CN104516784A - Method and system for forecasting task resource waiting time - Google Patents

Abstract

The invention discloses a method and system for predicting the waiting time of task resources. The invention relates to large-scale computing system resource management, optimization and allocation, and in particular to a method and system for predicting the waiting time of task resources. The method includes obtaining historical task records , delete the task records that have dependencies in the historical task record, and generate a new historical task record; use the autocorrelation function to obtain the task records in the new historical task record that have correlation with the predicted time period, and generate a task Record set; set the task resource waiting time threshold, obtain the number of task records whose task resource waiting time exceeds the task resource waiting time threshold in the task record set, and according to the total amount of task records in the task record set, through Bayeux The Adams method predicts the waiting time of task resources within the time period to be predicted. The invention can predict the availability of computing system resources and optimize task scheduling.

Description

Method and system for predicting task resource waiting time

technical field

The invention relates to resource management, optimization and allocation of large-scale computing systems (including supercomputing and cloud computing), and in particular to a method and system for predicting the waiting time of task resources.

Background technique

Many existing approaches use model-based approaches to characterize, predict, optimize, and allocate resource usage in large-scale computing systems. Specifically, researchers first select one or several dimensions related to system resource usage for data tracking observation (such as the remaining running time of jobs, the queuing time of jobs in system queues, etc.), and then apply some correlation A probability model is used to describe the probability distribution of such dimensional data. Next, researchers use the probability distribution properties presented by this model to predict the future performance of the system, so as to achieve resource optimization and reasonable allocation. For example, Downey applied the logarithm Uniform distribution (log uniform distribution) to measure the remaining running time of the job, Brevik and his collaborators proposed the concept of binomial method batch prediction (BinomialMethod Batch Predictor (BMBP)) to describe the waiting time of the system queue, Li and his collaborators applied mixed The Gaussian-mixture model is used to describe the running time distribution of all jobs in the system within a certain period of time.

The method of describing resource distribution based on a specific probability model will have the following practical problems in the practical application of large-scale computing system management and big data processing: the historical load records of massive data are difficult to obey a specific probability in practical applications In fact, by testing different actual data, the historical load records of big data not only do not obey a single probability distribution, but are even difficult to describe with a mixed probability distribution model; some commonly used probability models (such as the binomial distribution model ( binomialmodel) and its derivative models), the assumption that the properties of jobs submitted by users in a short time interval are independent of each other is often not true in reality. In fact, through specific research on the historical load data of actual supercomputers, we found that most users will Jobs with similar content and equivalent parameters are submitted multiple times in a short period of time, so it is often inaccurate to use this type of probability model to evaluate and predict resource consumption.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention proposes a method and system for predicting the waiting time of task resources.

The present invention proposes a method for predicting the waiting time of task resources, including:

Step 1, obtain historical task records, delete task records that have dependencies in the historical task records, and generate new historical task records;

Step 2, through the autocorrelation function, obtain the task records in the time period related to the time period to be predicted in the new historical task record, and generate a task record set;

Step 3: Set the task resource waiting time threshold, obtain the number of task records whose task resource waiting time in the task record set exceeds the task resource waiting time threshold, and according to the total amount of task records in the task record set, use Bayeux The Adams method predicts the waiting time of task resources within the time period to be predicted.

The method for predicting the waiting time of task resources, the specific steps of the step 1 include:

Step 11, judging whether there is a dependency relationship among the task records in the historical task record;

Step 12, delete task records with dependencies if they exist.

The method for predicting the waiting time of task resources, the step 11 includes:

Step 21, select time critical point t* and spatial critical point x*;

Step 22, if the submission time interval of two task records in the historical task record is within the time critical point t* and the proximity of parameter selection is within the spatial interval x*, and the pairing density is higher than the historical task record In addition to the density of pairings of the two task records, the two task records have a dependency within the scale (t*, x*).

The method for predicting the waiting time of task resources, the step 12 includes:

Step 31, arrange the task records in the historical task records according to the submission time from small to large;

Step 32, starting from the task record with the smallest submission time, deleting the task record with the smallest submission time to all task records whose parameter selection proximity to the task record with the smallest submission time within the next t* time is less than x*;

Step 33, update the historical task record and repeat the step 22 until traversing the historical task record and updating the historical task record.

The method for predicting the waiting time of task resources, the formula of the Bayesian method in the step 3 is:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Among them, N _{LTW, k-1, k-2} are the number of the task records whose waiting time of the task resource exceeds the threshold of the task resource waiting time, and N _{k-1, k-2} are the task records in the task record set The total amount of P _LTW,k is the task probability that the waiting time of the task resource exceeds the threshold of the waiting time of the task resource in the time period to be predicted.

