CN118093170A - A GPU resource scheduling method and system based on heterogeneous computing platform - Google Patents

Abstract

The present invention discloses a GPU resource scheduling method and system based on a heterogeneous computing platform. Firstly, a GPU is selected as a general resource and the GPU resources under the scattered heterogeneous computing platforms are concentrated to form a GPU resource pool for scheduling by a scheduler. Secondly, monitoring data of the heterogeneous platform is collected, and tools under different platforms are used to collect data. The collected data may have data anomalies due to equipment failure, network problems, software operation problems, etc., and anomaly detection algorithms can be used to identify and locate abnormal data, and timely measures can be taken to process them. Then, in view of the problems of large resource scheduling overhead and low efficiency of the heterogeneous platform, a prediction scheduling algorithm is used to predict the resource usage trend of the platform, and resources are reasonably allocated to the platform in advance. Finally, according to the above prediction results, resources are reasonably allocated by the scheduler to improve the utilization rate of GPU resources.

Description

A GPU resource scheduling method and system based on heterogeneous computing platform

Technical Field

The present invention belongs to the technical field of resource scheduling and middleware in computer science, and relates to a scheduling method and system for GPU resources under heterogeneous platforms, using middleware of different computing platforms to collect monitoring data, and particularly relates to collecting data from an openstack cluster of a cloud computing platform, a slurm cluster of a high-performance computing platform, and a kubernetes cluster of an artificial intelligence platform.

Background technique

With the development of information technology, data centers have an increasing demand for computing power, and need to make better use of computing power to process explosively growing data. Computing power has become the core productivity. As a computing power production center and supply center, data centers provide various computing power-based services to the outside world. GPUs have become an important computing resource with their powerful computing power, high memory bandwidth, and parallel computing and processing of massive data sets.

Currently, cluster scheduling for data centers has become a research focus, and different scheduling systems have emerged in different fields, such as artificial intelligence, big data analysis, and cloud computing. For a single computing platform, its scheduling system can effectively schedule tasks and allocate resources to improve resource utilization.

The GPU scheduling framework based on the artificial intelligence platform is mainly the Kubernetes framework. Kubernetes is a scheduling framework based on container technology that can virtualize a physical resource into multiple logical resources, which can be used by different applications. The GPU resource scheduling framework based on the high-performance computing platform is mainly Slurm, which is a cluster manager and job scheduling system widely used in high-performance computing clusters. It is popular for its many advantages such as open source, high fault tolerance and high scalability.

Common data anomaly detection algorithms are mainly based on machine learning and deep learning. Among them, data anomaly detection models based on machine learning include SR, SVR and Prophet, and data anomaly detection models based on deep learning mainly include VAE and LSTM. Different detection algorithms are suitable for different data scenarios, and different data anomaly detection models need to be selected according to specific data characteristics.

The ARIMA model is a time series model. The basic idea is to use the historical information of the data itself to predict future information. The model can be used to predict the resource usage of the tasks to be scheduled.

The existing resource scheduling of heterogeneous computing platforms has problems such as large differences in data feature values collected by each platform, high degree of data anomaly, and poor accuracy and timeliness of resource scheduling of heterogeneous computing platforms, resulting in waste of resources.

Summary of the invention

The present invention proposes a method and system for GPU resource scheduling for heterogeneous computing platforms, which are used to partially solve the problem of low GPU resource utilization of heterogeneous computing platforms existing in the prior art.

The technical solution adopted by the method of the present invention is: a GPU resource scheduling method based on a heterogeneous computing platform, comprising the following steps:

Step 1: Use GPU as a general resource and centralize the GPU resources scattered on heterogeneous computing platforms to form a GPU resource pool for scheduling by the scheduler;

Step 2: Collect monitoring data from heterogeneous platforms to provide data support for data anomaly detection and predictive scheduling;

Step 3: Detect the data and process the detected abnormal data;

Step 4: Predict resource usage trends, predict the resource usage trends of the platform, allocate resources to the platform in advance, and further make full use of GPU resources;

Step 5: Based on the prediction results of step 4, resources are scheduled between different computing platforms.

Preferably, in step 1, GPU virtualization technology is used to convert large cards to small cards, and remote GPU functions are used to break the boundaries of physical servers, extending the management and use of GPUs from a single server to the entire data center, realizing pooling of GPU resources of different heterogeneous platforms, and forming a GPU resource pool that can be scheduled by the scheduler.

Preferably, in step 2, the cloud computing platform data collection uses the Ceilometer framework of the openstack community, and the Ceilometer framework uses an agent to complete the data collection of cluster resource information.

