CN114064242A - Method, device and storage medium for adjusting scheduling parameters - Google Patents
Method, device and storage medium for adjusting scheduling parameters Download PDFInfo
- Publication number
- CN114064242A CN114064242A CN202111354335.XA CN202111354335A CN114064242A CN 114064242 A CN114064242 A CN 114064242A CN 202111354335 A CN202111354335 A CN 202111354335A CN 114064242 A CN114064242 A CN 114064242A
- Authority
- CN
- China
- Prior art keywords
- scheduling
- operator
- parameter
- parameters
- template
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/48—Program initiating; Program switching, e.g. by interrupt
- G06F9/4806—Task transfer initiation or dispatching
- G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The embodiment of the application relates to the technical field of computers, and provides a method, equipment and a storage medium for adjusting scheduling parameters. The method for adjusting the scheduling parameters is applied to the main control equipment and comprises the following steps: searching an operator scheduling template matched with the target equipment; 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 equipment for the target equipment to run a scheduling process corresponding to an operator according to the scheduling parameter; and receiving performance data for executing the scheduling process, which is fed back by the target equipment, adjusting the scheduling parameters according to the performance data and sending the scheduling parameters to the target equipment. Under various application scenes, the method can overcome the dependence on manpower, automatically carry out scheduling design on the operator scheduling process of any target equipment to obtain the optimal scheduling parameter, and is more efficient and faster.
Description
Technical Field
Embodiments of the present disclosure relate to the field of computer technologies, and in particular, to a method, a device, and a storage medium for adjusting scheduling parameters.
Background
With the great success of deep learning technology in many fields such as computer vision, speech recognition, natural language processing, etc. in recent years, the industry has also begun to deploy services related to deep learning model inference step by step on various types of hardware, such as Central Processing Units (CPUs), Graphics Processing Units (GPUs), smart chips, etc., where performance indexes such as inference delay, throughput, etc. of a deep learning model can be better promoted only when computing resources and storage resources of the hardware are reasonably and fully scheduled.
However, at present, scheduling optimization of deep learning model inference is mainly completed manually, and optimal scheduling cannot be generally achieved, so that efficiency is low, and reasonable resource calling cannot be performed quickly and efficiently.
Disclosure of Invention
The embodiment of the application mainly aims to provide a method, equipment and a storage medium for adjusting scheduling parameters, and aims to overcome the dependence on manual work in various application scenarios, automatically perform scheduling design on an operator scheduling process of any target equipment, obtain optimal scheduling parameters, and achieve high efficiency and high speed.
In order to at least achieve the above object, an embodiment of the present application provides a method for adjusting a scheduling parameter, which is applied to a master control device, and includes: searching an operator scheduling template matched with the target equipment; 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 equipment for the target equipment to run a scheduling process corresponding to an operator according to the scheduling parameter; and receiving performance data for executing the scheduling process, which is fed back by the target equipment, adjusting the scheduling parameters according to the performance data and sending the scheduling parameters to the target equipment.
In order to at least achieve the above object, an embodiment of the present application further provides a method for adjusting a scheduling parameter, which is applied to a target device, and includes: receiving scheduling parameters sent by the master control equipment; the scheduling parameters are generated according to an operator scheduling template matched with the target equipment and a scheduling parameter searching algorithm; operating a scheduling process corresponding to the operator according to the scheduling parameter; and feeding back performance data for executing the scheduling process to the main control equipment, so that the main control equipment adjusts the scheduling parameters according to the performance data and sends the scheduling parameters to the target equipment.
In order to at least achieve the above object, an embodiment of the present application further provides a master control device, including: the search module is used for searching an operator scheduling template matched with the target equipment; the scheduling parameter generating module is used for generating scheduling parameters according to the matched operator scheduling template and the scheduling parameter searching algorithm, and sending the scheduling parameters to the target equipment so that the target equipment can operate a scheduling process corresponding to an operator according to the scheduling parameters; and the iteration module is used for receiving the performance data which is fed back by the target equipment and used for executing the scheduling process, adjusting the scheduling parameters according to the performance data and sending the scheduling parameters to the target equipment.
In order to achieve at least the above object, an embodiment of the present application further provides a target device, including: the receiving module is used for receiving the scheduling parameters sent by the main control equipment; the scheduling parameters are generated according to an operator scheduling template matched with the target equipment and a scheduling parameter searching algorithm; the operation module is used for operating the scheduling process corresponding to the operator according to the scheduling parameter; and the feedback module is used for feeding back performance data for executing the scheduling process to the main control equipment, so that the main control equipment adjusts the scheduling parameters according to the performance data and sends the scheduling parameters to the target equipment.
In order to achieve at least the above object, an embodiment of the present application further provides an electronic device, including: 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 to enable the at least one processor to perform the method of adjusting scheduling parameters as described in any one of the above.