The present invention also proposes a system for predicting the waiting time of task resources, including:

Generate a new historical task record module, which is used to obtain historical task records, delete dependent task records in the historical task records, and generate new historical task records;

Generating a task record set, which is used to obtain the task records in the time period that has correlation with the time period to be predicted in the new historical task record through an autocorrelation function, and generate a task record set;

The prediction module is used to set the task resource waiting time threshold, obtain the number of task records whose task resource waiting time exceeds the task resource waiting time threshold in the task record set, and according to the total amount of task records in the task record set, pass The Bayesian method predicts the waiting time of task resources within the time period to be predicted.

In the system for predicting the waiting time of task resources, the module for generating new historical task records also includes:

A judging module, judging whether there is a dependency relationship among the task records in the historical task records;

Delete the dependency module, if it exists, delete the task records with dependencies.

In the system for predicting the waiting time of task resources, the specific functions of the judging module include: selecting time critical point t* and space critical point x*; if the submission time interval of two task records in the historical task record is within the time critical point point t* and the proximity of parameter selection is within the space interval x*, and the pairing density is higher than the pairing density of the historical task records except the two task records, then the two task records are in (t *, x*) have dependencies within the scale.

In the system for predicting the waiting time of task resources, the specific function of the deletion dependency module includes: arranging the task records in the historical task records according to the submission time from small to large; starting from the task record with the smallest submission time, delete From the task record with the smallest submission time to the next t* time, the parameter selection of the task record with the smallest submission time is less than x* task records; update the historical task record and repeat step 22 until traversal For the historical task record, update the historical task record.

In the system for predicting the waiting time of task resources, the formula of the Bayesian method in the prediction module is:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Among them, N _{LTW, k-1, k-2} are the number of the task records whose waiting time of the task resource exceeds the threshold of the task resource waiting time, and N _{k-1, k-2} are the task records in the task record set The total amount of P _LTW,k is the task probability that the waiting time of the task resource exceeds the threshold of the waiting time of the task resource in the time period to be predicted.

As can be seen from the above scheme, the present invention has the advantages of:

The present invention can be used to predict the availability of computing system resources, and can be used to optimize job scheduling, optimize task resource demand configuration, improve resource utilization efficiency, and predict the waiting time of jobs in the job execution queue. The experiments of the present invention show that The invention can achieve a reliable prediction rate of more than 89%.

Description of drawings

Fig. 1 is the general flowchart of the present invention;

Fig. 2 is the operation partition count chart;

Fig. 3 is a long-time waiting probability chart after denoising and de-correlation;

Figure 4 is the LTW monthly trend graph on the supercomputer Kraken;

Fig. 5 is the monthly long-time waiting probability automatic correlation value chart;

FIG. 6 is a flow chart of noise reduction processing for historical task records.

Wherein reference sign is:

Step 100 is an overall step of the present invention, including:

Step 101/102/103;

Step 200 is a specific step of noise reduction in the present invention, including:

Step 201/202/203.

detailed description

In the process of studying the historical load of the supercomputer Kraken, the inventor found that the waiting time of about 52.4% of the jobs (tasks) submitted by users in the job execution queue exceeded the actual running time. For each job, especially parallel jobs, there are two basic parameters that express its demand for computing resources: running time and the number of computing nodes (CPU, memory). From the user's point of view, the present invention hopes to adjust these two parameters under the prerequisite of ensuring the correct result, so that the job can be run quickly and reduce the waiting time in the job execution queue. In the process of statistical analysis, a user submits multiple similar jobs (such as the same execution program, different input data, etc.) within a short period of time, which greatly affects the effective statistical results, because the system can simultaneously execute The number of jobs is generally limited. The present invention uses a statistical method: space-time clustering detection method (space-time clustering detection method) (also known as Knox method (Knox method)), to analyze the task dependencies in the system load history record, which is different from the usual clustering ( clustering) method, the method used in the present invention puts more emphasis on the relationship between parameter space and time, thereby making data reduction (reduction) more effective, removing task records with dependencies, and achieving noise removal, and the load history records after denoising can be used Generate accurate long-time waiting probability charts. This long-time waiting probability table (Long-Time Waiting Chart) can present the macro trends of different parameter value selections and long-time waiting probabilities, and then give users an overall guidance on parameter selection. . Furthermore, the present invention can simultaneously generate a long-time waiting probability trend graph in units of months. Since the usage rate of computing resources is dynamic, it is not that more historical records are more helpful for the calculation of waiting probability, and often redundant data There is the negative effect of concealing true trend, so the present invention uses autocorrelation function AutoCorrelation Function (ACF, autocorrelation function has different characterizations in different fields, and the present invention adopts autocovariance function to describe autocorrelation coefficient, and autocovariance function is description The second-order mixed central moment between the values of time series X(t) at any two different moments t1 and t2 is used to describe the correlation between the fluctuations (relative and mean) of the values of X(t) at two moments Degree, also known as centralized autocorrelation function) to determine the time interval area with correlation, such as finding that the data of 3 months on the test system has a correlation with each other, finally, the present invention uses Bayesian method (BayesianFramework) to predict The wait time for a future task (according to its requirements for system resources) to acquire resources and run on this computing system.

Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings.

The following are general steps of the present invention, as shown in Figure 1, the concrete steps are as follows:

Step 101, use the space-time clustering detection method (space-time clustering detection method) (also known as Knox method) to analyze the task dependencies in the system load history records (historical task records), and remove task records with dependencies, so as to remove Noise, for the purpose of improving prediction accuracy.

Step 102, generate a long-time waiting probability chart (Long-Time Waiting Chart, long-time waiting chart for task resources) according to the load history record (new historical task record) after the noise is removed, and generate a long-time waiting probability trend in units of months at the same time picture. In addition, the AutoCorrelation Function (ACF) was used to identify time interval regions with correlations.

Step 103, use the Bayesian method (Bayesian Framework) to predict the waiting time for future tasks (according to their requirements for system resources) to obtain resources and run on the computing system, wherein the task resource waiting time threshold is set, and the task record set is obtained The number of task records whose waiting time of the task resource exceeds the threshold of the task resource waiting time, and predict the waiting time of the task resource within a certain period of time according to the total number of task records in the task record set by Bayesian method.

The present invention is further described below in conjunction with embodiment:

Discover the real long-time waiting pattern: job partition count table: first, obtain the statistical results of the resource usage of the job in the form of a two-dimensional chart (for each submitted job, especially parallel jobs, there are two basic parameters expressed Job requirements for computing resources: reserved running time (WallClock Requested (WCR)), the number of reserved computing nodes (CPU, memory) (Number of computeNodes Requested (NNR)), this embodiment only uses the above two Parameters, that is, the reserved running time and the number of reserved computing nodes, so two-dimensional icons are used to display the statistical results, and multi-dimensional cases can be deduced by analogy), and each parameter is counted in sections. The job data on (Kraken) (NNR is in the range of 1 to 4128 nodes, and WCR is in the range of 0 to 24 hours) is divided into several parameter partitions. The selection of the partition threshold is mainly based on the following two points: each region contains some actual Operate the parameter combination mode selected by the user (such as 1 hour running time and 10 computing nodes, 12 hours running time and 1 computing node, etc.); ensure that each partition covers a considerable job record, so as to ensure the accuracy of statistical results reliability and effectiveness. Figure 2 is the job partition count table generated based on 30 months of historical load data on the supercomputer Kraken. Noise reduction and de-correlation processing of historical data: use the space-time clustering detection method (space-time clustering detection method), also known as the Knox method (Knox method), to check the task dependencies in the system load history records, the specific method is as follows: First, select a time critical point t* and a space critical point x* based on prior experience, if the submission time interval of two jobs is within the time critical point t* and the proximity of parameter selection is within the space interval x*, then It is considered that two jobs are considered to be correlated, and if the density of such job pairings is much higher than that of other job pairs, the historical data is considered to have a strong correlation within the (t*, x*) scale, thus It is necessary to take noise reduction and de-correlation processing. In order to check the above phenomenon, the present invention applies the Knox statistical test method applied in epidemiology. If the Knox method shows that the historical data has a strong presence within the (t*, x*) scale , then apply the following algorithm to perform noise reduction processing on historical load data (historical task records), as shown in Figure 6:

Step 201, arrange the historical load data in ascending order of submission time, and execute step 202, starting from the first load data, delete all parameters from the submission time of the load data to the next t* time that are less than x All relevant load data of *, step 203, update the load data and repeat step 202 until all historical load data are traversed.

The present invention uses the above algorithm to perform noise reduction and de-correlation processing on the load history of the system, thereby reducing the amount of data and generating relatively independent job records.