Preferably, in step 2, historical data collection of the artificial intelligence platform uses the Resource metricsAPI and Custom Metrics API to collect resource usage data of the cluster.

Preferably, in step 2, historical resource usage information data of the high-performance computing platform is collected by deploying Promethes of the Slurm cluster. Promethes is an open source framework for monitoring and alarming, and the exporter of Promethes can be used to export various resource information of the high-performance computing platform.

Preferably, in step 3, an anomaly detection model of heterogeneous computing platform data is used to detect anomalies in the data and process the detected data anomalies;

The anomaly detection model based on heterogeneous computing platform data extracts the stability and periodicity characteristic values of the data of each application platform, and then the model routing selects a matching anomaly detection algorithm according to the characteristic values.

Preferably, in step 3, data anomaly detection is performed based on time series, wherein the most frequently occurring and widely used anomaly category is point anomaly;

The formula for judging data anomalies based on point anomalies is as follows:

Among them, _xt refers to the data feature value at time t, It refers to the expected value of the data at time t, and τ represents a threshold value greater than 0. If at time t, the absolute value of the difference between the characteristic value of the data and the expected value is greater than the given threshold, it means that the data collected at time t is abnormal and needs to be de-anomaly processed.

Preferably, in step 4, a resource usage trend forecast is performed;

The ARIMA model first inputs the collected time series, verifies the stability of the time series, and uses the difference model to perform difference operations to eliminate the volatility of the data and make the stability of the data meet the requirements. In this process, the order of the difference is determined; after the data stability meets the requirements, the white noise of the data is tested. In this process, the AR model and the MA model are used to perform white noise removal operations to make the white noise of the data meet a certain threshold, and the ARIMA model is fitted. The autoregressive model is used to determine the order p of the autoregressive part of the data and the moving average model is used to determine the order q representing the moving average part of the data. When the stability of the data and the white noise meet the requirements, the modeling of the ARIMA model is completed.

The technical solution adopted by the system of the present invention is: a GPU resource scheduling system based on a heterogeneous computing platform, comprising the following modules:

The GPU resource pool construction module is used to use GPU as a general resource and centralize the GPU resources scattered on the heterogeneous computing platform to form a GPU resource pool for scheduling by the scheduler;

Heterogeneous platform monitoring data collection module, used to collect monitoring data of heterogeneous platforms and provide data support for data anomaly detection and predictive scheduling;

An abnormal data detection and processing module is used to detect data and process the detected abnormal data;

The trend prediction module is used to predict the resource usage trend of the platform, allocate resources to the platform in advance, and further make full use of GPU resources;

The resource scheduling module is used to predict results and schedule resources between different computing platforms.

Compared with the prior art, the beneficial effects of the present invention include:

(1) The present invention decouples GPU nodes from the platform and pools them to achieve resource scheduling among heterogeneous computing platforms, thereby improving the overall resource utilization of the data center.

(2) The present invention proposes a data anomaly detection model based on a heterogeneous computing platform. In order to solve the problem of high anomaly levels and large differences in eigenvalues of data on different platforms, the data quality is improved by extracting data eigenvalues and selecting matching anomaly algorithms.

(3) The present invention proposes a method for resource prediction. To address the problem of high overhead and low efficiency in resource scheduling on heterogeneous computing platforms, the present invention proposes using the ARIMA model to perform resource prediction before scheduling to improve resource utilization.

BRIEF DESCRIPTION OF THE DRAWINGS

The technical solution of this invention is further described below using embodiments and specific implementation methods. In addition, some drawings are also used in the process of describing the technical solution. For those skilled in the art, other drawings and the intention of the invention can also be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of a GPU resource scheduling method based on a heterogeneous computing platform provided by an embodiment of the present invention;

FIG2 is a modeling flow chart of the ARIMA model in the present invention;

FIG3 is a schematic diagram of the structure of an openstack cluster collecting data in an embodiment of the present invention;

FIG4 is a schematic diagram of the structure of Kurbenets cluster data collection in an embodiment of the present invention;

FIG5 is a flow chart of a data anomaly detection model according to an embodiment of the present invention;

FIG6 is a flow chart of resource prediction in an embodiment of the present invention;

FIG. 7 is a schematic diagram showing the principle of resource scheduling in an embodiment of the present invention.

