Bare Metal Server Computing Power Scheduling Platform

Abstract

The present invention belongs to the technical field of computing power scheduling, and discloses a bare metal server computing power scheduling platform, including: when the r-th server receives task data sent by a user end, a network delay condition is set according to the task data; the r-th server records the network delay between R servers communicating with each other, and divides the servers within the network delay condition into a processing partition, and connects the servers in the same processing partition through a blockchain to form a partitioned blockchain; a first analysis module marks the server that receives the user end task data as a receiving server, the receiving server extracts feature data of the task data, inputs the feature data into a pre-trained split evaluation model, obtains a split evaluation result, generates task description data according to the split evaluation result, and publishes the task description data on the partitioned blockchain; computing resources are fully utilized, and efficient computing power resource scheduling is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of computing power scheduling, and more specifically, to a bare metal server computing power scheduling platform. Background Art

[0002] Bare Metal Server is a physical server with high performance, stability and security. In the traditional bare metal server computing power scheduling system, the server's computing resources usually rely on a central server for management and allocation. The central server is responsible for receiving task requests, evaluating the load of each server, and determining the allocation of tasks. However, this centralized architecture has obvious technical bottlenecks and risks, which are reflected in the following aspects:

[0003] Once the central server fails, the entire task allocation and scheduling will be interrupted, the system will easily be paralyzed, and the continuity and stability of data processing will be affected.

[0004] Scheduling systems usually lack dynamic analysis of task data, resulting in unreasonable allocation of server resources, overloaded or idle servers, and overall low efficiency.

[0005] Due to network latency and the lack of a flexible partitioning mechanism, servers may be assigned inappropriate tasks, increasing processing latency.

[0006] Faced with large and divisible tasks, it is difficult for the system to dynamically identify the need to split data, resulting in the lack of full utilization of parallel processing potential and prolonged data processing time.

[0007] In view of this, the present invention proposes a bare metal server computing power scheduling platform to solve the above problems. Summary of the invention

[0008] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purpose, the present invention provides the following technical solution: a bare metal server computing power scheduling platform, comprising:

[0009] Partitioning module, when the rth server receives the task data sent by the user end, it sets the network delay condition according to the task data; the rth server records the network delay between the R servers communicating with each other, and divides the servers within the network delay condition into a processing partition, r∈R, and connects the servers in the same processing partition through the blockchain to form a partitioned blockchain;

[0010] The first analysis module marks the server that receives the task data of the user end as a receiving server, and the receiving server extracts feature data of the task data, inputs the feature data into a pre-trained split evaluation model, obtains a split evaluation result, and the split evaluation result includes splittable and non-splittable, generates task description data according to the split evaluation result, and publishes the task description data to the partition blockchain;

[0011] In the second analysis module, each server in the same processing partition inputs its real-time load data and the task description data into the pre-trained task data evaluation model to obtain the required processing time, and publishes the required processing time on the partition blockchain;

[0012] The third analysis module records the required processing time for each server in the same processing partition, compares the required processing time published by each server with the required processing time published by other servers, and the server whose required processing time is not the shortest exits the processing partition; marks the server corresponding to the shortest required processing time as a processing server, the processing server receives the corresponding task data sent by the receiving server, and the processing server sends the processed data to the user end.

[0013] Furthermore, the method for generating task description data according to the split evaluation result includes:

[0014] If the split evaluation result is that the task cannot be split, the task description data includes the user terminal ID, the total data size, the total data identity ID, the data type and the data format;

[0015] If the splitting evaluation result is that it can be split, the task data is divided into q parts to obtain q parts of sub-data, where q is an integer greater than 1 and q is a preset value; the task description data includes the user terminal ID, sub-data size, sub-data identity ID, data type, and data format, and the sub-data identity ID includes the sub-data unique identity number and the position of the sub-data in the task data.

[0016] Further, if the task data is indivisible data, the required processing time for each server in the same processing partition is one;

[0017] If the task data is divisible data, the required processing time for each server in the same processing partition is q, and the q required processing times correspond one-to-one to the required processing time when processing q pieces of data.

