WO2023083058A1

WO2023083058A1 - Scheduling parameter adjusting method, devices, and storage medium

Info

Publication number: WO2023083058A1
Application number: PCT/CN2022/129029
Authority: WO
Inventors: 裘瑞涛; 金士英; 刘涛; 王永成; 韩炳涛; 屠要峰
Original assignee: 中兴通讯股份有限公司
Priority date: 2021-11-12
Filing date: 2022-11-01
Publication date: 2023-05-19
Also published as: CN114064242A

Abstract

Embodiments of the present application relate to the technical field of computers, and provide a scheduling parameter adjusting method, devices, and a storage medium. The scheduling parameter adjusting method is applied to a main control device, and comprises: searching for an operator scheduling template matched with a target device; generating a scheduling parameter according to the matched operator scheduling template and a scheduling parameter search algorithm, and sending the scheduling parameter to the target device, so that the target device runs, according to the scheduling parameter, a scheduling process corresponding to an operator; and receiving performance data fed back by the target device and used for executing the scheduling process, adjusting the scheduling parameter according to the performance data, and sending the scheduling parameter to the target device.

Description

Adjustment method, equipment and storage medium of scheduling parameters

technical field

The embodiments of the present application relate to the field of computer technology, and in particular, to a scheduling parameter adjustment method, device, and storage medium.

Background technique

With the great success of deep learning technology in computer vision, speech recognition, natural language processing and other fields in recent years, the industry has also begun to use various types of hardware, such as central processing unit (central processing unit, CPU), graphics processing, etc. Graphics processing unit (GPU), smart chips, etc., and gradually deploy deep learning model inference-related services. Among them, only when the computing resources and storage resources of the hardware are reasonably and fully scheduled, the inference delay and throughput of the deep learning model Performance indicators such as efficiency can be better improved.

However, in some cases, the scheduling optimization of deep learning model reasoning is mainly done manually, which usually cannot achieve optimal scheduling, and the efficiency is very low, and resources cannot be allocated quickly and efficiently.

Contents of the invention

The main purpose of the embodiments of this application is to propose a scheduling parameter adjustment method, device, and storage medium, aiming to overcome the dependence on manual work in various application scenarios, and automatically implement the operator scheduling process of any target device. Scheduling design, to obtain the optimal scheduling parameters, and more efficient and faster.

In order to at least achieve the above purpose, an embodiment of the present application provides a method for adjusting scheduling parameters, which is applied to the main control device, including: searching for an operator scheduling template that matches the target device; and scheduling according to the matched operator scheduling template and The parameter search algorithm generates scheduling parameters, and sends the scheduling parameters to the target device for the target device to run the scheduling process corresponding to the operator according to the scheduling parameters; execute the scheduling process after receiving feedback from the target device performance data, adjusting the scheduling parameter according to the performance data and sending it to the target device.

In order to at least achieve the above purpose, the embodiment of the present application also proposes a scheduling parameter adjustment method, which is applied to the target device, including: receiving the scheduling parameter sent by the master device; wherein the scheduling parameter is based on matching with the target device The operator scheduling template and the scheduling parameter search algorithm are generated; the scheduling process corresponding to the operator is run according to the scheduling parameters; the performance data of the scheduling process is fed back to the main control device for the main control device The performance data adjusts the scheduling parameter and sends it to the target device.

In order to at least achieve the above purpose, the embodiment of the present application also proposes a main control device, including: a search module, configured to search for an operator scheduling template that matches the target device; a scheduling parameter generation module, configured to The sub-scheduling template and the scheduling parameter search algorithm generate scheduling parameters, and send the scheduling parameters to the target device for the target device to run the scheduling process corresponding to the operator according to the scheduling parameters; the iteration module is used to receive the The performance data for executing the scheduling process fed back by the target device is adjusted, and the scheduling parameter is adjusted according to the performance data and sent to the target device.

In order to at least achieve the above purpose, an embodiment of the present application also proposes a target device, including: a receiving module, configured to receive a scheduling parameter sent by a master device; wherein, the scheduling parameter is based on an operator matched with the target device A scheduling template and scheduling parameter search algorithm are generated; an operation module is used to run the scheduling process corresponding to the operator according to the scheduling parameters; a feedback module is used to feed back the performance data of the scheduling process to the main control device for the The master control device adjusts the scheduling parameter according to the performance data and sends it to the target device.

