Cloud computing system based on deep learning and big data

Abstract

The application discloses a cloud computing system based on deep learning and big data, relates to the technical field of cloud platform task scheduling, and sets a task collection module to collect task data received by a cloud platform and classify tasks according to processors used; sets a cloud platform data collection module to collect data storage spaces, CPU computing power and GPU computing power of the cloud platform in advance; sets a historical data collection module to collect historical task processing data of the cloud platform; sets a deep reinforcement learning model training module to train a deep reinforcement learning model capable of making decisions on task selection according to the historical task processing data of the cloud platform; and sets a task scheduling module to use the trained deep reinforcement learning model to make decisions on tasks of the cloud platform; and the problems of intelligent scheduling of tasks of the cloud platform are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of cloud computing, and relates to a deep reinforcement learning technology, in particular to a cloud computing system based on deep learning and big data. BACKGROUND

[0002] At present, cloud computing technology is one of the hottest topics in the field of computer services. Large enterprises such as IBM and Google and small private enterprises and even some technology personnel who are willing to pursue new technologies are deploying or researching cloud computing in order to consolidate or enhance their position in the industry. In the cloud computing environment, virtualization technology is used to virtualize the server as a whole into a data resource pool. Since the data resources are various and large in scale, cloud computing data resource scheduling has become one of the hotspots of cloud computing research.

[0003] Most of the current task scheduling algorithms use fixed scheduling methods such as first-come-first-served or short-job-first. It is difficult to meet the increasingly complex task scheduling conditions. Considering that each task scheduling of the cloud platform is a decision, the use of deep reinforcement learning can better train a task scheduling decision model.

[0004] Therefore, the cloud computing system based on deep learning and big data is proposed. SUMMARY

[0005] The application aims to at least solve one of the technical problems existing in the prior art. To this end, the cloud computing system based on deep learning and big data is proposed, which sets a task collection module to collect task data received by the cloud platform and classify tasks according to the used processors; sets a cloud platform data collection module to collect data storage space, CPU computing power and GPU computing power of the cloud platform in advance; sets a historical data collection module to collect historical task processing data of the cloud platform; sets a deep reinforcement learning model training module to train a deep reinforcement learning model capable of making decisions on task selection according to the historical task processing data of the cloud platform; and sets a task scheduling module to use the trained deep reinforcement learning model to make decisions on task selection of the cloud platform, thereby solving the problem of intelligent scheduling of tasks of the cloud platform.

[0006] To achieve the above-mentioned purpose, the cloud computing system based on deep learning and big data according to the first aspect of the application is proposed, which comprises a task collection module, a cloud platform data collection module, a historical data collection module, a deep reinforcement learning model training module and a task scheduling module; wherein each module is connected through an electrical way.

[0007] The task collection module is mainly used to collect task data received by the cloud platform and classify tasks.

[0008] The task collection module divides all tasks into two categories; specifically, tasks using CPU operations are marked as class C; tasks using GPU operations are marked as class G; further, all class C tasks are saved in task set C; all class G tasks are saved in task set G; tasks in task set C are marked as c, and tasks in task set G are marked as g; for task c, the storage space required is marked as Mc; the required computing power size is marked as Cc; the processing time is marked as Tc; for task g, the storage space required is marked as Mg; the required computing power size is marked as Cg; the processing time is marked as Tg; the task collection module sends the collected tasks in task set C and task set G to the task scheduling module;

[0009] The cloud platform data collection module is mainly used to collect the basic data of the cloud platform in advance;

[0010] The basic data collected in advance by the cloud platform data collection module includes: the upper limit of the storage space of the cloud platform, the upper limit of the CPU unit time operation computing power, and the upper limit of the GPU unit time operation computing power; the upper limit of the storage space of the cloud platform is marked as Mmax; the CPU unit time operation computing power is marked as Cmax; and the GPU unit time operation computing power is marked as Gmax;

[0011] The cloud platform data collection module sends the collected cloud platform basic data to the deep reinforcement learning model training module;

[0012] The historical data collection module is mainly used to collect historical task processing data of the cloud platform;

[0013] The task processing data collected in advance by the historical data collection module includes data processing task information received by the cloud platform in the past and cloud platform information; the data processing task information collected by the historical data collection module includes the type of each task, the task arrival time, the storage space required by the task, the unit time calculation computing power required by the task, and the calculation time required by the task; historical tasks are saved in the form of a task set; when a new task arrives at the cloud platform, the cloud platform first adds it to the task set; when a task is called and executed by the cloud platform from the task set, it is deleted from the task set;