To achieve at least the above object, an embodiment of the present application further provides a computer-readable storage medium, which stores a computer program, and the computer program, when executed by a processor, implements the method for adjusting the scheduling parameter as described in any one of the above.
The method for adjusting scheduling parameters provided in the embodiment of the application searches an operator scheduling template matched with target equipment, generates scheduling parameters according to the matched operator scheduling template and a scheduling parameter search algorithm, sends the scheduling parameters to the target equipment, allows the target equipment to run a scheduling process corresponding to an operator according to the scheduling parameters, receives performance data fed back by the target equipment for executing the scheduling process, adjusts the scheduling parameters according to the performance data and sends the scheduling parameters to the target equipment until the performance data is converged, i.e. decouples the scheduling design process into a determined operator scheduling template which can be understood and executed by a machine, determines the scheduling parameters and runs the three parts, thereby being capable of handing the original manually-implemented scheduling design process to the machine for completion, overcoming the dependence on manual work, reducing the workload of manual work participation in the scheduling design process, the design efficiency can be improved, more practical application scenes can be covered as much as possible, the applicability and the practicability are enhanced, the scheduling design can be carried out on any remote equipment in various application scenes, the optimal scheduling parameters can be obtained efficiently and quickly, and the reasoning speed of any deep learning network model can be automatically accelerated. In addition, the three processes of determining the operator scheduling template, determining the scheduling parameter and running are respectively completed by the main control device and the target device, wherein the determined scheduling parameter is generated according to the matched operator scheduling template and the scheduling parameter search algorithm and is completed by the main control device without running on the target device, so that the problem that the target device is poor in calculation performance, such as a user terminal and a CPU (Central processing Unit) of an edge device, and the efficiency is low or even cannot be realized is solved.
Drawings
One or more embodiments are illustrated by the corresponding figures in the drawings, which are not meant to be limiting.
Fig. 1 is a schematic flowchart of a method for adjusting scheduling parameters applied to a master device according to an embodiment of the present application;
fig. 2 is a flowchart illustrating a method for adjusting scheduling parameters applied to a target device according to an embodiment of the present application;
fig. 3 is a schematic structural diagram of a master device provided in another embodiment of the present application;
FIG. 4 is a schematic 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 Description
The method for adjusting scheduling parameters provided in the embodiment of the application searches an operator scheduling template matched with target equipment, generates scheduling parameters according to the matched operator scheduling template and a scheduling parameter search algorithm, sends the scheduling parameters to the target equipment, allows the target equipment to run a scheduling process corresponding to an operator according to the scheduling parameters, receives performance data fed back by the target equipment for executing the scheduling process, adjusts the scheduling parameters according to the performance data and sends the scheduling parameters to the target equipment until the performance data is converged, i.e. decouples the scheduling design process into a determined operator scheduling template which can be understood and executed by a machine, determines the scheduling parameters and runs the three parts, thereby being capable of handing the original manually-implemented scheduling design process to the machine for completion, overcoming the dependence on manual work, reducing the workload of manual work participation in the scheduling design process, the design efficiency can be improved, more practical application scenes can be covered as much as possible, the applicability and the practicability are enhanced, the scheduling design can be carried out on any remote equipment in various application scenes, the optimal scheduling parameters can be obtained efficiently and quickly, and the purpose of automatically accelerating the inference speed of any deep learning network model is achieved. In addition, the three processes of determining the operator scheduling template, determining the scheduling parameter and running are respectively completed by the main control device and the target device, wherein the determined scheduling parameter is generated according to the matched operator scheduling template and the scheduling parameter search algorithm and is completed by the main control device without running on the target device, so that the problem that the target device is poor in calculation performance, such as a user terminal and a CPU (Central processing Unit) of an edge device, and the efficiency is low or even cannot be realized is solved.
To make the objects, 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, it will be appreciated by those of ordinary skill in the art that in the examples of the present application, numerous technical details are set forth in order to provide a better understanding of the present application. However, the technical solution claimed in the present application can be implemented without these technical details and various changes and modifications based on the following embodiments. The following embodiments are divided for convenience of description, and should not constitute any limitation to the specific implementation manner of the present application, and the embodiments may be mutually incorporated and referred to without contradiction.
An embodiment of the present application provides a method for adjusting a scheduling parameter, which is applied to a master control device, where the master control device may be an electronic device such as a computer and a server, and as shown in fig. 1, the method specifically includes:
In this embodiment, the operator refers to various operations in the depth model, such as convolution, pooling, splicing, upsampling, and the like. The operator scheduling template is a description of a scheduling process of an operator in a certain specific environment, and at least includes operator characteristic information and operator operation information, for example, for convolution operation, the operator scheduling template may include information such as a convolution kernel, an operation environment, that is, hardware that is depended on during operation.