In order to at least achieve the above purpose, an embodiment of the present application also proposes an electronic device, including: at least one processor; and a memory connected to the at least one processor in communication; wherein, the memory stores information that can be used by the Instructions executed by at least one processor, where the instructions are executed by the at least one processor, so that the at least one processor can execute the scheduling parameter adjustment method described in any one of the preceding items.

In order to at least achieve the above purpose, the embodiment of the present application also proposes a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the method for adjusting the scheduling parameter as described in any one of the preceding items is implemented.

Description of drawings

One or more embodiments are exemplified by pictures in the accompanying drawings, and these exemplifications are not intended to limit the embodiments.

FIG. 1 is a schematic flowchart of a method for adjusting scheduling parameters applied to a master control device provided in an embodiment of the present application;

FIG. 2 is a schematic flowchart of a method for adjusting scheduling parameters applied to a target device provided in an embodiment of the present application;

FIG. 3 is a schematic structural diagram of a master control device provided in another embodiment of the present application;

FIG. 4 is a schematic structural diagram of a target device provided in another embodiment of the present application;

Fig. 5 is a schematic structural diagram of an electronic device provided in another embodiment of the present application.

Detailed ways

In the scheduling parameter adjustment method provided by the embodiment of the present application, after searching for the operator scheduling template matching the target device, generating scheduling parameters according to the matching operator scheduling template and scheduling parameter search algorithm, and sending the scheduling parameters to the target device, Provide the target device to run the scheduling process corresponding to the operator according to the scheduling parameters, and then receive the performance data of the execution scheduling process fed back by the target device, adjust the scheduling parameters according to the performance data and send it to the target device until the performance data converges, that is, the scheduling design process of reasoning Decoupling into three parts: determining operator scheduling template, determining scheduling parameters, and running that can be understood and executed by the machine, so that the scheduling design process that was originally implemented manually can be handed over to the machine to complete, overcoming the dependence on manual labor, and reducing the scheduling design process While reducing the workload of manual participation, it can also improve design efficiency, cover as many actual application scenarios as possible, enhance applicability and practicality, and enable scheduling for any remote device in various application scenarios Design, efficiently and quickly obtain the optimal scheduling parameters, so as to automatically accelerate the inference speed of any deep learning network model. In addition, the three processes of determining the operator scheduling template, determining the scheduling parameters, and running are handed over to the main control device and the target device to complete respectively. Among them, determining the scheduling parameters is to generate scheduling parameters according to the matching operator scheduling template and scheduling parameter search algorithm , put it on the main control device to complete, without running on the target device, and avoid the problem that the target device is a device with poor computing performance, such as user terminal, edge device CPU, etc., which is inefficient or even impossible to realize.

In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that in each embodiment of the application, many technical details are provided for readers to better understand the application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can also be realized. The division of the following embodiments is for the convenience of description, and should not constitute any limitation to the specific implementation of the present application, and the embodiments can be combined and referred to each other on the premise of no contradiction.

On the one hand, the embodiments of the present application provide a method for adjusting scheduling parameters, which is applied to a master control device. The master control device may be an electronic device such as a computer or a server, as shown in FIG. 1 , and specifically includes the following steps.

Step 101, searching for an operator scheduling template matching the target device.

In this embodiment, the operator refers to various operations in the depth model, such as convolution, pooling, splicing, and upsampling. An operator scheduling template is a description of the scheduling process of an operator in a specific environment, including at least operator feature information and operator running information. For example, for convolution operations, an operator scheduling template can include convolution kernels, running The environment refers to information such as the hardware that the runtime depends on.

In one example, when the target device is provided with an operator scheduling template database and multiple operator scheduling templates are stored in the operator scheduling template database, at this time, searching for an operator scheduling template that matches the target device can be done by It is realized in the following manner: according to the hardware information of the target device and the corresponding relationship between the preset operator scheduling template and the hardware information, query the preset operator scheduling template database to obtain the operator scheduling template that matches the target device. For example, for an operator whose convolution kernel is 3×3 and whose runtime hardware is an ARM processor, you can use the query condition "type==conv and kernel_size>1and env==arm" in the operator scheduling template The database queries the corresponding operator scheduling template.