Long time waiting probability chart: The present invention defines a long time waiting (LTW: Long Time Waiting) threshold (task resource waiting time threshold), such as 1 hour, based on the new data after noise reduction and correlation removal, the establishment is as shown in Figure 2 For each partition, calculate the proportion of jobs whose waiting time is greater than the LTW threshold to all jobs in the partition, thereby generating a heatmap as shown in FIG. 3 .

Predicting task resource waiting time: The long-time waiting probability chart after noise reduction and decorrelation generated in Figure 3 can provide a macro parameter selection criterion. Taking the Kraken supercomputer as an example, under the same service calculation amount (service unit), the parameter selection scheme that reserves a small amount of computing time and a large number of computing nodes will generate less cost than the scheme that reserves a large amount of computing time and a small number of computing nodes. Waiting for a long time, thereby producing more efficient supercomputer user resource utilization efficiency, but, in order to provide a more refined and time-sensitive user guide, the present invention also needs to consider the time factor, and the specific steps are as follows:

Use the autocorrelation function to determine the time interval area with correlation: in the time series records of LTW probability in months as shown in Figure 4, the value at a certain moment often has a strong correlation with its adjacent historical data , the present invention uses autocorrelation function AutoCorrelation Function (ACF) to determine the time interval area with correlation. Specifically, the present invention uses an autocovariance function to describe the autocorrelation coefficient. The autocovariance function is the second-order mixed central moment between the values of the time series X(t) at any two different moments t1 and t2, and is used to describe the fluctuation of the values of X(t) at two moments ( relative to the mean), also known as the centralized autocorrelation function. Its definition is

$\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; [</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ,</span>$

Among them, E represents the expected value, and Xi represents the value of the random variable at t( _i ). μ _i represents the expected value at t(i), X _i+k represents the random variable value at t(i+k), μ _i+k represents the expected value at t(i+k), σ ² stands for variance.

Figure 5 shows the autocorrelation function value of the monthly long-time wait probability based on the supercomputer Kraken's job history: From Figure 5, it can be seen that the long-time wait probability (LTW probability) of the current month is different from that of the previous two months The LTW probability of the LTW has a significant statistical correlation, and then use the historical correlation time period data and the Bayesian method to predict the long-term waiting probability of the next time period. For the part of the time interval with correlation, the average value of the Beta-binomial hierarchical model (Beta-binomial hierarchical model) is applied to predict the probability of long waiting in the next time period. Taking the supercomputer Kraken as an example, because the autocorrelation function of the historical data shows that the long waiting time of the next month is related to the data of the previous two months, the average value of the Beta-binomial distribution model is equal to:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Among them, N _LTW,k-1,k-2 is the number of long-waiting jobs in the previous two months, N _k-1,k-2 is the total number of jobs in the previous two months, P _LTW,k is the number of long-waiting jobs in the current month Time waiting probability. Therefore, the present invention can predict the long-time waiting probability of each partition in the next month.

The present invention also proposes a system for predicting the waiting time of task resources, including the following modules:

Generate a new historical task record module, which is used to obtain historical task records, delete dependent task records in the historical task records, and generate new historical task records;

generating a task record set, which is used to obtain the task records in the time period related to the certain time period in the new historical task record through an autocorrelation function, and generate a task record set;

The prediction module is used to set the task resource waiting time threshold, obtain the number of task records whose waiting time of the task resource in the task record set exceeds the task resource waiting time threshold, and according to the total number of task records in the task record set, pass The Bayesian method predicts the waiting time of task resources within a certain period of time.

Judging module, judging whether there is a dependency relationship among the task records in the historical task record, wherein the time critical point t* and the spatial critical point x* are selected; if the submission time interval of the two task records in the historical task record is within the time critical point Within t* and the proximity of parameter selection is within the spatial interval x*, and the pairing density is higher than the pairing density of the historical task records except the two task records, then the two task records are in (t* , there is a dependency relationship within the scale of x*).

Delete the dependency module, if it exists, delete the task records with dependencies, where the task records in the historical task records are arranged in descending order of submission time; starting from the task record with the smallest submission time, delete the task record with the smallest submission time From the task record to the next t* time, the parameters of the task record with the minimum submission time are selected as task records whose proximity is less than x*; update the historical task record and repeat step 22 until the historical task record is traversed , to update the historical task record.

The formula for the Bayesian method in the forecasting module is:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Among them, N _LTW,k-1,k-2 is the number of the task records whose waiting time of the task resource exceeds the threshold of the task resource waiting time, and N _k-1,k-2 is the number of the task records in the task record set The total quantity, P _LTW,k is the task probability that the waiting time of the task resource exceeds the waiting time threshold of the task resource within the certain period of time.

Claims (10)

1. A method for predicting task resource waiting time, characterized in that, comprising: Step 1, obtain historical task records, delete task records that have dependencies in the historical task records, and generate new historical task records; Step 2, through the autocorrelation function, obtain the task records in the time period related to the time period to be predicted in the new historical task record, and generate a task record set; Step 3: Set the task resource waiting time threshold, obtain the number of task records whose task resource waiting time in the task record set exceeds the task resource waiting time threshold, and according to the total amount of task records in the task record set, use Bayeux The Adams method predicts the waiting time of task resources within the time period to be predicted. 2. the method for predicting task resource waiting time as claimed in claim 1, is characterized in that, the concrete steps of this step 1 comprise: Step 11, judging whether there is a dependency relationship among the task records in the historical task record; Step 12, delete task records with dependencies if they exist. 3. The method for predicting task resource waiting time as claimed in claim 2, is characterized in that, this step 11 comprises: Step 21, select time critical point t* and spatial critical point x*; Step 22, if the submission time interval of two task records in the historical task record is within the time critical point t* and the proximity of parameter selection is within the spatial interval x*, and the pairing density is higher than the historical task record In addition to the density of pairings of the two task records, the two task records have a dependency within the scale (t*, x*). 4. The method for predicting task resource waiting time as claimed in claim 2, is characterized in that, this step 12 comprises: Step 31, arrange the task records in the historical task records according to the submission time from small to large; Step 32, starting from the task record with the smallest submission time, deleting the task record with the smallest submission time to all task records whose parameter selection proximity to the task record with the smallest submission time within the next t* time is less than x*; Step 33, update the historical task record and repeat the step 22 until traversing the historical task record and updating the historical task record. 5. the method for predicting task resource waiting time as claimed in claim 1, is characterized in that, the formula of this Bayesian method in this step 3 is: ${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}$ Among them, N _{LTW, k-1, k-2} are the number of the task records whose waiting time of the task resource exceeds the threshold of the task resource waiting time, and N _{k-1, k-2} are the task records in the task record set The total amount of P _LTW,k is the task probability that the waiting time of the task resource exceeds the threshold of the waiting time of the task resource in the time period to be predicted. 6. A system for predicting the waiting time of task resources, characterized in that it comprises: Generate a new historical task record module, which is used to obtain historical task records, delete dependent task records in the historical task records, and generate new historical task records; Generate a task record set, which is used to obtain task records in the new historical task record in a time period that is correlated with the time period to be predicted by an autocorrelation function, and generate a task record set; The prediction module is used to set the task resource waiting time threshold, obtain the number of task records whose task resource waiting time exceeds the task resource waiting time threshold in the task record set, and according to the total amount of task records in the task record set, pass The Bayesian method predicts the waiting time of task resources within the time period to be predicted. 7. The system of predicting task resource waiting time as claimed in claim 6, is characterized in that, this generation new historical task record module also comprises: A judging module, judging whether there is a dependency relationship among the task records in the historical task records; Delete the dependency module, if it exists, delete the task records with dependencies. 8. The system for predicting task resource waiting time as claimed in claim 7, wherein the specific functions of the judging module include: selecting a time critical point t* and a spatial critical point x*; The submission time interval of a task record is within the time critical point t* and the proximity of parameter selection is within the space interval x*, and the pairing density is higher than the pairing density of the historical task record except the two task records , then the two task records have dependencies within the (t*, x*) scale. 9. The system for predicting the waiting time of task resources as claimed in claim 7, wherein the specific function of the deletion dependency module includes: arranging the task records in the historical task records according to the submission time from small to large; Starting from the task record with the smallest submission time, delete the task record with the smallest submission time to all the task records whose parameter selection proximity to the task record with the smallest submission time within the next t* time is less than x*; update the history Task records and repeat step 22 until the historical task records are traversed and the historical task records are updated. 10. The system of predicting task resource waiting time as claimed in claim 6, is characterized in that, the formula of this Bayesian method in this prediction module is: ${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; LTW</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}}}$ Among them, N _{LTW, k-1, k-2} are the number of the task records whose waiting time of the task resource exceeds the threshold of the task resource waiting time, and N _{k-1, k-2} are the task records in the task record set The total amount of P _LTW,k is the task probability that the waiting time of the task resource exceeds the threshold of the waiting time of the task resource in the time period to be predicted.