In an example, when the target device is provided with an operator scheduling template database and the operator scheduling template database stores a plurality of operator scheduling templates, searching for an operator scheduling template matching the target device may be implemented as follows: and inquiring in a preset operator scheduling template database to obtain an operator scheduling template matched with the target equipment according to the hardware information of the target equipment and the corresponding relation between the preset operator scheduling template and the hardware information. For example, for an operator with a convolution kernel of 3 × 3 and ARM processor as the runtime-dependent hardware, the operator scheduling template database may be queried for the corresponding operator scheduling template through a query condition "type and kernel _ size >1and env ═ ARM".
It should be noted that the operator scheduling templates in the operator scheduling template database need to cover various application scenarios as much as possible, that is, covers operators having various operator characteristic information and operator operation information, for example, for a convolution operator, the operator scheduling template database needs to define at least corresponding operator scheduling templates of each subdivided scenario in environments such as CPU, ARM processor, GPU, etc. with convolution kernel length equal to 1and greater than 1and operation environment of x86 or x 84.
It should be further noted that, a corresponding operator scheduling template is queried in the operator scheduling template database, and if a completely matched operator scheduling template cannot be found, the operator scheduling template with the highest matching degree may be used as the matched operator scheduling template, for example, for a convolution operator of a CPU with a convolution kernel of 7 × 7 and an operating environment of x84, if the operator scheduling template in the operator scheduling template database only has a convolution kernel of 1 × 1and a convolution kernel of 3 × 3, and the operating environments are x86, the CPU, the ARM processor, and the GPU, the matched operator scheduling template at this time is the operator scheduling template corresponding to the CPU with a convolution kernel of 3 × 3 and an operating environment of x 86.
It can be understood that, the deep learning model generally includes several operators, and the scheduling design should actually be a scheduling process designed for multiple operators, so in one example, before searching for an operator scheduling template matching the target device, the method for adjusting the scheduling parameter further includes: and splitting the deep learning model related to the scheduling parameter to be acquired into single operators. Accordingly, searching for an operator scheduling template matching the target device includes: and searching an operator scheduling template of the operator obtained by splitting and matched with the target equipment.
In an example, a deep learning model needing inference acceleration in a target device is a face recognition model obtained by training based on a Convolutional Neural Network (CNN), that is, a deep learning model related to scheduling parameters to be obtained is a face recognition model, the face recognition model is firstly split, 32 convolution operators with convolution kernels of 11 × 11, 1 pooling operator, 32 convolution operators of 9 × 9, 16 convolution operators of 7 × 7, 16 convolution operators of 5 × 5, 1 full-link operator, and 1 loss function operator are sequentially obtained, wherein a CPU is used when the target device runs the deep learning model, and then corresponding operator scheduling templates are searched and matched in an operator scheduling template database according to operator characteristic information and operator running information of each operator.
It should be noted that, in this embodiment, the number of operators and operator scheduling templates is not limited, and the deep learning model related to the scheduling parameter to be acquired in the target device is composed of how many operators or how many operators, and the corresponding number of operator scheduling templates needs to be searched, where if the deep learning model is composed of 78 operations, the 78 operator scheduling templates corresponding to the 78 operations are searched, or if the deep learning model is composed of 98 operations, and the 98 operations correspond to 75 operators, the operator scheduling templates corresponding to the 75 operators included in the 98 operations are searched, where operators with different operator characteristic information may be considered as different types of operators, and, for example, convolution operators with different convolution kernels may be considered as different types of operators. Of course, the above is only a specific example, and the number of the operator scheduling templates and the depth model may also be in other relationships, which is not described in detail here.
And 102, 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 equipment for the target equipment to run a scheduling process corresponding to the operator according to the scheduling parameter.
In this embodiment, the scheduling parameter search algorithm is an algorithm for finding an optimal solution in an optimization problem, such as a simulated annealing algorithm, a gradient descent algorithm, a global traversal algorithm, and the like, and the scheduling parameter search algorithm is not limited in this embodiment.
In this embodiment, generating the scheduling parameter according to the operator scheduling template and the scheduling parameter search algorithm may be implemented as follows: generating an operator scheduling parameter set according to the scheduling parameters exposed by the matched operator scheduling template, wherein the scheduling parameters exposed by the operator scheduling template refer to feasible scheduling parameters in a preset operator scheduling process; the operator scheduling parameter set comprises a plurality of groups of scheduling parameters, and each group of scheduling parameters comprises scheduling parameters required in the primary scheduling process of the operator; and searching a group of scheduling parameters in the operator scheduling parameter set by using a scheduling parameter search algorithm, and taking the searched group of scheduling parameters as the generated scheduling parameters.