Detailed ways

In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

Please refer to FIG1 . This embodiment provides a GPU resource scheduling method based on a heterogeneous computing platform, including the following steps:

Step 1: Use GPU as a general resource and centralize the GPU resources scattered in heterogeneous computing platforms to form a GPU resource pool for scheduling by the scheduler. Use GPU virtualization technology to convert large cards to small cards. GPU virtualization technology can use the graphics card virtualization technology provided by NVIDIA, namely VCUDA technology, to achieve fine-grained division, reorganization and reuse of GPU resources, support fine-grained GPU sharing, and support dynamic allocation and automatic release of GPU resources. Use remote GPU functions to break the boundaries of physical servers, expand the management and use of GPUs from a single server to the entire data center, and achieve pooling of GPU resources on different heterogeneous platforms to form a GPU resource pool that can be scheduled by the scheduler.

Step 2: Collect monitoring data of the heterogeneous computing platform, that is, resource information of the heterogeneous computing platform.

The data collection method uses an openstack cluster, a Kurbenetes cluster, and a Slurm cluster to collect data.

In one implementation, the flow chart of data collection in the openstack cluster is shown in FIG3 . The Ceilometer framework of the openstack community is mainly responsible for data collection, and its data collection is mainly completed by agents. Ceilometer includes compute, central, and other agents. The resource information data collection based on the openstack cluster is described as follows.

Compute agent is responsible for collecting resource usage data of VM instances deployed on each computing node of the openstack platform.

The Central agent is responsible for collecting resource usage information for OpenStack components that do not publish messages through the message queue. It can also collect resource usage information for hardware-related resources.

The resource information collected above is sent to the collector through the message queue for aggregation, and the collected information is stored in the database.

In one implementation, a flow chart of Kurbenetes cluster collecting resource information is shown in FIG4 .

Cluster resource usage data is collected based on Resource metrics API and Custom Metrics API. Kubelet acts as a node-level and application-level indicator collector to monitor resource usage of each node.

Metrics-server stores the latest indicator values captured from Kubelet locally and exposes the main indicator API to external clients, responsible for providing the main indicator API service. By accessing the resource indicator API, you can view the resource usage of pods and nodes in the cluster, and summarize the collected resource usage in the database.

In one embodiment, the slurm cluster crawls the metrics data provided by the HTTP service from each client by deploying Prometheus. Prometheus is an open source framework for monitoring and alerting. For the monitoring framework, the system uses the Prometheus database to store the metrics data of the job, and the Prometheus exporter collects the metrics of the entire Slurm scheduling system, including various resource information.

Step 3: Detect and process the collected data for anomalies. The flowchart is shown in Figure 5. The specific steps are described as follows:

In one embodiment, an anomaly detection model for heterogeneous computing platform data is used to detect anomalies in the data and process the detected data anomalies;

This embodiment proposes an anomaly detection model based on heterogeneous computing platform data. This model extracts the stability and periodicity characteristic values of each application platform data, and then the model routing selects a matching anomaly detection algorithm based on the characteristic values. The process of model routing is the process of selecting an anomaly detection algorithm, which improves data quality and provides data support for the scheduling algorithm.

Commonly used models for data anomaly detection methods: For data with different eigenvalues, different anomaly detection models can be used to detect anomalies in the data. In the present invention, the data anomaly detection models used include SR, SVR, Prophet, VAE, and LSTM algorithm models. The above models are suitable for data anomaly detection scenarios with different requirements for detection time, accuracy, computational consumption, and scope of application, to achieve efficient data anomaly detection.

In one embodiment, based on machine learning and deep learning, data anomaly detection is performed and the detected data anomaly is processed;

It includes the following sub-steps:

Step 3.1: Take the data collected in step 2 as input.

Step 3.2: Clean the collected data.

Step 3.2.1: Remove duplicate records from the data. In the collected data set, there may be some identical records. The redundant records have no task value, so it is necessary to remove the redundant duplicate data. You can use the drop_duplicate() function in the Pandas library to remove duplicate records.

Step 3.2.2: Fill missing values. The actual collected data is often incomplete and often contains missing values. It is necessary to process the missing values to make the data more complete. You can use the fillna() function in the Pandas library to replace missing values with specific data feature values.

Step 3.2.3: Data correction, in order to unify data standards, reduce secondary errors, and reduce the impact of unstable data collection caused by network fluctuations, the data needs to be corrected.

Step 3.3: Feature extraction. Extract features from the cleaned data. The quality of data feature extraction will directly affect the effect of anomaly detection. Data feature extraction mainly extracts the stability and periodicity of the data. Stability detection can use the ADF test, which is a time series stationarity test method. It determines whether the data is stable by testing whether the sequence has a unit root. Periodicity detection uses Fourier coefficients as the relevant judgment basis.