It should be noted that the operator scheduling template in the operator scheduling template database needs to cover a variety of application scenarios as much as possible, that is, operators with various operator feature information and operator running information, for example, For convolution operators, the operator scheduling template database must at least define the corresponding subdivision scenarios for each subdivided scene in which the length of the convolution kernel is equal to 1 or greater than 1, and the operating environment is x86 or x84 CPU, ARM processor, GPU, etc. Operator scheduling template.

It should also be noted that, if the corresponding operator scheduling template cannot be found in the operator scheduling template database, if a completely matching operator scheduling template cannot be found, the operator scheduling template with the highest matching degree can be used as the matched operator scheduling template. For example, for a convolution operator with a CPU with a convolution kernel of 7×7 and an operating environment of x84, if only the convolution kernel is 1×1 and the convolution kernel is 3×3 in the operator scheduling template database, the operating environment Operator scheduling templates for x86 CPUs, ARM processors, and GPUs. At this time, the matching operator scheduling templates are operator scheduling templates corresponding to CPUs with 3×3 convolution kernels and x86 operating environments.

It can be understood that a deep learning model usually includes several operators, and the scheduling design should actually design the scheduling process of multiple operators. Therefore, in an example, before searching for the operator scheduling template that matches the target device The method for adjusting the scheduling parameters further includes: splitting the deep learning model involved in the scheduling parameters to be acquired into a single operator. Correspondingly, searching for an operator scheduling template that matches the target device includes: searching for an operator scheduling template that matches the target device and obtains operators obtained through splitting.

In one example, the deep learning model that needs inference acceleration in the target device is a face recognition model trained based on Convolutional Neural Network (CNN), that is, the deep learning model involved in the scheduling parameters to be obtained is face recognition model, first split the face recognition model, and obtain 32 convolution operators with convolution kernels of 11×11, 1 pooling operator, 32 convolution operators of 9×9, and 16 convolution operators of 7 ×7 convolution operator, 16 5×5 convolution operators, 1 fully connected operator, and 1 loss function operator. Among them, the target device uses the CPU when running the deep learning model, and then, according to The operator feature information and operator running information of each operator above are searched and matched against the corresponding operator scheduling template in the operator scheduling template database.

It should be noted that this embodiment does not limit the number of operators and operator scheduling templates. How many operators or how many types of operators are involved in the deep learning model involved in the scheduling parameters to be obtained in the target device? Search for the corresponding number of operator scheduling templates. If the deep learning model consists of 78 operations, search for the 78 operator scheduling templates corresponding to the 78 operations. Or, if the deep learning model consists of 98 operations, and this The 98 operations correspond to 75 types of operators, and the operator scheduling templates corresponding to the 75 types of operators included in the 98 operations are searched. Among them, operators with different operator feature information can be considered as different types of operators, such as Convolution operators with different convolution kernels can be considered as different kinds of operators. Of course, the above is only a specific example, and the number of operator scheduling templates and the depth model can also have other relationships, which will not be described here.

Step 102, generate scheduling parameters according to the matching operator scheduling template and scheduling parameter search algorithm, and send the scheduling parameters to the target device, so that the target device runs the scheduling process corresponding to the operator according to the scheduling parameters.

In this embodiment, the scheduling parameter search algorithm is an algorithm for finding an optimal solution in an optimization problem, such as simulated annealing algorithm, gradient descent algorithm, global traversal algorithm, etc. This embodiment does not limit the scheduling parameter search algorithm.

In this embodiment, generating scheduling parameters based on the operator scheduling template and the scheduling parameter search algorithm can be realized in the following way: Generate an operator scheduling parameter set according to the scheduling parameters exposed by the matching operator scheduling template, where the operator scheduling template exposes The scheduling parameters refer to the preset feasible scheduling parameters in the operator scheduling process; the operator scheduling parameter set includes multiple sets of scheduling parameters, and each set of scheduling parameters includes the scheduling parameters required in the operator's one scheduling process; The scheduling parameter search algorithm searches out a set of scheduling parameters in the operator scheduling parameter set, and uses the searched set of scheduling parameters as the generated scheduling parameters.