In one example, the scheduling parameters of a certain operator include a parameter a and a parameter B, the feasible value range of the parameter a in the scheduling process is { a1, a2, … …, an }, and the feasible value range of the parameter B is { B1, B2, … …, bm }, then the parameters a of the scheduling parameters exposed by the operator scheduling template corresponding to the operator include a1, a2, … …, an, the parameters B include B1, B2, … …, bm, the operator scheduling parameter set C, i.e. scheduling parameters are { (a1, b1), (a1, b2), … …, (a1, bm), (a2, b1), … …, (a2, bm), … …, (an, bm) }, then an optimal solution is found in the set C based on a scheduling parameter search algorithm, that is, the optimal combination of the parameter a and the parameter B, wherein the optimal solution may be the combination of the parameter a and the parameter B with the shortest execution time, the combination of the parameter a and the parameter B with the least system resource requirement, and the like.
Certainly, the above description is given by way of example for the case that the exposed scheduling parameters can determine specific numerical values, that is, the exposed scheduling parameters may be exhausted, in this embodiment, the exposed scheduling parameters may also include continuous scheduling parameters within a certain range, that is, the exposed scheduling parameters are not exhausted, at this time, an operator scheduling parameter set is still generated according to the exposed scheduling parameters, and then, the operator scheduling parameter set is searched based on a scheduling parameter search algorithm, which is not described here any more.
It should be noted that, the scheduling parameter search algorithm in the target device may actually be various, so as to select an appropriate scheduling parameter search algorithm according to the actual situation.
Therefore, in one example, after generating the operator scheduling parameter set, before searching out a set of scheduling parameters in the operator scheduling parameter set by using the scheduling parameter search algorithm, the method for adjusting the scheduling parameters further includes: selecting a scheduling parameter search algorithm from a preset scheduling search algorithm database according to the size of a parameter search space formed based on an operator scheduling parameter set; the parameter search space is obtained based on an operator scheduling parameter set, and the scheduling search algorithm database comprises a plurality of scheduling parameter search algorithms. Correspondingly, a group of scheduling parameters is searched in the operator scheduling parameter set by using a scheduling parameter search algorithm, and the method comprises the following steps: and searching a group of scheduling parameters in the operator scheduling parameter set by using the selected scheduling parameter searching algorithm.
In particular, according to the size of a parameter search space formed based on an operator scheduling parameter set, one scheduling parameter search algorithm is selected from a preset scheduling search algorithm database, and the method can be implemented as follows: estimating the time required by performance data convergence according to the size of the parameter search space; under the condition that the time required by performance data convergence is greater than a preset threshold value, selecting a scheduling parameter search algorithm biased to global uniform search; and under the condition that the time required by the performance data convergence is less than or equal to a preset threshold value, selecting a scheduling parameter searching algorithm for searching a local optimal solution within a specified time.
In an example, the scheduling parameters of a certain operator may be exhausted, that is, the feasible values of the scheduling parameters may be described in an enumerated manner, at this time, the search space is considered to be relatively small, and the generated scheduling parameters may be obtained by using a global traversal algorithm selected from a preset scheduling search algorithm database. It should be noted that, since the global traversal algorithm compares every feasible solution and then determines the optimal solution, the obtained scheduling parameter can be ensured to be the current optimal solution by the global traversal algorithm, and the searching accuracy is extremely high.
In another example, the search space of the scheduling parameter of an operator is relatively large, and the execution time of the scheduling parameter search algorithm is required in advance, at this time, the convergence time of the scheduling parameter search algorithm needs to be estimated, and under the condition that the estimated convergence time is less than the execution time, a scheduling parameter search algorithm with high search accuracy, such as global traversal and the like, can be preferentially selected in the scheduling search algorithm database, and under the condition that the estimated convergence time is not less than the execution time, a scheduling parameter search algorithm with high search efficiency, such as the steepest descent method and the like, can be preferentially selected in the scheduling search algorithm database. For example, when the size of the parameter search space corresponding to a certain 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 a preset threshold T, the global traversal algorithm may be selected, and otherwise, optimization algorithms such as simulated annealing may be selected.
Of course, the above is only a specific example, and in practice, a suitable scheduling parameter search algorithm may be selected from the scheduling search algorithm database according to requirements during implementation, and is not described here any more.
It should be noted that the deep learning model in the target device usually includes a plurality of operators, and therefore, there may be a plurality of operator scheduling templates obtained through matching in step 101, and considering that there is an association relationship between operators, when the deep learning model is run in the target device, scheduling processes of the operators will have an influence on each other, and therefore, when generating the scheduling parameters, the influence between the operator scheduling templates needs to be considered, that is, the scheduling parameter search algorithm is specific to all matched operator scheduling templates, rather than a single operator scheduling template. In particular, different numbers of operator scheduling templates of the same class corresponding to the deep learning model may also bring different optimal scheduling parameters.