[0014] The cloud platform information includes the remaining storage space of the cloud platform at each task arrival time, the CPU remaining unit time processing computing power, the GPU remaining unit time processing computing power, and the storage space consumed by each task being processed, the unit time computing power, and the remaining processing time; the remaining storage space of the cloud platform is marked as Rm, the CPU remaining unit time processing computing power is marked as Rc, and the GPU remaining unit time processing computing power is marked as Rg;

[0015] The historical data collection module sends the collected cloud platform historical task processing data to the deep reinforcement learning model training module;

[0016] The deep reinforcement learning model training module is mainly used for training a deep reinforcement learning model for selecting an optimal data processing task from a task set for processing;

[0017] The deep reinforcement learning model needs to generate a reward function for the result of each decision; the decision is to select a plurality of data processing tasks from the task set for processing; wherein the reward function includes a reward and a penalty;

[0018] The reward generated by the deep reinforcement learning model training module for each decision can include the number of selected data processing tasks and the total waiting time of selected data processing tasks; the selected data processing task set is marked as S; the number of tasks in the task set is marked as N; the waiting time of each task c or g in the task set S is the current time minus the time when it arrives at the cloud platform; the waiting time of tasks c and g is marked as Wc and Wg respectively; then the total waiting time W of all tasks in the task set S is W = ∑ c∈S Wc+∑ g∈S Wg; then the reward function J can be represented as J = a*N+b*W; wherein a and b are preset reward coefficients;

[0019] The penalty generated by the deep reinforcement learning model training module for each decision can include: the penalty when the total required storage space of the tasks in the task set S exceeds the remaining storage space Rm of the cloud platform, the penalty when the total CPU computing power requirement exceeds the remaining CPU computing power per unit time Rc of the cloud platform, the penalty when the total GPU computing power requirement exceeds the remaining GPU computing power per unit time Rg of the cloud platform, the total waiting time of the remaining tasks and the average size of the remaining tasks; wherein the average size of the remaining tasks is marked as V; then the calculation formula of the average size V is Wherein o and k are both preset proportion coefficients; it can be understood that when the value of the average size V is larger, it means that the sizes of the remaining tasks are larger, so the penalty is heavier, and therefore the compensation for completing the data processing tasks with larger sizes is completed;

[0020] The deep reinforcement learning model training module uses the difference between the reward function and the penalty function as the training target, uses the historical task data sent by the historical data collection module to train the deep reinforcement learning model, and marks the trained deep reinforcement learning model as M; the deep reinforcement learning model training module sends the trained deep reinforcement learning model M to the task scheduling module;

[0021] The task scheduling module mainly uses a deep reinforcement learning model M to select optimal tasks from a task set for processing.

[0022] When the task scheduling module receives the deep reinforcement learning model M, and the cloud platform completes a task, leaves out spare storage space, CPU computing power and GPU computing power, the deep reinforcement learning model M is used to select optimal task combinations from the task set, and the CPU and GPU of the cloud platform are used for processing according to the types of the tasks.

[0023] Compared with the prior art, the present application has the following advantages:

[0024] The task collection module is arranged to collect task data received by the cloud platform and classify the tasks according to the processors used; the cloud platform data collection module is arranged to collect the data storage space, CPU computing power and GPU computing power of the cloud platform in advance; the historical data collection module is arranged to collect historical task processing data of the cloud platform; the deep reinforcement learning model training module is arranged to train a deep reinforcement learning model capable of making decisions on task selection according to the historical task processing data of the cloud platform; and the task scheduling module is arranged to use the trained deep reinforcement learning model to make decisions on the tasks of the cloud platform; thereby solving the problem of intelligent scheduling of tasks of the cloud platform. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The present application is a schematic diagram. DETAILED DESCRIPTION

[0026] The technical solutions of the present application will be described in detail below with reference to the embodiments, obviously, the described embodiments are only part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] As shown in Figure 1 The cloud computing system based on deep learning and big data includes a task collection module, a cloud platform data collection module, a historical data collection module, a deep reinforcement learning model training module and a task scheduling module; wherein the modules are connected by electrical means.

[0028] The task collection module is mainly used to collect task data received by the cloud platform and classify the tasks.