[0018] Further, the method for the processing server to receive the corresponding task data sent by the receiving server is:

[0019] The receiving server determines the sub-data according to the required processing time corresponding to each processing server, and sends the sub-data to the corresponding processing server, and the required processing time corresponds to the sub-data one by one.

[0020] Furthermore, the network delay condition includes a network speed threshold and a network speed standard deviation threshold, and the network speed threshold and the network speed standard deviation threshold are both preset values;

[0021] Methods for dividing processing partitions include:

[0022] The network speed when R servers communicate with each other in a unit time is recorded, and R network speed analysis sets are established one by one. The average network speed and standard deviation of the network speed analysis sets are calculated.

[0023] Furthermore, the method for setting the network delay condition according to the task data includes:

[0024] Inputting the total data size, total data identity ID, data type and data format of the task data into a pre-trained condition setting model to obtain a network delay condition;

[0025] The training methods for conditioning models include:

[0026] Collect k groups of condition evaluation data in advance, the condition evaluation data including condition setting feature data and network delay conditions corresponding to the condition setting feature data, the condition setting feature data including total data size, total data identity ID, data type and data format;

[0027] Convert k groups of conditional evaluation data into feature vectors and divide them into training set and test set. Use the conditional setting feature data in the training set as the input of the conditional setting model, use the network delay conditions in the training set as the output of the conditional setting model, use the test set to test the conditional setting model, and output the conditional setting model that meets the prediction accuracy. The conditional setting model is a deep neural network or a gradient boosting decision tree model.

[0028] Furthermore, the characteristic data includes data type, data format, data dependency, data structure complexity and data integrity constraints;

[0029] The training method of the split evaluation model includes:

[0030] Collect k groups of evaluation data in advance, where k is an integer greater than 1, and the evaluation data includes feature data and split evaluation results corresponding to the feature data, and convert the evaluation data into feature vectors;

[0031] Each group of feature vectors is used as the input of the split evaluation model, and the split evaluation model takes a group of predicted split evaluation results corresponding to each group of feature data as output, and takes the actual split evaluation results corresponding to each group of feature data as the prediction target; minimizing the sum of prediction errors of all feature data is used as the training target; the split evaluation model is trained until the sum of prediction errors reaches convergence, and the training is stopped; the split evaluation model is a decision tree model, a random forest model or a gradient boosting tree model.

[0032] Furthermore, the training method of the task data evaluation model includes:

[0033] Collect H groups of task evaluation data in advance. The task evaluation data includes evaluation feature data and the required processing time corresponding to the evaluation feature data. The evaluation feature data includes load data and task description data. The load data includes CPU data, memory usage, total memory capacity, hard disk utilization, total hard disk capacity and other data. The CPU data includes CPU utilization, clock speed, number of cores, number of threads, cache capacity and floating-point computing capability.

[0034] The H group of task evaluation data are converted into feature vectors and divided into training set and test set. The evaluation feature data in the training set is used as the input of the task data evaluation model, and the required processing time in the training set is used as the output of the task data evaluation model. The task data evaluation model is tested with the test set, and the task data evaluation model that meets the prediction accuracy is output. The task data evaluation model is a deep neural network or a gradient boosting decision tree model.

[0035] Furthermore, the processing server sends the processed data to the corresponding user terminal according to the user terminal ID.

[0036] Furthermore, if the task data is divisible, the user terminal merges the sub-data according to their positions in the task data, and then displays them on the user terminal.

[0037] The technical effects and advantages of the bare metal server computing power scheduling platform of the present invention are as follows:

[0038] The decentralized nature of the platform avoids relying on central servers for computing power scheduling. The server adaptively sets the network delay conditions according to the task data, thereby automatically forming processing partitions. Even if a server fails, other servers can continue to execute the scheduling tasks, greatly reducing the risk of single point failures. This structural design ensures the high availability and continuity of computing power scheduling; according to the network delay conditions of the task, qualified servers are organized into a processing partition. Servers in the same processing partition are interconnected through blockchain, and determine whether they can be split for processing based on the characteristic data of the task. The adaptive partitioning and judgment mechanism realizes flexible computing power scheduling, ensuring that each partition operates under low latency and sufficient resources, greatly improving data processing efficiency.