It will be appreciated that the scheduling templates for all matched operators are mainly related to the objective function in the scheduling parameter search algorithm, etc. Therefore, the above description is only given by taking a single operator scheduling template as an example, and can be generalized to the case of multiple operator scheduling templates, which does not mean that the present embodiment can only be implemented for a single operator scheduling template. For example, when determining the search space, the search space formed by combining the scheduling parameters of the operators included in the deep learning model may be used, which is not described herein any more.
And 103, receiving the performance data of the execution scheduling process fed back by the target equipment, adjusting the scheduling parameters according to the performance data and sending the scheduling parameters to the target equipment.
Specifically, when performance data fed back by target equipment is received, comparing the historically received performance data according to the currently received performance data to detect whether the performance is improved, and judging that a search algorithm is converged under the condition that the performance is not improved, wherein at the moment, a scheduling parameter corresponding to a test item with the optimal performance of historical scheduling operation needs to be selected as the selection of the optimal scheduling parameter; under the condition that the performance is detected to be improved, judging that the search algorithm is not converged, at this time, there still may be a better scheduling parameter combination in the search space, and adjusting the scheduling parameter search algorithm according to a certain strategy to select another group of scheduling parameter combination and send the scheduling parameter combination to the target device for execution, wherein the strategy for adjusting the scheduling parameter search algorithm may be an optimization direction determined according to an execution effect, or may be adding a certain disturbance to the scheduling parameter search algorithm to make it continue to iterate to select another scheduling parameter combination in another direction, which is not described herein any more.
It should be noted that for the target identification, it is practical to continuously receive the performance data returned by the target device, then adjust the scheduling parameter according to the performance parameter, and then send the adjusted scheduling parameter to the target device until the performance data converges, that is, a satisfactory scheduling parameter is obtained, that is, the optimal scheduling parameter is determined by cycling until the optimal scheduling parameter is obtained, so that the optimality of the scheduling parameter is ensured.
As known from the background art, the deep learning model reasoning realized by manually designing the scheduling process is often inefficient and cannot obtain optimal scheduling. It is worth mentioning that, in the embodiment, the scheduling design process is decoupled into three parts, namely, the determined operator scheduling template, the determined scheduling parameter and the operation, which can be understood and executed by the machine, so that the automated scheduling design is realized, the dependence on manpower is overcome, further, because the scheduling design is realized automatically, the limitation of manpower is avoided, the optimal scheduling can be designed for all operators on any hardware, even if the deep learning model usually contains a large number of operators of different types and operators of the same type, the realization modes of the different optimal scheduling parameters of the operators have differences, the different types of the used hardware can influence the optimal scheduling of hardware resources, even if the optimal scheduling of the same operator on the same type of hardware with different types has differences, a large amount of design work can be completed by the huge computing power of the machine, coverage of scheduling designs for a variety of application scenarios is achieved.
Another aspect of the embodiments of the present application further provides a method for adjusting a scheduling parameter, which is applied to a target device, where the target device may be an electronic device such as a computer and a server, and as shown in fig. 2, the method specifically includes:
It should be noted that, because there may be a plurality of operator scheduling templates matched with the target device, the received scheduling parameter may be a scheduling parameter of a single operator, or may be a scheduling parameter of multiple operators.
And step 202, operating a scheduling process corresponding to the operator according to the scheduling parameter.
Specifically, during operation, the target device also monitors the operation process to obtain performance data.
And 203, feeding back the performance data for executing the scheduling process to the master control device, so that the master control device adjusts the scheduling parameters according to the performance data and sends the scheduling parameters to the target device.
In addition, it should be understood that the above steps of the various methods are divided for clarity, and the implementation may be combined into one step or split into some steps, and the steps are divided into multiple steps, so long as the same logical relationship is included in the protection scope of the present patent; it is within the scope of the patent to add insignificant modifications to the algorithms or processes or to introduce insignificant design changes to the core design without changing the algorithms or processes.
Another aspect of the embodiments of the present application further provides a master device, as shown in fig. 3, including:
and the searching module 301 is used for searching the operator scheduling template matched with the target device.
And the scheduling parameter generating module 302 is configured to generate a scheduling parameter according to the matched operator scheduling template and the scheduling parameter search algorithm, and send the scheduling parameter to the target device, so that the target device runs a scheduling process corresponding to the operator according to the scheduling parameter.
And the iteration module 303 is configured to receive the performance data of the scheduling execution process fed back by the target device, adjust the scheduling parameter according to the performance data, and send the scheduling parameter to the target device.