[0029] It can be understood that each task received by the cloud platform has the needs of data storage, computing power and computing time length; and each task uses a CPU and a GPU for selection; when each task is sent to the cloud platform, the data storage space required by the task, the type of processor used, the size of the computing power required and the processing time length can be estimated;

[0030] In a preferred embodiment, the task collection module divides all tasks into two categories; specifically, tasks using CPU operations are marked as class C; tasks using GPU operations are marked as class G; further, all class C tasks are saved in task set C; all class G tasks are saved in task set G; tasks in task set C are marked as c, and tasks in task set G are marked as g; for task c, the storage space required is marked as Mc; the size of the computing power required is marked as Cc; the processing time length is marked as Tc; for task g, the storage space required is marked as Mg; the size of the computing power required is marked as Cg; the processing time length is marked as Tg; the task collection module sends the tasks in task set C and task set G collected to the task scheduling module;

[0031] The cloud platform data collection module is mainly used for pre-collection of basic data of the cloud platform;

[0032] It can be understood that the main function of the cloud platform is to process received data processing tasks; the cloud platform needs a certain data storage space, computing power and computing time length to process each task; and the data storage space and computing power of the cloud platform are limited;

[0033] In a preferred embodiment, the basic data pre-collected by the cloud platform data collection module includes: an upper limit of the storage space of the cloud platform, an upper limit of the CPU operation power per unit time and an upper limit of the GPU operation power per unit time; the upper limit of the storage space of the cloud platform is marked as Mmax; the CPU operation power per unit time is marked as Cmax; and the GPU operation power per unit time is marked as Gmax;

[0034] The cloud platform data collection module sends the collected basic data of the cloud platform to the deep reinforcement learning model training module;

[0035] The historical data collection module is mainly used for collection of historical task processing data of the cloud platform;

[0036] In a preferred embodiment, the historical data collection module pre-collects task processing data including historical data processing task information received by the cloud platform and cloud platform information; specifically, the historical data collection module collects data processing task information including the type of each task, task arrival time, storage space required by the task, computing power required by the task per unit time, and computing time required by the task; the historical task is saved in the form of a task set; when a new task arrives at the cloud platform, the cloud platform first adds it to the task set; when the task is called and executed by the cloud platform from the task set, it is deleted from the task set;

[0037] The cloud platform information includes the remaining storage space of the cloud platform at each task arrival time, the CPU remaining unit time processing power, the GPU remaining unit time processing power, and the storage space consumed by each task being processed, the unit time power, and the remaining processing time; the remaining storage space of the cloud platform is marked as Rm, the CPU remaining unit time processing power is marked as Rc, and the GPU remaining unit time processing power is marked as Rg;

[0038] The historical data collection module sends the collected cloud platform historical task processing data to the deep reinforcement learning model training module;

[0039] The deep reinforcement learning model training module is mainly used to train a deep reinforcement learning model for selecting the optimal data processing task from the task set for processing;

[0040] It can be understood that the deep reinforcement learning model needs to generate a reward function for the result of each decision; the decision is to select several data processing tasks from the task set for processing at the same time; wherein the reward function includes reward and punishment;

[0041] In a preferred embodiment, the reward generated by the deep reinforcement learning model training module for each decision can include the number of selected data processing tasks and the total waiting time of selected data processing tasks; the selected data processing task set is marked as S; the number of tasks in the task set is marked as N; the waiting time of each task c or g in the task set S is the current time minus the time when it arrives at the cloud platform; The waiting time of tasks c and g is marked as Wc and Wg respectively; then the total waiting time W of all tasks in the task set S is W = ∑ c∈S Wc+∑ g∈S Wg; then the reward function J can be represented as J = a*N+b*W; wherein a and b are preset reward coefficients;

[0042] It can be understood that the reward setting will preferentially select data processing tasks with long waiting time and small task size from the task set; it can cause data processing tasks with large task size to be unable to be executed; therefore, compensation is made by setting a penalty function for each decision;

[0043] The penalty generated by the deep reinforcement learning model training module for each decision can include: a penalty when the total demand storage space of the tasks in the task set S exceeds the remaining storage space Rm of the cloud platform, a penalty when the total CPU computing power demand exceeds the remaining CPU unit time computing power Rc of the cloud platform, a penalty when the total GPU computing power demand exceeds the remaining GPU unit time computing power Rg of the cloud platform, the total length of time that the remaining tasks need to wait, and the average size of the remaining tasks; wherein the average size of the remaining tasks is marked as V; the calculation formula of the average size V is Wherein o and k are both preset proportion coefficients; it can be understood that when the value of the average size V is large, it indicates that the sizes of the remaining tasks are large, and therefore the penalty is heavier, and thus compensation for completing data processing tasks with large task size is made;

[0044] The deep reinforcement learning model training module uses the historical task data sent by the historical data collection module to train the deep reinforcement learning model, and marks the trained deep reinforcement learning model as M; the deep reinforcement learning model training module sends the trained deep reinforcement learning model M to the task scheduling module;

[0045] The task scheduling module mainly uses the deep reinforcement learning model M to select the optimal tasks from the task set for processing;

[0046] In a preferred embodiment, after the task scheduling module receives the deep reinforcement learning model M, when the cloud platform completes a task and leaves empty storage space, CPU computing power, and GPU computing power, the task scheduling module uses the deep reinforcement learning model M to select the optimal task combination from the task set, and processes the tasks using the CPU and GPU of the cloud platform according to the types of the tasks.