In an example, the scheduling parameters of an operator include parameter A and parameter B. During the scheduling process, the feasible value range of parameter A is {a1, a2,...,an}, and the feasible value range of parameter B is {b1, b2,...,bm}, then the parameter A includes a1, a2,..., an, and the parameter B includes b1, b2,..., bm in the scheduling parameters exposed by the operator scheduling template corresponding to the operator , operator scheduling parameter set C, that is, the scheduling parameters are {(a1, b1), (a1, b2), ..., (a1, bm), (a2, b1), ..., (a2, bm), ... ..., (an, bm)}, and then search for the optimal solution in the set C based on the scheduling parameter search algorithm, that is, the optimal combination of parameter A and parameter B, where the optimal solution can be parameter A and parameter B with the shortest execution time The combination of B may be a combination of parameter A and parameter B that require the least system resources.

Of course, the above description is for the case where the exposed scheduling parameters can determine specific values, that is, they can be exhausted. In this embodiment, the exposed scheduling parameters can also include continuous scheduling parameters within a certain range, that is, they cannot be exhausted. At this time, the operator scheduling parameter set is still generated according to the exposed scheduling parameters, and then the operator scheduling parameter set is searched based on the scheduling parameter search algorithm, which will not be described here.

It should be noted that there may actually be multiple scheduling parameter search algorithms in the target device, so that an appropriate scheduling parameter search algorithm can be selected according to actual conditions.

Therefore, in an example, after the operator scheduling parameter set is generated, before a scheduling parameter search algorithm is used to search for a set of scheduling parameters in the operator scheduling parameter set, the scheduling parameter adjustment method further includes: according to the operator-based scheduling The size of the parameter search space formed by the parameter set, select a scheduling parameter search algorithm in the preset scheduling search algorithm database; wherein, the parameter search space is obtained based on the operator scheduling parameter set, and the scheduling search algorithm database includes a variety of scheduling parameter search algorithm. Correspondingly, searching for a set of scheduling parameters in the operator scheduling parameter set by using the scheduling parameter search algorithm includes: searching for a set of scheduling parameters in the operator scheduling parameter set by using the selected scheduling parameter search algorithm.

In particular, according to the size of the parameter search space formed based on the operator scheduling parameter set, a scheduling parameter search algorithm is selected from the preset scheduling search algorithm database, which can be realized in the following way: predict the performance according to the size of the parameter search space The time required for data convergence; when the time required for performance data convergence is greater than the preset threshold, select a scheduling parameter search algorithm that is biased towards global uniform search; when the time required for performance data convergence is less than or equal to the preset threshold, Select a scheduling parameter search algorithm that searches for a local optimal solution within a specified time.

In an example, the scheduling parameters of an operator can be exhausted, that is, the feasible values of the scheduling parameters can be explained by enumerating. At this time, the search space is considered to be relatively small, and it can be selected from the preset scheduling search algorithm database. The global traversal algorithm gets the generated scheduling parameters. It is worth mentioning that, since the global traversal algorithm will compare each feasible solution, and then determine the optimal solution, therefore, the global traversal algorithm can ensure that the obtained scheduling parameters are the current optimal solution, and the accuracy of the search extremely high.

In another example, the search space of scheduling parameters of an operator is relatively large, and the execution time of the scheduling parameter search algorithm is required in advance. At this time, it is necessary to estimate the convergence time of the scheduling parameter search algorithm. In the case where the estimated convergence time is less than the execution time, the scheduling parameter search algorithm with high search accuracy can be preferentially selected in the scheduling search algorithm database, such as global traversal. If the estimated convergence time is not less than the execution time, A scheduling parameter search algorithm with high search efficiency can be preferentially selected in the scheduling search algorithm database, such as the steepest descent method. For example, when the parameter search space size corresponding to an operator scheduling template is 100, and the evaluation time of each target device is 5s, the total running time of the algorithm is about 500s. If the time is less than the preset threshold T, then You can choose the global traversal algorithm, and vice versa, you can choose optimization algorithms such as simulated annealing.

Of course, the above is only a specific example, and in actual implementation, a suitable scheduling parameter search algorithm can be selected from the scheduling search algorithm database according to requirements, and details will not be repeated here.

It should be noted that the deep learning model in the target device usually contains several operators. Therefore, there may be multiple operator scheduling templates obtained through the matching in step 101. Considering the actual relationship between operators, the target When the deep learning model is running on the device, the scheduling process of operators will affect each other. Therefore, the influence between several operator scheduling templates needs to be considered when generating scheduling parameters. That is to say, the scheduling parameter search algorithm is for all matching In terms of operator scheduling templates, rather than a single operator scheduling template. In particular, different numbers of scheduling templates of the same type of operators corresponding to deep learning models will also lead to differences in optimal scheduling parameters.

