 resource real-time status index specifically includes following 5 indexs:
4.1 cluster CPU idleness, for describing the overall service condition of cluster CPU;
4.2 cluster core cpu sums, for describing the core total number of cluster CPU;
4.3 cluster free memories, the summation of the memory remaining space size for describing each node of cluster;
4.4 cluster disk remaining spaces, the summation of the disk remaining space size for describing each node of cluster;
4.5 cross-domain network availability bandwidths, for describing the service condition of the network bandwidth between two clusters.
6. a kind of cross-domain calculating task based on Intellisense dispatches system characterized by comprising
Model training module, based on label data training decision-tree model;
Task execution time evaluator, the execution time based on relative time complexity estimation calculating task;
Change in resources trend prediction device predicts the change in resources trend indicator in each domain based on resource historical record and ARIMA algorithm;
Resource real-time indicators collector obtains the resource real-time status index in each domain using resource status interface;
Data Migration time evaluator, the transit time in each domain is moved to based on available bandwidth estimated data;
Task optimal execution domain decision-making device is based on decision-tree model and overall target decision task optimal execution domain.
7. a kind of cross-domain calculating task based on Intellisense according to claim 6 dispatches system, which is characterized in that mould
Type training module includes:
Initial data unit for constructing initial labels data, and is divided into training set and test set;
Model unit obtains decision-tree model for being input in decision tree training algorithm and being arranged training parameter for training set;
Index unit show that decision-tree model is commented for decision-tree model and test set to be input in decision tree assessment algorithm
Estimate index;
Adjustment unit reuses model list for the adjusting training parameter when decision-tree model evaluation index is unsatisfactory for requiring
Member and index unit, until index is met the requirements;Alternatively, adjustment label rule, reuses initial data unit, model
Unit and index unit, until index is met the requirements.
8. a kind of cross-domain calculating task based on Intellisense according to claim 6 dispatches system, which is characterized in that appoint
Business executes time evaluator
Fitting unit chooses a kind of benchmark algorithm, and fits the time complexity curve of the benchmark algorithm;
Computing unit, the time complexity according to task to be evaluated relative to benchmark algorithm, calculating are gone out on missions the expected time
T。
9. a kind of cross-domain calculating task based on Intellisense according to claim 6 dispatches system, which is characterized in that money
Source trend implement body includes:
Data capture unit obtains the resource historical data of this domain the past period;
Data Computation Unit calculates the resources data of this domain following a period of time using ARIMA algorithm;
Time acquisition unit obtains the expected time T of current time t0 and current task;
Indicator calculating unit intercepts the data in the section [t0, t0+T] from resources data, calculates variation tendency index;
Each domain uses said units, respectively calculates the variation tendency index in this domain.
10. a kind of cross-domain calculating task based on Intellisense according to claim 6 dispatches system, which is characterized in that
Resource real-time indicators collector includes:
Cluster CPU idleness unit, for describing the overall service condition of cluster CPU;
The total counting unit of cluster core cpu, for describing the core total number of cluster CPU;
Cluster free memory unit, the summation of the memory remaining space size for describing each node of cluster;
Cluster disk remaining space unit, the summation of the disk remaining space size for describing each node of cluster;
Cross-domain network availability bandwidth unit, for describing the service condition of the network bandwidth between two clusters.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811643211.1A CN109960573B (en) | 2018-12-29 | 2018-12-29 | Cross-domain computing task scheduling method and system based on intelligent perception |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811643211.1A CN109960573B (en) | 2018-12-29 | 2018-12-29 | Cross-domain computing task scheduling method and system based on intelligent perception |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109960573A true CN109960573A (en) | 2019-07-02 |
CN109960573B CN109960573B (en) | 2021-01-08 |