[0047] The above embodiments are only used to illustrate the technical method of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A cloud computing system based on deep learning and big data, characterized in that, It includes a task collection module, a cloud platform data collection module, a historical data collection module, a deep reinforcement learning model training module, and a task scheduling module; the modules are connected electrically. The task collection module is used to collect task data received by the cloud platform and classify the tasks; the task collection module sends the tasks in the collected task set C and task set G to the task scheduling module; The cloud platform data collection module is used to collect basic data of the cloud platform in advance; the cloud platform data collection module sends the collected basic data of the cloud platform to the deep reinforcement learning model training module; The historical data collection module is used to collect historical task processing data of the cloud platform; the task processing data collected in advance by the historical data collection module includes data processing task information received by the cloud platform in the past and cloud platform information; the historical data collection module sends the collected historical task processing data of the cloud platform to the deep reinforcement learning model training module. The deep reinforcement learning model training module is used to train a deep reinforcement learning model that selects the optimal data processing task from the task set for processing; the deep reinforcement learning model training module sends the trained deep reinforcement learning model M to the task scheduling module; The task scheduling module uses a deep reinforcement learning model M to select several optimal tasks from the task set for processing. The task collection module categorizes all tasks into two types: tasks using CPU computation are labeled as type C, and tasks using GPU computation are labeled as type G. Further, all type C tasks are stored in task set C, and all type G tasks are stored in task set G. Tasks in task set C are labeled c, and tasks in task set G are labeled g. For task c, its required storage space is labeled Mc, its required computing power is labeled Cc, and its processing time is labeled Tc. For task g, its required storage space is labeled Mg, its required computing power is labeled Cg, and its processing time is labeled Tg. The basic data pre-collected by the cloud platform data collection module includes: the cloud platform's storage space limit, CPU's computing power limit per unit time, and GPU's computing power limit per unit time; the cloud platform's storage space limit is marked as Mmax; CPU's computing power limit per unit time is marked as Cmax; and GPU's computing power limit per unit time is marked as Gmax. The deep reinforcement learning model training module aims to maximize the difference between reward and punishment, and uses historical task data sent by the historical data collection module to train the deep reinforcement learning model; and marks the trained deep reinforcement learning model as M. The reward generated by each decision in the deep reinforcement learning model training module includes the number of selected data processing tasks and the total waiting time of the selected data processing tasks. The set of selected data processing tasks is labeled S; the number of tasks in the task set is labeled N; the waiting time of each task c or g in task set S is the current time minus its arrival time on the cloud platform; the waiting times of tasks c and g are labeled Wc and Wg, respectively; then the total waiting time W of all tasks in task set S is... The reward function J is then expressed as J = a * N + b * W, where a and b are preset reward coefficients. The penalties generated by the deep reinforcement learning model training module for each decision include: penalties when the total storage space required by the tasks in the task set S exceeds the remaining storage space Rm of the cloud platform; penalties when the total CPU computing power required exceeds the remaining CPU computing power per unit time Rc of the cloud platform; penalties when the total GPU computing power required exceeds the remaining GPU computing power per unit time Rg of the cloud platform; the total waiting time for the remaining tasks; and the average size of the remaining tasks; wherein the average size of the remaining tasks is denoted as V; and the formula for calculating the average size V is as follows: Where o and k are both preset proportional coefficients.

2. The cloud computing system based on deep learning and big data according to claim 1, characterized in that, The historical data collection module collects data processing task information including the type of each task, the task arrival time, the storage space required by the task, the computing power required per unit time, and the computing time required by the task; historical tasks are saved in the form of task sets; when a new task arrives at the cloud platform, the cloud platform first adds it to the task set; when a task is retrieved and executed by the cloud platform from the task set, it is deleted from the task set. The cloud platform information includes the remaining storage space, CPU remaining processing power per unit time, GPU remaining processing power per unit time, and the storage space, processing power per unit time, and remaining processing time consumed by each task being processed at the time of each task arrival. The remaining storage space of the cloud platform is marked as Rm, the remaining CPU processing power per unit time is marked as Rc, and the remaining GPU processing power per unit time is marked as Rg.

3. The cloud computing system based on deep learning and big data according to claim 1, characterized in that, After receiving the deep reinforcement learning model M, the task scheduling module completes a task on the cloud platform. When there is spare storage space, CPU computing power, and GPU computing power, it uses the deep reinforcement learning model M to select the optimal task combination from the task set and uses the cloud platform's CPU and GPU to process the task according to its type.

Images