[0039] In the processing partition, the server predicts the required processing time based on its own real-time load and task description data, and publishes the prediction results on the blockchain. Each server compares the required processing time. The server whose processing time is not the shortest will automatically exit the current partition. The competitive elimination mechanism ensures that the optimal server for processing the task is responsible for execution, reduces resource waste, avoids inefficient servers occupying resources, and improves the efficiency of the entire partition.

[0040] When the task data is divisible, it is automatically divided and distributed to multiple servers for parallel processing. After the processing is completed, the user end automatically merges the processed data to achieve rapid display. The splitting and parallel processing mechanism is particularly suitable for big data tasks, which significantly shortens the processing time and achieves the best resource utilization. The blockchain is used to connect servers in the same processing partition to achieve decentralized task time prediction and release. The processing time information of each server is transparent and open, and is updated in real time on the blockchain, which prevents information tampering and ensures the openness and fairness of task scheduling.

[0041] Decentralized blockchain connections and dynamic adaptive allocation of processing partitions reduce the risk of task data loss. When a server exits a processing partition due to a failure, other servers can quickly reorganize a new processing partition and continue processing tasks, ensuring the security and integrity of task data. Through adaptive scheduling, the server evaluates its own computing power and task requirements in real time, automatically selects and joins the appropriate processing partition. This scheduling method not only optimizes computing power allocation and makes full use of computing resources, but also improves task response speed and achieves efficient computing power resource scheduling.

[0042] In summary, through innovative adaptive partition scheduling and blockchain connection methods, a decentralized, efficient, and secure data processing system is realized, which has stronger robustness and flexibility and is suitable for processing large-scale and complex data tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of a bare metal server computing power scheduling platform according to Embodiment 1 of the present invention;

[0044] Figure 2 This is a flow chart of a bare metal server computing power scheduling method according to Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Example 1

[0047] See also Figure 1 As shown, the bare metal server computing power scheduling platform described in this embodiment includes a partitioning module, a first analysis module, a second analysis module and a third analysis module, and each module is connected by wire and / or wirelessly.

[0048] Partitioning module, when the r-th server receives the task data sent by the user end, it sets the network delay condition according to the task data; the r-th server records the network delay between the R servers communicating with each other, and divides the servers within the network delay condition into a processing partition, r∈R, and connects the servers in the same processing partition through the blockchain to form a partitioned blockchain; the network delay condition includes a network speed threshold and a network speed standard deviation threshold, and the network speed threshold and the network speed standard deviation threshold are both preset values.

[0049] Methods for dividing processing partitions include:

[0050] The network speed when R servers communicate with each other in a unit time is recorded, and R network speed analysis sets are established one by one. The average network speed and standard deviation of the network speed analysis sets are calculated.

[0051] The servers whose average network speed is within the network speed threshold range and whose network speed standard deviation is within the network speed standard deviation threshold range are divided into a processing partition; this ensures that the communication network of all servers in a processing partition is within the set network delay condition range, which facilitates efficient processing of task data between each other.

[0052] The method for setting the network delay condition according to the task data includes:

[0053] The total data size, total data identity ID, data type and data format of the task data are input into the pre-trained condition setting model to obtain the network delay condition.

[0054] The training methods for conditioning models include:

[0055] K groups of condition evaluation data are collected in advance, the condition evaluation data include condition setting feature data and network delay conditions corresponding to the condition setting feature data, the condition setting feature data include total data size, total data identity ID, data type and data format.

[0056] The k groups of conditional evaluation data are converted into feature vectors and divided into training set and test set. The conditional setting feature data in the training set is used as the input of the conditional setting model, and the network delay conditions in the training set are used as the output of the conditional setting model. The conditional setting model is tested with the test set, and the conditional setting model that meets the prediction accuracy is output. The conditional setting model is a deep neural network (DNN) or a gradient boosting decision tree model.

[0057] It should be noted that when training the condition setting model, when selecting the condition evaluation data, the network delay condition corresponding to the condition setting feature data is optimal, that is, the network delay condition corresponding to the shortest processing time, which is convenient for screening out servers that meet the network delay conditions and building processing partitions.