It can be understood that the scheduling template for all matched operators is mainly related to the objective function in the scheduling parameter search algorithm. Therefore, the above only uses a single operator scheduling template as an example for illustration, and can be extended to the case of multiple operator scheduling templates, which does not mean that this embodiment can only be implemented for a single operator scheduling template. For example, when determining the search space, the search space may be formed by combining the scheduling parameters of the various operators included in the deep learning model, which will not be described here.

Step 103, receiving the performance data of executing the scheduling process fed back by the target device, adjusting the scheduling parameters according to the performance data and sending it to the target device.

Specifically, when receiving the performance data fed back by the target device, first compare the historically received performance data according to the currently received performance data to detect whether the performance has improved. If it is detected that the performance has not been improved, It is determined that the search algorithm has converged. At this time, it is necessary to select the scheduling parameter corresponding to the test item with the best historical scheduling performance as the choice of the optimal scheduling parameter; when it is detected that the performance has been improved, it is determined that the search algorithm has not converged. , there may still be a better combination of scheduling parameters in the search space, and it is necessary to adjust the scheduling parameter search algorithm according to a certain strategy to select another set of scheduling parameter combinations and send them to the target device for execution. Among them, adjusting the scheduling parameter search algorithm The strategy can be an optimization direction determined according to the execution effect, or it can add a certain disturbance to the scheduling parameter search algorithm to make it iteratively select other scheduling parameter combinations in another direction, which will not be described here.

It should be noted that, for the target device, it actually continuously receives the performance data returned by the target device, then adjusts the scheduling parameters according to the performance parameters, and then sends the adjusted scheduling parameters to the target device until the performance data converges, that is, obtains Satisfactory scheduling parameters, that is, the optimal scheduling parameters are determined through loops until the optimal scheduling parameters are obtained, which ensures the optimality of the scheduling parameters.

It can be seen from the background technology that implementing deep learning model reasoning by manually designing the scheduling process is often inefficient and cannot be optimally scheduled. It is worth mentioning that this embodiment realizes automatic scheduling design by decoupling the scheduling design process into three parts, which can be understood and executed by the machine: determining operator scheduling template, determining scheduling parameters, and running. Dependence, and because the scheduling design is realized automatically, it avoids the limitation of manpower, and can design the optimal scheduling for all operators on any hardware, even though the deep learning model usually contains a large number of different types of operators, and the types For the same operator, different parameters have different optimal scheduling implementation methods. Different types of hardware used will also affect the optimal scheduling of hardware resources. Even the same operator is optimal on the same type of hardware with different models. There are also differences in scheduling, and a large amount of design work can be completed by the huge computing power of the machine, and the coverage of scheduling design for various application scenarios can be realized.

On the other hand, the embodiments of the present application also provide a method for adjusting scheduling parameters, which is applied to a target device. The target device may be an electronic device such as a computer or a server, as shown in FIG. 2 , which specifically includes the following steps.

Step 201, receiving the scheduling parameters sent by the master device; wherein, the scheduling parameters are generated according to the operator scheduling template and the scheduling parameter search algorithm matching the target device.

It should be noted that since there may be multiple operator scheduling templates matched by the target device, the received scheduling parameters may be the scheduling parameters of a single operator or the scheduling parameters of multiple operators.

Step 202, run the scheduling process corresponding to the operator according to the scheduling parameters.

Specifically, during the running process, the target device also monitors the running process to obtain performance data.

Step 203: Feedback the performance data of executing the scheduling process to the main control device, so that the main control device can adjust the scheduling parameters according to the performance data and send it to the target device.

In addition, it should be understood that the division of steps in the above methods is only for clarity of description, and may be combined into one step or split into multiple steps during implementation. As long as the same logical relationship is included, all Within the scope of protection of this patent; adding insignificant modifications or introducing insignificant designs to the algorithm or process, but not changing the core design of the algorithm and process are all within the scope of protection of the patent.

On the other hand, the embodiments of the present application also provide a master control device, as shown in FIG. 3 , including the following modules.

The search module 301 is configured to search for an operator scheduling template that matches the target device.

The scheduling parameter generating module 302 is configured to generate scheduling parameters according to the matching operator scheduling template and scheduling parameter search algorithm, and send the scheduling parameters to the target device for the target device to run the scheduling process corresponding to the operator according to the scheduling parameters.