Step 3.4, the model routing process is the recommendation process of the anomaly detection algorithm. Commonly used data anomaly detection algorithms are based on machine learning and deep learning. The data anomaly detection algorithms based on machine learning mainly include SR, SVR and Prophet, and the data anomaly detection algorithms based on deep learning include VAE and LSTM. Different algorithms are suitable for different data anomaly detection models. The usage scenarios of various algorithms are shown in Table 1.

Table 1 Comparison of anomaly detection algorithms

Based on the characteristic coefficients of the indicators obtained in the previous step, namely the stability and periodicity of the data, the model routing process will recommend the most suitable anomaly detection algorithm based on the characteristic coefficients of the data and other algorithm indicators (such as accuracy, detection time, and detection consumption).

Step 3.5: The detector calls the algorithm recommended by the model routing to detect data anomalies, retains the data with normal detection results, and discards the data with abnormal detection results.

The data monitored by the data center is usually based on time series, and the detected data generally has the attribute of timestamp. There are different forms of data anomalies based on time series, among which the most frequent and widely used anomaly category is point anomaly.

The formula for judging data anomalies based on point anomalies is as follows:

Among them, _xt refers to the data characteristic value at time t, It refers to the expected value of the data at time t, and τ represents a threshold value greater than 0. If at time t, the absolute value of the difference between the characteristic value of the data and the expected value is greater than the given threshold, it means that the data collected at time t is abnormal and needs to be de-anomaly processed.

Step 4: Using the above data after removing anomalies, use the resource prediction model and algorithm to predict the resource usage trend. In the present invention, the ARIMA model is used to predict resources and plan the resource allocation in advance.

The full name of the ARIMA model is the Autoregressive Difference Moving Average model. Its basic idea is to use the historical information of the data itself to predict the future. It attempts to extract the time series patterns hidden behind the data through the autocorrelation and difference of the data, and then use these patterns to predict future data. In this way, the resource usage trend can be predicted. The ARIMA model consists of three parts, namely the autoregressive model AR, the difference process I and the moving average model MA.

In one embodiment, the ARIMA model modeling flow chart is shown in Figure 2; first, the collected time series is input, the time series is checked for stationarity, and the differential model is used to perform differential operations to eliminate the volatility of the data and make the data stationarity meet the requirements. In this process, the order of the differential is determined. After the data stationarity meets the requirements, the white noise of the data is tested. In this process, the AR model and the MA model are used to perform white noise removal operations so that the white noise of the data meets a certain threshold, and the ARIMA model is fitted. The order p of the autoregressive part of the data is determined using the autoregressive model and the order q representing the moving average part of the data is determined using the moving average model. When the stationarity and white noise of the data meet the requirements, the modeling of the ARIMA model is completed, and the resource usage trend can be predicted based on the data after the abnormality is removed and the relevant parameters obtained in the modeling process.

The three main components of the ARIMA model are described as follows:

1) Autoregressive model AR is a statistical model for analyzing time series data, which is used to describe the relationship between a variable and its past values. For a time series data, the P-order form of the AR model can be expressed as:

Where _Yt is the observed value at time t, c is a constant, is the autoregressive coefficient, Y _t-1 is the observed value at time t-1, and the last term is the error.

2) Moving average model MA describes the relationship between the data at the current time point and the past noise. MA is based on the assumption of white noise sequence. White noise is a special time series model. The data at each time point are independent and identically distributed, and have a constant mean and variance. Given a white noise sequence η _t , the MA model is defined as:

Y _t = μ + η _t + θ ₁ η _t-1 + θ ₂ η _t-2 + ........ + θ _q η _tq

Among them, _Yt is the observed value of the time series we are interested in at time t, μ is the mean or expected value of the time series, _ηt , ηt _-1 ,.... _ηtq is the white noise at the corresponding time, _θ1 , _θ2 ,... _θq are the corresponding parameters of the model, which measure the influence of white noise on the current time point.

3) The difference process I is a mathematical operation that describes the difference between adjacent data points in a set of numerical sequences. For a time series Y _t , the first-order difference can be defined as:

ΔY _t =Y _t -Y _t-1

Among them, Y _t-1 is the time series at time t-1, and ΔY _t is the difference between the two time series.

The specific process is shown in Figure 6, and the steps are described as follows:

Step 4.1: Input historical data information, that is, the data after de-abnormalization in the previous step.