[0058] The first analysis module marks the server that receives the user-side task data as a receiving server. The receiving server extracts feature data of the task data, inputs the feature data into a pre-trained split evaluation model, obtains a split evaluation result, and the split evaluation result includes splittable and non-splittable. Task description data is generated according to the split evaluation result, and the task description data is published on the partition blockchain. At this time, any server in the same processing partition can know the task description data, which is convenient for any server in the same processing partition to estimate the required processing time of the task processing data corresponding to the task description data according to its own real-time load data.

[0059] The characteristic data includes data type, data format, data dependency, data structure complexity and data integrity constraints.

[0060] Data types include structured data (such as relational databases, CSV), semi-structured data (such as JSON, XML) and unstructured data (such as text, images, audio). The data type affects the splitting method; data formats include CSV, JSON, XML, images, audio and video. The specific data format helps to determine the parsing and splitting methods; data dependencies include no dependency (data points are independent of each other), sequential dependency (processing order is important) and hierarchical dependency (such as parent-child relationship). The dependencies within the data affect the feasibility of splitting; data structure complexity uses nesting levels and the number of fields to quantify the complexity of the data structure; data integrity constraints include no special constraints and existence of constraints, which need to be processed as a whole. Whether there are data integrity or constraint requirements affects the splitting.

[0061] The training method of the split evaluation model includes:

[0062] Collect k groups of evaluation data in advance, where k is an integer greater than 1. The evaluation data includes feature data and split evaluation results corresponding to the feature data. The evaluation data is converted into feature vectors, for example:

[0063] Data type: 0 (structured data), Data format: 0 (CSV), Data dependency: 0 (no dependency), Data structure complexity: 15, Processing task type: 0 (independent calculation), Task parallelism: 1 (parallelizable), Sequential processing requirement: 0 (no sequential processing required), Data partition feasibility: 0.9, Split evaluation result: 1 (splittable).

[0064] Each group of feature vectors is used as the input of the split evaluation model, and the split evaluation model takes a group of predicted split evaluation results corresponding to each group of feature data as output, and takes the actual split evaluation results corresponding to each group of feature data as the prediction target; minimizing the sum of prediction errors of all feature data is used as the training target; the split evaluation model is trained until the sum of prediction errors reaches convergence and the training is stopped.

[0065] The split evaluation model is specifically a decision tree model, a random forest model or a gradient boosting tree model; it includes an input layer, a hidden layer and an output layer; each hidden layer includes multiple neurons, each neuron is connected to the neurons in the next layer, and the connection contains weights, which determine the importance and influence of data transmission in the neural network; an activation function is applied to each neuron between the hidden layer and the output layer, and the activation function introduces nonlinearity, allowing the network to learn more complex patterns and features.

[0066] It should be noted that publishing the task description data on the partitioned blockchain does not mean publishing the task data on the partitioned blockchain. The task description data is a text description of the task data. The task description data is very small compared to the task data, which can reduce the occupancy of parallel transmission resources.

[0067] Methods for generating task description data based on split evaluation results include:

[0068] If the splitting evaluation result is not splittable, the task description data includes the user terminal ID, total data size, total data identity ID, data type and data format. The total data size is in bytes and is used to evaluate the load required for processing.

[0069] If the split evaluation result is that it can be split, the task data is divided into q parts to obtain q sub-data, where q is an integer greater than 1 and is a preset value. The value of q is preferably half of the number of servers in the processing partition. If half of the number is not an integer, it is rounded down. Too many splits will increase the overhead of parallel network transmission and affect transmission efficiency. The task description data includes the user terminal ID, sub-data size, sub-data identity ID, data type, and data format. The sub-data identity ID includes the sub-data unique identity number and the position of the sub-data in the task data (to facilitate the user terminal to merge the sub-data that has been processed later).

[0070] In the second analysis module, each server in the same processing partition inputs its real-time load data and the task description data into the pre-trained task data evaluation model (the task data evaluation model is pre-deployed in R servers), obtains the required processing time, and publishes the required processing time on the partition blockchain. Similarly, at this time, any server in the same processing partition can know it.

[0071] It should be noted that publishing the task description data on the partitioned blockchain does not mean publishing the task data on the partitioned blockchain. The task description data is a text description of the task data. The task description data is very small compared to the task data, which can reduce the occupation of network resources.