The iteration module 303 is configured to receive the performance data of the execution scheduling process fed back by the target device, adjust the scheduling parameters according to the performance data and send it to the target device.

It is not difficult to find that this embodiment is a device embodiment corresponding to the method embodiment applied to the main control device, and this embodiment can be implemented in cooperation with the method embodiment applied to the main control device. The relevant technical details mentioned in the embodiment of the method applied to the master control device are still valid in this embodiment, and will not be repeated here to reduce repetition. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the method embodiment applied to the master control device.

It is worth mentioning that all the modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, or a part of a physical unit, or multiple physical units. Combination of units. In addition, in order to highlight the innovative part of the present application, units that are not closely related to solving the technical problem proposed in the present application are not introduced in this embodiment, but this does not mean that there are no other units in this embodiment.

On the other hand, the embodiment of the present application provides a target device, as shown in FIG. 4 , including the following modules.

The receiving module 401 is configured to receive the scheduling parameters sent by the main control device; wherein, the scheduling parameters are generated according to the operator scheduling template and the scheduling parameter search algorithm matched with the target device.

The running module 402 is configured to run the scheduling process corresponding to the operator according to the scheduling parameters.

The feedback module 403 is configured to feed back the performance data of the scheduling process to the main control device, so that the main control device adjusts the scheduling parameters according to the performance data and sends them to the target device.

It is not difficult to find that this embodiment is a device embodiment corresponding to the method embodiment applied to the target device, and this embodiment can be implemented in cooperation with the method embodiment applied to the target device. The relevant technical details mentioned in the embodiment of the method applied to the target device are still valid in this embodiment, and will not be repeated here in order to reduce repetition. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied in the method embodiment applied to the target device.

On the other hand, the embodiment of the present application also provides an electronic device, as shown in FIG. 5 , including: at least one processor 501; and a memory 502 communicatively connected to the at least one processor 501; An instruction to be executed by at least one processor 501, the instruction is executed by at least one processor 501, so that at least one processor 501 can execute the scheduling parameter adjustment method described in any one of the above method embodiments.

Wherein, the memory 502 and the processor 501 are connected by a bus, and the bus may include any number of interconnected buses and bridges, and the bus connects one or more processors 501 and various circuits of the memory 502 together. The bus may also connect together various other circuits such as peripherals, voltage regulators, and power management circuits, all of which are well known in the art and therefore will not be further described herein. The bus interface provides an interface between the bus and the transceivers. A transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing means for communicating with various other devices over a transmission medium. The data processed by the processor 501 is transmitted on the wireless medium through the antenna, and further, the antenna also receives the data and transmits the data to the processor 501 .

Processor 501 is responsible for managing the bus and general processing, and may also provide various functions including timing, peripheral interface, voltage regulation, power management and other control functions. And the memory 502 may be used to store data used by the processor 501 when performing operations.

Another aspect of the embodiment of the present application provides a computer-readable storage medium storing a computer program. When the computer program is executed by the processor, the scheduling parameter adjustment method described in any one of the above method embodiments is implemented.

That is, those skilled in the art can understand that all or part of the steps in the method of the above-mentioned embodiments can be completed by instructing related hardware through a program, the program is stored in a storage medium, and includes several instructions to make a device ( It may be a single-chip microcomputer, a chip, etc.) or a processor (processor) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. .

In addition, an embodiment of the present invention also provides a computer program product, the computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program includes program instructions, and when the program instructions are executed When executed by a computer, the computer is made to execute the method in any of the above method embodiments.

Those of ordinary skill in the art can understand that the above-mentioned embodiments are specific embodiments for realizing the present application, and in practical applications, various changes can be made to it in form and details without departing from the spirit and spirit of the present application. scope.