Step 4.2: Perform preliminary data analysis using the difference model, autoregressive model, and moving average model to obtain the time order in order to determine the time order of the ARIMA model.

Step 4.2.1: Historical data information is a time series, and its information attribute carries the time value of the collection moment. Using the difference model, calculate the difference between two adjacent time series. After processing by the difference model, data fluctuations can be eliminated to a certain extent.

Step 4.2.2: Use the autoregressive model AR to process the autoregressive part of the historical data information, and make preliminary predictions based on it. Process the historical data information to make it more stable. The AR model takes into account the impact of observations from several periods in the past on the current value.

Step 4.2.3: Use the MA model, or moving average model, to handle temporary, sudden changes or noisy time series data. The MA model takes into account the impact of past forecast errors on current values.

Step 4.3: Obtain the time order from the previous step, use this time order as the time order of the ARIMA model, use the ARIMA model to make predictions, and obtain the prediction results.

Step 5: Schedule GPU resources based on the prediction results.

The specific process is shown in Figure 7, and the steps are described as follows:

Step 5.1: The scheduling module obtains various conditions of GPU resources through the scheduler of the heterogeneous platform, including the usage of GPU resources, the type of GPU, the number of GPUs, and which heterogeneous platform the GPU belongs to.

Step 5.2: The scheduling module records the GPU resources and other resources required for each task, records the status of each task, and monitors the completion of the task.

Step 5.3: Based on the resource prediction results of step 4 and the usage of GPU resources, determine whether there are GPU resources that meet the task requirements. If not, wait until GPU resources that meet the task are available for allocation. If yes, proceed to step 5.4.

Step 5.4: According to the GPU resource record table of the scheduling module, the scheduling module allocates the required GPU resources for the task by calling the schedulers of each platform, that is, the GPU resources required for a task come from different platforms.

Step 5.5: The scheduling module monitors the running status of the job. If the job is running, it calls the scheduler of each platform to release the GPU resources and return the GPU resources to the GPU resource pool.

This embodiment also provides a GPU resource scheduling system based on a heterogeneous computing platform, including the following modules:

The GPU resource pool construction module is used to use GPU as a general resource and centralize the GPU resources scattered on the heterogeneous computing platform to form a GPU resource pool for scheduling by the scheduler;

Heterogeneous platform monitoring data collection module, used to collect monitoring data of heterogeneous platforms and provide data support for data anomaly detection and predictive scheduling;

An abnormal data detection and processing module is used to detect data and process the detected abnormal data;

The trend prediction module is used to predict the resource usage trend of the platform, allocate resources to the platform in advance, and further make full use of GPU resources;

The resource scheduling module is used to predict results and schedule resources between different computing platforms.

The GPU resources of different heterogeneous computing platforms are relatively scattered, and it is necessary to concentrate the GPU resources of different homogeneous platforms to form a GPU resource pool, and a scheduler is used to uniformly schedule and allocate the resource pool in the GPU to achieve the purpose of making full use of GPU resources. The present invention can use the data tools of different heterogeneous computing platforms to collect data. The three common distributed cluster scheduling platforms in data centers include cloud computing platforms, artificial intelligence platforms, and high-performance computing platforms, and use the components provided by the three platforms to collect data. In the above three platforms, each has its own component to collect data. The cloud computing platform mainly uses the openstack cluster to collect data, the artificial intelligence platform mainly uses the kubernetes cluster to collect data, and the high-performance computing platform uses the Slurn cluster to collect data.

The data collected by different data platforms are diverse and complex, resulting in high anomaly levels and large differences in eigenvalues of heterogeneous platform data. For the collected data, the present invention first needs to clean the data, filter out some data that does not meet the requirements, and use a model to extract the stability and periodic eigenvalues of the data of each application platform. Then, the model routing selects a matching anomaly algorithm based on the eigenvalues of the data. Commonly used anomaly detection algorithms are mainly based on machine learning and deep learning, which perform anomaly detection on the data and process the detected data anomalies.

The resource scheduler selects a suitable node, allocates GPU resources according to the number and type of GPUs required for the task, and migrates the task to a node that meets the resource allocation requirements for execution. If the demand for GPU resources of other tasks is consistent with that of the current task, the two tasks can use the GPU resources of the node in a time-sharing manner, thereby achieving the purpose of sharing GPU resources and improving the utilization rate of GPU resources.

It should be understood that the embodiments described above are part of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be arbitrarily combined with each other to form a feasible technical solution. Such combination is not restricted by the sequence of steps and/or the structural composition mode, but must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the protection scope required by the present invention.