[0072] If the task data is indivisible data, the required processing time for each server in the same processing partition is one, that is, each server processes the entire task data according to its own real-time load data, and the predicted required processing time is published on the partition blockchain.

[0073] If the task data is divisible data, the required processing time for each server in the same processing partition is q, and the q required processing times correspond one to one to the required processing time when processing q pieces of data; when each server processes q pieces of data according to its own real-time load data, the predicted q required processing times are published on the partition blockchain.

[0074] The training method of the task data evaluation model includes:

[0075] Collect H groups of task evaluation data in advance. The task evaluation data includes evaluation feature data and the required processing time corresponding to the evaluation feature data. The evaluation feature data includes load data and task description data. The load data includes CPU data, memory usage, total memory capacity, hard disk utilization, total hard disk capacity and other data. The CPU data includes CPU utilization, clock speed, number of cores, number of threads, cache capacity, floating-point computing capability and other data.

[0076] The H group of task evaluation data are converted into feature vectors and divided into training set and test set. The evaluation feature data in the training set is used as the input of the task data evaluation model, and the required processing time in the training set is used as the output of the task data evaluation model. The task data evaluation model is tested with the test set, and the task data evaluation model that meets the prediction accuracy is output. The task data evaluation model uses a deep neural network (DNN) or a gradient boosting decision tree model.

[0077] The third analysis module records the required processing time for each server in the same processing partition, compares the required processing time published by each server with the required processing time published by other servers, and the server whose required processing time is not the minimum exits the processing partition and performs the next round of re-partitioning, thereby reducing the occupied time of servers that do not meet the requirements and reducing the waste of server resources; the server corresponding to the minimum required processing time is marked as a processing server, the processing server receives the corresponding task data sent by the receiving server, and the processing server sends the processed data to the user terminal; specifically, the processing server sends the processed data to the corresponding user terminal according to the user terminal ID.

[0078] If the task data is divisible, the user terminal first merges the processed data (the user terminal merges the sub-data according to their positions in the task data) and then displays it on the user terminal; if the task data is indivisible, this step can be omitted and the data can be displayed directly.

[0079] It should be noted that, for “marking the server corresponding to the minimum required processing time as a processing server, and the processing server receiving the corresponding task data sent by the receiving server”, further examples are given as follows:

[0080] If the task data is not divisible, the number of processing servers is any server in the same processing partition. If the task data is divisible, the number of sub-data is q, and the number of processing servers is also q. For example, when q is 3, the number of processing servers is also 3.

[0081] The first processing server takes 1 minute to process the first sub-data prediction, 2 minutes to process the second sub-data prediction, and 3 minutes to process the third sub-data prediction. The second processing server takes 2 minutes to process the first sub-data prediction, 1 minute to process the second sub-data prediction, and 3 minutes to process the third sub-data prediction. The third processing server takes 3 minutes to process the first sub-data prediction, 2 minutes to process the second sub-data prediction, and 3 minutes to process the first sub-data prediction.

[0082] The first sub-data is sent to the first processing server for processing, the second sub-data is sent to the second processing server for processing, and the third sub-data is sent to the third processing server for processing, so as to process the task data in the shortest time and improve the processing efficiency of the task data.

[0083] The decentralized nature of the platform avoids relying on a central server for computing power scheduling. The server adaptively sets the network delay conditions according to the task data, thereby automatically forming processing partitions. Even if a server fails, other servers can continue to execute the scheduling task, greatly reducing the risk of single point failure. This structural design ensures the high availability and continuity of computing power scheduling; according to the network delay conditions of the task, qualified servers are organized into a processing partition. Servers in the same processing partition are interconnected through blockchain, and determine whether they can be split for processing based on the characteristic data of the task. The adaptive partitioning and judgment mechanism realizes flexible computing power scheduling, ensuring that each partition operates under low latency and sufficient resources, greatly improving data processing efficiency.

[0084] In the processing partition, the server predicts the required processing time based on its own real-time load and task description data, and publishes the prediction results on the blockchain. Each server compares the required processing time. The server whose processing time is not the shortest will automatically exit the current partition. The competitive elimination mechanism ensures that the optimal server for processing the task is responsible for execution, reduces resource waste, avoids inefficient servers occupying resources, and improves the efficiency of the entire partition.