Family
ID=67023414
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811643211.1A Active CN109960573B (en) | 2018-12-29 | 2018-12-29 | Cross-domain computing task scheduling method and system based on intelligent perception |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109960573B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110688207A (en) * | 2019-09-05 | 2020-01-14 | 烽火通信科技股份有限公司 | Embedded task scheduling method and system |
CN110705780A (en) * | 2019-09-27 | 2020-01-17 | 科大国创软件股份有限公司 | IT performance index prediction method based on intelligent algorithm |
CN111191113A (en) * | 2019-09-29 | 2020-05-22 | 西北大学 | Data resource demand prediction and adjustment method based on edge computing environment |
CN111930789A (en) * | 2020-09-21 | 2020-11-13 | 北京东方通软件有限公司 | Automatic scheduling method and device for middleware of database access layer |
CN112579273A (en) * | 2020-12-15 | 2021-03-30 | 京东数字科技控股股份有限公司 | Task scheduling method and device and computer readable storage medium |
CN114936086A (en) * | 2022-07-26 | 2022-08-23 | 之江实验室 | Task scheduler, task scheduling method and task scheduling device under multi-computing center scene |
CN116738239A (en) * | 2023-08-11 | 2023-09-12 | 浙江菜鸟供应链管理有限公司 | Model training method, resource scheduling method, device, system, equipment and medium |
Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101308468A (en) * | 2008-06-13 | 2008-11-19 | 南京邮电大学 | Grid calculation environment task cross-domain control method |
CN101604261A (en) * | 2009-07-08 | 2009-12-16 | 深圳先进技术研究院 | The method for scheduling task of supercomputer |
EP2742426A1 (en) * | 2011-09-29 | 2014-06-18 | NEC Laboratories America, Inc. | Network-aware coordination of virtual machine migrations in enterprise data centers and clouds |
CN104102533A (en) * | 2014-06-17 | 2014-10-15 | 华中科技大学 | Bandwidth aware based Hadoop scheduling method and system |
CN104239417A (en) * | 2014-08-19 | 2014-12-24 | 天津南大通用数据技术股份有限公司 | Dynamic adjustment method and dynamic adjustment device after data fragmentation in distributed database |
CN105205154A (en) * | 2015-09-24 | 2015-12-30 | 浙江宇视科技有限公司 | Data migration method and device |
CN105868222A (en) * | 2015-09-17 | 2016-08-17 | 乐视网信息技术(北京)股份有限公司 | Task scheduling method and device |
CN106022245A (en) * | 2016-05-16 | 2016-10-12 | 中国资源卫星应用中心 | Multi-source remote sensing satellite data parallel processing system and method based on algorithm classification |
CN108055119A (en) * | 2017-12-11 | 2018-05-18 | 北方工业大学 | Safe motivational techniques and system based on block chain in a kind of intelligent perception application |
CN108304250A (en) * | 2018-03-05 | 2018-07-20 | 北京百度网讯科技有限公司 | Method and apparatus for the node for determining operation machine learning task |
CN108509276A (en) * | 2018-03-30 | 2018-09-07 | 南京工业大学 | Video task dynamic migration method in edge computing environment |
CN108769182A (en) * | 2018-05-24 | 2018-11-06 | 国网上海市电力公司 | A kind of prediction executes the Combinatorial Optimization dispatching method of task execution time |
CN108900358A (en) * | 2018-08-01 | 2018-11-27 | 重庆邮电大学 | Virtual network function dynamic migration method based on deepness belief network resource requirement prediction |
-
2018
- 2018-12-29 CN CN201811643211.1A patent/CN109960573B/en active Active
Patent Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101308468A (en) * | 2008-06-13 | 2008-11-19 | 南京邮电大学 | Grid calculation environment task cross-domain control method |
CN101604261A (en) * | 2009-07-08 | 2009-12-16 | 深圳先进技术研究院 | The method for scheduling task of supercomputer |
EP2742426A1 (en) * | 2011-09-29 | 2014-06-18 | NEC Laboratories America, Inc. | Network-aware coordination of virtual machine migrations in enterprise data centers and clouds |
CN104102533A (en) * | 2014-06-17 | 2014-10-15 | 华中科技大学 | Bandwidth aware based Hadoop scheduling method and system |
CN104239417A (en) * | 2014-08-19 | 2014-12-24 | 天津南大通用数据技术股份有限公司 | Dynamic adjustment method and dynamic adjustment device after data fragmentation in distributed database |
CN105868222A (en) * | 2015-09-17 | 2016-08-17 | 乐视网信息技术(北京)股份有限公司 | Task scheduling method and device |
CN105205154A (en) * | 2015-09-24 | 2015-12-30 | 浙江宇视科技有限公司 | Data migration method and device |
CN106022245A (en) * | 2016-05-16 | 2016-10-12 | 中国资源卫星应用中心 | Multi-source remote sensing satellite data parallel processing system and method based on algorithm classification |
CN108055119A (en) * | 2017-12-11 | 2018-05-18 | 北方工业大学 | Safe motivational techniques and system based on block chain in a kind of intelligent perception application |
CN108304250A (en) * | 2018-03-05 | 2018-07-20 | 北京百度网讯科技有限公司 | Method and apparatus for the node for determining operation machine learning task |