It is obvious that this embodiment is a device embodiment corresponding to the method embodiment applied to the master control device, and this embodiment may be implemented in cooperation with the method embodiment applied to the master control device. Related technical details mentioned in the method embodiment applied to the main control device are still valid in this embodiment, and are not described herein again in order to reduce repetition. Accordingly, the related technical details mentioned in the present embodiment can also be applied to the method embodiment applied to the master control device.
It should be noted that, all the modules involved in this embodiment are logic modules, and in practical application, one logic unit may be one physical unit, may also be a part of one physical unit, and may also be implemented by a combination of multiple physical units. In addition, in order to highlight the innovative part of the present application, a unit that is not so closely related to solving the technical problem proposed by the present application is not introduced in the present embodiment, but this does not indicate that there is no other unit in the present embodiment.
Another aspect of the embodiments of the present application further provides a target device, as shown in fig. 4, including:
a receiving module 401, configured to receive a scheduling parameter sent by a master control device; and the scheduling parameters are generated according to an operator scheduling template matched with the target equipment and a scheduling parameter searching algorithm.
And an operation module 402, configured to operate a scheduling process corresponding to the operator according to the scheduling parameter.
The feedback module 403 is configured to feed back performance data for executing a scheduling process to the master control device, so that the master control device adjusts a scheduling parameter according to the performance data and sends the scheduling parameter to the target device.
It is to be understood that the present embodiment is an embodiment of an apparatus corresponding to an embodiment of a method applied to a target apparatus, and the present embodiment may be implemented in cooperation with an embodiment of a method applied to a target apparatus. Related technical details mentioned in the method embodiment applied to the target device are still valid in this embodiment, and are not described herein again in order to reduce repetition. Accordingly, the related technical details mentioned in the present embodiment can also be applied in the method embodiment applied to the target device.
It should be noted that, all the modules involved in this embodiment are logic modules, and in practical application, one logic unit may be one physical unit, may also be a part of one physical unit, and may also be implemented by a combination of multiple physical units. In addition, in order to highlight the innovative part of the present application, a unit that is not so closely related to solving the technical problem proposed by the present application is not introduced in the present embodiment, but this does not indicate that there is no other unit in the present embodiment.
Another aspect of the embodiments of the present application further provides an electronic device, as shown in fig. 5, including: at least one processor 501; and a memory 502 communicatively coupled to the at least one processor 501; the memory 502 stores instructions executable by the at least one processor 501, and the instructions are executed by the at least one processor 501, so that the at least one processor 501 can perform the method for adjusting the scheduling parameter according to any one of the method embodiments.
The memory 502 and the processor 501 are coupled by a bus, which may include any number of interconnected buses and bridges that couple one or more of the various circuits of the processor 501 and the memory 502 together. The bus may also connect various other circuits such as peripherals, voltage regulators, power management circuits, and the like, which are well known in the art, and therefore, will not be described any further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver may be one element or a plurality of elements, such as a plurality of receivers and transmitters, providing a means for communicating with various other apparatus over a transmission medium. The data processed by the processor 501 is transmitted over a wireless medium via an antenna, which further receives the data and transmits the data to the processor 501.
The processor 501 is responsible for managing the bus and general processing and may also provide various functions including timing, peripheral interfaces, voltage regulation, power management, and other control functions. And memory 502 may be used to store data used by processor 501 in performing operations.
In another aspect, an embodiment of the present application further provides a computer-readable storage medium storing a computer program. The computer program, when executed by a processor, implements the method for adjusting scheduling parameters described in any of the above method embodiments.
That is, as can be understood by those skilled in the art, all or part of the steps in the method for implementing the embodiments described above may be implemented by a program instructing related hardware, where the program is stored in a storage medium and includes several instructions to enable a device (which may be a single chip, a chip, or the like) or a processor (processor) to execute all or part of the steps of the method described in the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.
It will be understood by those of ordinary skill in the art that the foregoing embodiments are specific examples for carrying out the present application, and that various changes in form and details may be made therein without departing from the spirit and scope of the present application in practice.
Claims (11)
1. A method for adjusting scheduling parameters is applied to a master control device, and comprises the following steps:
searching an operator scheduling template matched with the target equipment;
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 equipment for the target equipment to run a scheduling process corresponding to an operator according to the scheduling parameter;
and receiving performance data for executing the scheduling process, which is fed back by the target equipment, adjusting the scheduling parameters according to the performance data and sending the scheduling parameters to the target equipment.
2. The method according to claim 1, wherein the searching for the operator scheduling template matching the target device comprises:
inquiring in a preset operator scheduling template database to obtain an operator scheduling template matched with the target equipment according to the hardware information of the target equipment and the corresponding relation between the preset operator scheduling template and the hardware information;
and a plurality of operator scheduling templates are stored in the operator scheduling template database.