[0085] When the task data is divisible, the data is automatically divided and distributed to multiple servers for parallel processing. After the processing is completed, the user end automatically merges the processed data to achieve rapid display. The split and parallel processing mechanism is particularly suitable for big data tasks, significantly shortening the processing time and achieving optimal resource utilization. The blockchain is used to connect servers in the same processing partition to achieve decentralized task time prediction and release. The processing time information of each server is transparent and open, and is updated in real time on the blockchain, preventing information tampering and ensuring the openness and fairness of task scheduling.

[0086] Decentralized blockchain connections and dynamic adaptive allocation of processing partitions reduce the risk of task data loss. When a server exits a processing partition due to a failure, other servers can quickly reorganize a new processing partition and continue processing tasks, ensuring the security and integrity of task data. Through adaptive scheduling, the server evaluates its own computing power and task requirements in real time, automatically selects and joins the appropriate processing partition. This scheduling method not only optimizes computing power allocation and makes full use of computing resources, but also improves task response speed and achieves efficient computing power resource scheduling.

[0087] In summary, through innovative adaptive partition scheduling and blockchain connection methods, a decentralized, efficient, and secure data processing system is realized, which has stronger robustness and flexibility and is suitable for processing large-scale and complex data tasks.

[0088] Example 2

[0089] See also Figure 2 As shown, the present application also provides a bare metal server computing power scheduling method, including:

[0090] When the rth server receives the task data sent by the user, it sets the network delay condition according to the task data; the rth server records the network delay between the R servers, and divides the servers within the network delay condition into a processing partition, r∈R, and connects the servers in the same processing partition through the blockchain to form a partitioned blockchain;

[0091] The server receiving the task data of the user terminal is marked as a receiving server, the receiving server extracts feature data of the task data, inputs the feature data into a pre-trained split evaluation model, obtains a split evaluation result, the split evaluation result includes splittable and non-splittable, generates task description data according to the split evaluation result, and publishes the task description data to the partition blockchain;

[0092] Each server in the same processing partition inputs its real-time load data and the task description data into the pre-trained task data evaluation model to obtain the required processing time, and publishes the required processing time on the partition blockchain;

[0093] Each server in the same processing partition records the required processing time, and the required processing time published by each server is compared with the required processing time published by other servers. The server whose required processing time is not the shortest exits the processing partition; the server corresponding to the shortest required processing time is marked as a processing server, and the processing server receives the corresponding task data sent by the receiving server, and the processing server sends the processed data to the user end.

[0094] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0095] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. Bare metal server computing power scheduling platform, characterized by: include: Partitioning module, when the rth server receives the task data sent by the user end, it sets the network delay condition according to the task data; the rth server records the network delay between the R servers communicating with each other, and divides the servers within the network delay condition into a processing partition, r∈R, and connects the servers in the same processing partition through the blockchain to form a partitioned blockchain; The method for setting the network delay condition according to the task data includes: Inputting the total data size, total data identity ID, data type and data format of the task data into a pre-trained condition setting model to obtain a network delay condition; The first analysis module marks the server that receives the task data of the user end as a receiving server, and the receiving server extracts feature data of the task data, inputs the feature data into a pre-trained split evaluation model, obtains a split evaluation result, and the split evaluation result includes splittable and non-splittable, generates task description data according to the split evaluation result, and publishes the task description data to the partition blockchain; The characteristic data includes data type, data format, data dependency, data structure complexity and data integrity constraints; Methods for generating task description data based on split evaluation results include: If the split evaluation result is that the task cannot be split, the task description data includes the user terminal ID, the total data size, the total data identity ID, the data type and the data format; If the splitting evaluation result is splittable, the task data is divided into q parts to obtain q sub-data, where q is an integer greater than 1 and q is a preset value; the task description data includes the user terminal ID, sub-data size, sub-data identity ID, data type, and data format, and the sub-data identity ID includes the sub-data unique identity number and the position of the sub-data in the task data; In the second analysis module, each server in the same processing partition inputs its real-time load data and the task description data into the pre-trained task data evaluation model to obtain the required processing time, and publishes the required processing time on the partition blockchain; The third analysis module records the required processing time for each server in the same processing partition, compares the required processing time published by each server with the required processing time published by other servers, and the server whose required processing time is not the shortest exits the processing partition; marks the server corresponding to the shortest required processing time as a processing server, the processing server receives the corresponding task data sent by the receiving server, and the processing server sends the processed data to the user end.