It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims (9)

1. The GPU resource scheduling method based on the heterogeneous computing platform is characterized by comprising the following steps of:

Step 1: using the GPU as a general resource and centralizing the GPU resources scattered on the heterogeneous computing platform to form a GPU resource pool for scheduling by a scheduler;

Step 2: collecting monitoring data of a heterogeneous platform, and providing data support for data anomaly detection and prediction scheduling;

Step 3: detecting the data and processing the detected abnormal data;

Step 4: predicting the use trend of resources, predicting the use trend of the resources of the platform, distributing the resources for the platform in advance, and further fully utilizing the GPU resources;

step 5: and (3) scheduling the resources among different computing platforms according to the prediction result of the step (4).

2. The heterogeneous computing platform-based GPU resource scheduling method of claim 1, wherein: in the step 1, the large card is changed into the small card by using a GPU virtualization technology, the boundary of a physical server is broken by using a remote GPU function, management and use of the GPU are expanded from a single server to the whole data center, pooling of GPU resources of different heterogeneous platforms is realized, and a GPU resource pool which can be scheduled by a scheduler is formed.

3. The heterogeneous computing platform-based GPU resource scheduling method of claim 1, wherein: in the step2, the data collection of the cloud computing platform uses Ceilometer frames of an opentack community, and Ceilometer frames use agents to complete the data collection of cluster resource information.

4. The heterogeneous computing platform-based GPU resource scheduling method of claim 1, wherein: in step 2, historical data of the artificial intelligence platform is collected, and Resource usage data of the clusters is collected based on Resource METRICS API and Custom METRICS API.

5. The heterogeneous computing platform-based GPU resource scheduling method of claim 1, wherein: in step 2, historical resource usage information data of the high-performance computing platform is collected through Promethes of the deployment Slurm cluster.

6. The heterogeneous computing platform-based GPU resource scheduling method of claim 1, wherein: in the step 3, based on an anomaly detection model of heterogeneous computing platform data, anomaly detection is carried out on the data, and detected data anomalies are processed;

according to the anomaly detection model based on the heterogeneous computing platform data, the stability and the periodic characteristic values of each application platform data are extracted, and then a model route selects a matched anomaly detection algorithm according to the characteristic values.

7. The heterogeneous computing platform-based GPU resource scheduling method of claim 1, wherein: in the step 3, data anomaly detection is carried out based on a time sequence, wherein the anomaly category with the highest occurrence frequency and the widest use range is point anomaly;

the formula for judging data abnormality according to the point abnormality is as follows:

Wherein x _t is a characteristic value of data at time t, x _t is an expected value of the data at time t, τ is a threshold value greater than 0, and if the absolute value of the subtraction of the characteristic value and the expected value of the data at time t is greater than a given threshold value, it is indicated that the data acquired at time t is abnormal, and the data at the time t needs to be processed for abnormality removal.

8. The heterogeneous computing platform-based GPU resource scheduling method of any of claims 1-7, wherein: in step 4, predicting the using trend of the resource by using an ARIMA model;

The ARIMA model firstly inputs the acquired time sequence, performs stability verification on the time sequence, and performs differential operation by using a differential model so as to eliminate the fluctuation of data, so that the stability of the data meets the requirement, and in the process, the differential order is determined; after the data stability meets the requirement, the white noise of the data is checked, in the process, the AR model and the MA model are used for carrying out the white noise removing operation, so that the white noise of the data meets a certain threshold value, the ARIMA model is fitted, the order p of the autoregressive part of the data is determined by using the autoregressive model, the order q of the moving average part of the data is determined by using the moving average model, and when the data stability and the white noise meet the requirement, the modeling of the ARIMA model is finished.

9. GPU resource scheduling system based on heterogeneous computing platform, characterized by comprising the following modules:

The GPU resource pool construction module is used for taking the GPU as a general resource and dispersing the GPU resources in the GPU resource sets of the heterogeneous computing platforms to form a GPU resource pool for a scheduler to schedule;

The heterogeneous platform monitoring data acquisition module is used for acquiring monitoring data of the heterogeneous platform and providing data support for data anomaly detection and prediction scheduling;

the abnormal data detection processing module is used for detecting the data and processing the detected abnormal data;

The trend prediction module is used for predicting the use trend of the resources, predicting the use trend of the resources of the platform, distributing the resources for the platform in advance, and further fully utilizing the GPU resources;

and the resource scheduling module is used for predicting results and scheduling the resources among different computing platforms.