Claims

A method for adjusting scheduling parameters, wherein, applied to a master control device, comprising:

Search for an operator scheduling template that matches the target device;

generating scheduling parameters according to the matched operator scheduling template and scheduling parameter search algorithm, and sending the scheduling parameters to the target device, so that the target device runs a scheduling process corresponding to the operator according to the scheduling parameters;

The performance data for executing the scheduling process fed back by the target device is received, and the scheduling parameter is adjusted according to the performance data and sent to the target device.
The method for adjusting scheduling parameters according to claim 1, wherein the searching for an operator scheduling template that matches the target device includes:

According to the hardware information of the target device and the corresponding relationship between the preset operator scheduling template and the hardware information, query the preset operator scheduling template database to obtain the operator scheduling template that matches the target device;

Wherein, the operator scheduling template database stores a plurality of operator scheduling templates.
The method for adjusting scheduling parameters according to claim 1, wherein said generating scheduling parameters according to said operator scheduling template and scheduling parameter search algorithm comprises:

According to the scheduling parameters exposed by the matched operator scheduling template, an operator scheduling parameter set is generated; the operator scheduling parameter set includes multiple sets of scheduling parameters, and each set of scheduling parameters includes the desired scheduling parameters;

A group of scheduling parameters is searched in the set of operator scheduling parameters by using the scheduling parameter search algorithm, and the searched group of scheduling parameters is used as the generated scheduling parameters.
The method for adjusting scheduling parameters according to claim 3, wherein after said generating the operator scheduling parameter set, a group of scheduling is searched in the operator scheduling parameter set using the scheduling parameter search algorithm Before the parameters, also include:

According to the size of the parameter search space formed based on the operator scheduling parameter set, select a scheduling parameter search algorithm from the preset scheduling search algorithm database; wherein, the parameter search space is obtained based on the operator scheduling parameter set , the scheduling search algorithm database includes multiple scheduling parameter search algorithms;

The searching out a set of scheduling parameters in the operator scheduling parameter set by using the scheduling parameter search algorithm includes:

A set of scheduling parameters is searched in the set of operator scheduling parameters by using the selected scheduling parameter search algorithm.
The method for adjusting scheduling parameters according to claim 4, wherein, according to the size of the parameter search space formed based on the operator scheduling parameter set, a scheduling parameter search algorithm is selected from a preset scheduling search algorithm database ,include:

Estimate the time required for the convergence of the performance data according to the size of the parameter search space;

When the time required for the convergence of the performance data is greater than a preset threshold, select a scheduling parameter search algorithm that is biased towards global uniform search;

In the case that the time required for the convergence of the performance data is less than or equal to a preset threshold, a scheduling parameter search algorithm that searches for a local optimal solution within a specified time is selected.
The method for adjusting scheduling parameters according to any one of claims 1 to 5, wherein, before the searching for an operator scheduling template that matches the target device, further comprising:

Split the deep learning model involved in the scheduling parameters to be obtained into a single operator;

The searching for an operator scheduling template that matches the target device includes:

Search for the operator scheduling template of the split operator that matches the target device.
A method for adjusting scheduling parameters, wherein, applied to a target device, comprising:

receiving scheduling parameters sent by the master control device; wherein, the scheduling parameters are generated according to an operator scheduling template and a scheduling parameter search algorithm that match the target device;

Run the scheduling process corresponding to the operator according to the scheduling parameters;

The performance data of executing the scheduling process is fed back to the master control device, so that the master control device adjusts the scheduling parameter according to the performance data and sends it to the target device.
A master control device, including:

A search module, configured to search for operator scheduling templates that match the target device;

A scheduling parameter generating module, configured to generate scheduling parameters according to the matched operator scheduling template and scheduling parameter search algorithm, and send the scheduling parameters to the target device for the target device to run the operator according to the scheduling parameters Sub-corresponding scheduling process;

An iteration module, configured to receive performance data for executing the scheduling process fed back by the target device, adjust the scheduling parameters according to the performance data, and send the scheduling parameters to the target device.
A target device, comprising:

A receiving module, configured to receive scheduling parameters sent by the master control device; wherein, the scheduling parameters are generated according to an operator scheduling template and a scheduling parameter search algorithm that match the target device;

A running module, configured to run a scheduling process corresponding to the operator according to the scheduling parameters;

A feedback module, configured to feed back performance data of executing the scheduling process to the main control device, so that the main control device adjusts the scheduling parameters according to the performance data and sends them to the target device.
An electronic device, comprising:

at least one processor; and,

a memory communicatively coupled to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, the instructions are executed by the at least one processor, so that the at least one processor can perform the operation described in any one of claims 1 to 6 The method for adjusting the scheduling parameters described above, or perform the method for adjusting the scheduling parameters according to claim 7.
A computer-readable storage medium storing a computer program, wherein, when the computer program is executed by a processor, the method for adjusting a scheduling parameter according to any one of claims 1 to 6 is realized, or the method according to claim 7 is realized The adjustment method of the scheduling parameters.