CN108509276A (en) * | 2018-03-30 | 2018-09-07 | 南京工业大学 | Video task dynamic migration method in edge computing environment |
CN108769182A (en) * | 2018-05-24 | 2018-11-06 | 国网上海市电力公司 | A kind of prediction executes the Combinatorial Optimization dispatching method of task execution time |
CN108900358A (en) * | 2018-08-01 | 2018-11-27 | 重庆邮电大学 | Virtual network function dynamic migration method based on deepness belief network resource requirement prediction |
Non-Patent Citations (1)
Title |
---|
李国玉: "云平台多任务并发调度策略推荐***的设计与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110688207A (en) * | 2019-09-05 | 2020-01-14 | 烽火通信科技股份有限公司 | Embedded task scheduling method and system |
CN110688207B (en) * | 2019-09-05 | 2022-03-11 | 烽火通信科技股份有限公司 | Embedded task scheduling method and system |
CN110705780A (en) * | 2019-09-27 | 2020-01-17 | 科大国创软件股份有限公司 | IT performance index prediction method based on intelligent algorithm |
CN111191113A (en) * | 2019-09-29 | 2020-05-22 | 西北大学 | Data resource demand prediction and adjustment method based on edge computing environment |
CN111191113B (en) * | 2019-09-29 | 2024-01-23 | 西北大学 | Data resource demand prediction and adjustment method based on edge computing environment |
CN111930789A (en) * | 2020-09-21 | 2020-11-13 | 北京东方通软件有限公司 | Automatic scheduling method and device for middleware of database access layer |
CN112579273A (en) * | 2020-12-15 | 2021-03-30 | 京东数字科技控股股份有限公司 | Task scheduling method and device and computer readable storage medium |
CN112579273B (en) * | 2020-12-15 | 2023-05-30 | 京东科技控股股份有限公司 | Task scheduling method and device and computer readable storage medium |
CN114936086A (en) * | 2022-07-26 | 2022-08-23 | 之江实验室 | Task scheduler, task scheduling method and task scheduling device under multi-computing center scene |
CN116738239A (en) * | 2023-08-11 | 2023-09-12 | 浙江菜鸟供应链管理有限公司 | Model training method, resource scheduling method, device, system, equipment and medium |
CN116738239B (en) * | 2023-08-11 | 2023-11-24 | 浙江菜鸟供应链管理有限公司 | Model training method, resource scheduling method, device, system, equipment and medium |
Also Published As
Publication number | Publication date |
---|---|
CN109960573B (en) | 2021-01-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109960573A (en) | A kind of cross-domain calculating task dispatching method and system based on Intellisense | |
Bao et al. | Online job scheduling in distributed machine learning clusters | |
Liu et al. | A hierarchical framework of cloud resource allocation and power management using deep reinforcement learning | |
KR102562260B1 (en) | Commitment-aware scheduler | |
US10761897B2 (en) | Predictive model-based intelligent system for automatically scaling and managing provisioned computing resources | |
Zhao et al. | Independent tasks scheduling based on genetic algorithm in cloud computing | |
CN104572307B (en) | The method that a kind of pair of virtual resource carries out flexible scheduling | |
CN101216710A (en) | Self-adapting selection dynamic production scheduling control system accomplished through computer | |
CN109324875A (en) | A kind of data center server power managed and optimization method based on intensified learning | |
US11381483B2 (en) | Maintenance recommendation for containerized services | |
CN109960578A (en) | A kind of offline dispatching method of data center resource based on deeply study | |
CN102207883A (en) | Transaction scheduling method of heterogeneous distributed real-time system | |
CN109947532A (en) | A kind of big data method for scheduling task in education cloud platform | |
Hu et al. | FlowTime: Dynamic scheduling of deadline-aware workflows and ad-hoc jobs | |
CN111813524A (en) | Task execution method and device, electronic equipment and storage medium | |
CN115185650A (en) | Task scheduling method for heterogeneous edge computational power network | |
CN113535387A (en) | Heterogeneous sensing GPU resource allocation and scheduling method and system | |
CN106844175B (en) | A kind of cloud platform method for planning capacity based on machine learning | |
CN105117281B (en) | A kind of method for scheduling task of task based access control application signal and processor cores Executing Cost value | |
Zhang et al. | Monitoring-based task scheduling in large-scale SaaS cloud | |
Zhao et al. | RAS: a task scheduling algorithm based on resource attribute selection in a task scheduling framework | |
Subramanian et al. | Real time non-linear cloud workload forecasting using the holt-winter model | |
Lu et al. | Reinforcement learning-based auto-scaling algorithm for elastic cloud workflow service | |
Zheng et al. | DOSP: an optimal synchronization of parameter server for distributed machine learning | |
Blanco-Fernández et al. | The benefits of coordination in adaptive virtual teams |
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 |