2. The bare metal server computing power scheduling platform according to claim 1, characterized in that: If the task data is indivisible data, the required processing time for each server in the same processing partition is one; If the task data is divisible data, the required processing time for each server in the same processing partition is q, and the q required processing times correspond one-to-one to the required processing time when processing q pieces of data.

3. The bare metal server computing power scheduling platform according to claim 2, characterized in that: The method for the processing server to receive the corresponding task data sent by the receiving server is: The receiving server determines the sub-data according to the required processing time corresponding to each processing server, and sends the sub-data to the corresponding processing server, and the required processing time corresponds to the sub-data one by one.

4. The bare metal server computing power scheduling platform according to claim 1, characterized in that: The network delay condition includes a network speed threshold and a network speed standard deviation threshold, and both the network speed threshold and the network speed standard deviation threshold are preset values; Methods for dividing processing partitions include: Record the network speed when R servers communicate with each other within a unit time, establish R network speed analysis sets one by one, and calculate the network speed average and network speed standard deviation of the network speed analysis sets.

5. The bare metal server computing power scheduling platform according to claim 1, characterized in that: The training methods for conditioning models include: Collect k groups of condition evaluation data in advance, the condition evaluation data including condition setting feature data and network delay conditions corresponding to the condition setting feature data, the condition setting feature data including total data size, total data identity ID, data type and data format; Convert k groups of conditional evaluation data into feature vectors and divide them into training set and test set. Use the conditional setting feature data in the training set as the input of the conditional setting model, use the network delay conditions in the training set as the output of the conditional setting model, use the test set to test the conditional setting model, and output the conditional setting model that meets the prediction accuracy. The conditional setting model is a deep neural network or a gradient boosting decision tree model.

6. The bare metal server computing power scheduling platform according to claim 1, characterized in that: The training method of the split evaluation model includes: Collect k groups of evaluation data in advance, where k is an integer greater than 1, and the evaluation data includes feature data and split evaluation results corresponding to the feature data, and convert the evaluation data into feature vectors; Each group of feature vectors is used as the input of the split evaluation model, and the split evaluation model takes a group of predicted split evaluation results corresponding to each group of feature data as output, and takes the actual split evaluation results corresponding to each group of feature data as the prediction target; minimizing the sum of prediction errors of all feature data is used as the training target; the split evaluation model is trained until the sum of prediction errors reaches convergence, and the training is stopped; the split evaluation model is a decision tree model, a random forest model or a gradient boosting tree model.

7. The bare metal server computing power scheduling platform according to claim 1, characterized in that: The training method of the task data evaluation model includes: Collect H groups of task evaluation data in advance. The task evaluation data includes evaluation feature data and required processing time corresponding to the evaluation feature data. The evaluation feature data includes load data and task description data. The load data includes CPU data, memory usage, total memory capacity, hard disk utilization and total hard disk capacity. The CPU data includes CPU utilization, clock speed, number of cores, number of threads, cache capacity and floating-point computing capability. The H group of task evaluation data are converted into feature vectors and divided into training set and test set. The evaluation feature data in the training set is used as the input of the task data evaluation model, and the required processing time in the training set is used as the output of the task data evaluation model. The task data evaluation model is tested with the test set, and the task data evaluation model that meets the prediction accuracy is output. The task data evaluation model is a deep neural network or a gradient boosting decision tree model.

8. The bare metal server computing power scheduling platform according to claim 1, characterized in that: The processing server sends the processed data to the corresponding user terminal according to the user terminal ID.

9. The bare metal server computing power scheduling platform according to claim 1, characterized in that: If the task data is divisible, the user terminal merges the sub-data according to their positions in the task data, and then displays them on the user terminal.

Images