3. The method according to claim 1, wherein the generating the scheduling parameter according to the operator scheduling template and the scheduling parameter search algorithm comprises:
generating an operator scheduling parameter set according to the scheduling parameters exposed by the matched operator scheduling template; the operator scheduling parameter set comprises a plurality of groups of scheduling parameters, and each group of scheduling parameters comprises scheduling parameters required in one scheduling process of the operator;
and searching a group of scheduling parameters in the operator scheduling parameter set by using the scheduling parameter search algorithm, and taking the searched group of scheduling parameters as the generated scheduling parameters.
4. The method according to claim 3, wherein after the generating the operator scheduling parameter set, before the searching out a set of scheduling parameters in the operator scheduling parameter set by the scheduling parameter searching algorithm, further comprises:
selecting a scheduling parameter search algorithm from a preset scheduling search algorithm database according to the size of a parameter search space formed based on the operator scheduling parameter set; the parameter search space is obtained based on the operator scheduling parameter set, and the scheduling search algorithm database comprises a plurality of scheduling parameter search algorithms;
the searching out a group of scheduling parameters in the operator scheduling parameter set by the scheduling parameter searching algorithm comprises:
and searching a group of scheduling parameters in the operator scheduling parameter set by using the selected scheduling parameter searching algorithm.
5. The method according to claim 4, wherein selecting a scheduling parameter search algorithm from a preset scheduling search algorithm database according to the size of a parameter search space formed based on the operator scheduling parameter set comprises:
estimating the time required by the performance data convergence according to the size of the parameter search space;
under the condition that the time required by the performance data convergence is greater than a preset threshold value, selecting a scheduling parameter search algorithm biased to global uniform search;
and under the condition that the time required by the performance data convergence is less than or equal to a preset threshold value, selecting a scheduling parameter searching algorithm for searching a local optimal solution within a specified time.
6. The method according to any one of claims 1 to 5, wherein before the searching for the operator scheduling template matching the target device, the method further comprises:
splitting a deep learning model related to a scheduling parameter to be acquired into single operators;
the searching operator scheduling template matched with the target device comprises the following steps:
and searching an operator scheduling template of the operator obtained by splitting and matched with the target equipment.
7. A method for adjusting scheduling parameters, applied to a target device, includes:
receiving scheduling parameters sent by the master control equipment; the scheduling parameters are generated according to an operator scheduling template matched with the target equipment and a scheduling parameter searching algorithm;
operating a scheduling process corresponding to the operator according to the scheduling parameter;
and feeding back performance data for executing the scheduling process to the main control equipment, so that the main control equipment adjusts the scheduling parameters according to the performance data and sends the scheduling parameters to the target equipment.
8. A master device, comprising:
the search module is used for searching an operator scheduling template matched with the target equipment;
the scheduling parameter generating module is used for generating scheduling parameters according to the matched operator scheduling template and the scheduling parameter searching algorithm, and sending the scheduling parameters to the target equipment so that the target equipment can operate a scheduling process corresponding to an operator according to the scheduling parameters;
and the iteration module is used for receiving the performance data which is fed back by the target equipment and used for executing the scheduling process, adjusting the scheduling parameters according to the performance data and sending the scheduling parameters to the target equipment.
9. A target device, comprising:
the receiving module is used for receiving the scheduling parameters sent by the main control equipment; the scheduling parameters are generated according to an operator scheduling template matched with the target equipment and a scheduling parameter searching algorithm;
the operation module is used for operating the scheduling process corresponding to the operator according to the scheduling parameter;
and the feedback module is used for feeding back performance data for executing the scheduling process to the main control equipment, so that the main control equipment adjusts the scheduling parameters according to the performance data and sends the scheduling parameters to the target equipment.
10. An electronic device, comprising:
at least one processor; and the number of the first and second groups,
a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of adjusting a scheduling parameter of any one of claims 1 to 6 or to perform the method of adjusting a scheduling parameter of claim 7.
11. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the method for adjusting the scheduling parameter of any one of claims 1 to 6, or implements the method for adjusting the scheduling parameter of claim 7.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111354335.XA CN114064242A (en) | 2021-11-12 | 2021-11-12 | Method, device and storage medium for adjusting scheduling parameters |
| PCT/CN2022/129029 WO2023083058A1 (en) | 2021-11-12 | 2022-11-01 | Scheduling parameter adjusting method, devices, and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111354335.XA CN114064242A (en) | 2021-11-12 | 2021-11-12 | Method, device and storage medium for adjusting scheduling parameters |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114064242A true CN114064242A (en) | 2022-02-18 |
Family
ID=80272481
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111354335.XA Pending CN114064242A (en) | 2021-11-12 | 2021-11-12 | Method, device and storage medium for adjusting scheduling parameters |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN114064242A (en) |
| WO (1) | WO2023083058A1 (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023083058A1 (en) * | 2021-11-12 | 2023-05-19 | 中兴通讯股份有限公司 | Scheduling parameter adjusting method, devices, and storage medium |
| CN116304720A (en) * | 2023-05-18 | 2023-06-23 | 之江实验室 | Cost model training method and device, storage medium and electronic equipment |
| CN116755782A (en) * | 2023-08-18 | 2023-09-15 | 腾讯科技（深圳）有限公司 | Method, device, equipment, storage medium and program product for instruction scheduling |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117170879B (en) * | 2023-11-01 | 2024-03-12 | 之江实验室 | Device management device and method for intelligent chip |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180341851A1 (en) * | 2017-05-24 | 2018-11-29 | International Business Machines Corporation | Tuning of a machine learning system |
| CN111796917A (en) * | 2019-04-09 | 2020-10-20 | 华为技术有限公司 | Operator operation scheduling method and device |
| EP4012630A4 (en) * | 2019-09-17 | 2022-11-23 | Huawei Cloud Computing Technologies Co., Ltd. | Model optimization method and apparatus, storage medium, and device |
| CN111752716A (en) * | 2020-06-29 | 2020-10-09 | 北京小米松果电子有限公司 | Model using method, data processing method and device |
| CN114064242A (en) * | 2021-11-12 | 2022-02-18 | 中兴通讯股份有限公司 | Method, device and storage medium for adjusting scheduling parameters |
-
2021
- 2021-11-12 CN CN202111354335.XA patent/CN114064242A/en active Pending
-
2022
- 2022-11-01 WO PCT/CN2022/129029 patent/WO2023083058A1/en unknown
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023083058A1 (en) * | 2021-11-12 | 2023-05-19 | 中兴通讯股份有限公司 | Scheduling parameter adjusting method, devices, and storage medium |
| CN116304720A (en) * | 2023-05-18 | 2023-06-23 | 之江实验室 | Cost model training method and device, storage medium and electronic equipment |
| CN116304720B (en) * | 2023-05-18 | 2023-08-25 | 之江实验室 | Cost model training method and device, storage medium and electronic equipment |
| CN116755782A (en) * | 2023-08-18 | 2023-09-15 | 腾讯科技（深圳）有限公司 | Method, device, equipment, storage medium and program product for instruction scheduling |
| CN116755782B (en) * | 2023-08-18 | 2023-10-20 | 腾讯科技（深圳）有限公司 | Method, device, equipment, storage medium and program product for instruction scheduling |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2023083058A1 (en) | 2023-05-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114064242A (en) | Method, device and storage medium for adjusting scheduling parameters | |
| US20220197953A1 (en) | Model pushing method and device, model requesting method and device, storage medium and electronic device | |
| CN109325041B (en) | Service data processing method, device, computer equipment and storage medium | |
| EP3734475A1 (en) | Method and device for training data, storage medium, and electronic device | |
| US20160253220A1 (en) | Data center operation | |
| US10956417B2 (en) | Dynamic operation scheduling for distributed data processing | |
| CN111338770A (en) | Task scheduling method, server and computer readable storage medium | |
| CN113934763B (en) | SQL query method and device for distributed database | |
| CN108776833B (en) | Data processing method, system and computer readable storage medium | |
| EP3299968A1 (en) | Big data calculation method and system | |
| JP3152279B2 (en) | Sub-optimal assignment determination method | |
| CN111831355A (en) | Weight precision configuration method, device, equipment and storage medium | |
| US10768974B2 (en) | Specifying an order of a plurality of resources in a transaction according to distance | |
| CN111831359A (en) | Weight precision configuration method, device, equipment and storage medium | |
| EP3217300B1 (en) | Extraction program, extraction device and extraction method | |
| CN114091688B (en) | Computing resource obtaining method and device, electronic equipment and storage medium | |
| US20240095529A1 (en) | Neural Network Optimization Method and Apparatus | |
| CN114611675A (en) | Data processing method, data processing device, electronic device and storage medium | |
| EP4142238A1 (en) | Pico base station configuration method and apparatus, and storage medium and electronic apparatus | |
| CN111831356B (en) | Weight precision configuration method, device, equipment and storage medium | |
| WO2023155783A1 (en) | Model adjustment method and apparatus, service processing method and apparatus, and device and storage medium | |
| CN115348184B (en) | Internet of things data security event prediction method and system | |
| CN109242680B (en) | Method for dynamically adjusting block chain consensus | |
| CN115544033B (en) | Method, device, equipment and medium for updating check repeat vector library and checking repeat data | |
| CN114185938B (en) | Project traceability analysis method and system based on digital finance and big data traceability |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |