CN110262995A - Executive body creation system and executive body creation method - Google Patents

Abstract

The present disclosure discloses an execution body creation system, including: a task topology map generation component, which is used to decompose the job to be completed into a series of tasks executed by the execution body in the heterogeneous architecture, and when the job is decomposed, Based on the inherent relationship between the decomposed tasks, a task relationship topology diagram is generated, the task nodes in the task topology diagram contain all the node attributes required to execute the corresponding task, and the task topology diagram generation component includes a specified component node selection unit , used to select a specific task node having a specified component including an additional node attribute according to a specific task in the process of generating the task relationship topology diagram, and the additional node attribute changes the way the specific task node executes the corresponding task; and an executive creation component , based on the task relationship topology graph and all node attributes of each task node or all node attributes containing specified components, create a corresponding execution body for each task node in the computing resource, and make the specific task node with additional node attributes The corresponding specific executive executes the corresponding task according to the additional node attributes.

Description

Executive body creation system and executive body creation method

technical field

The present disclosure relates to a system and method for fine-grained management and hardware optimization of storage space in a data processing network, and more specifically, to an executive creation system and execution creation method in a data processing network.

Background technique

With the development of machine learning and the gradual deepening of artificial neural network research, the concept of deep learning has been widely concerned and applied. Deep learning is a special kind of machine learning. It uses a network hierarchical structure to express the learning objects, combines simple concepts into abstract concepts, and realizes abstract concept expression through simple concept calculations. At present, deep learning has made great progress in the fields of image recognition, speech recognition and natural language processing. Deep learning involves a large number of model parameters, resulting in a huge amount of calculation, and the scale of training data is large, so it needs to consume more computing resources.

With the rise of big data computing and deep learning, various coprocessors are usually used to share the data processing function of the CPU. Such as GPU (Graphic Processing Unit), APU, etc. GPU has a highly parallel structure, so GPU is more efficient than CPU in processing graphics data and complex algorithms. When the CPU performs computing tasks, only one piece of data is processed at a time, and there is no true parallelism, while the GPU has multiple processor cores, which can process multiple pieces of data in parallel at a time. Compared with the CPU, the GPU has more ALUs (ArithmeticLogic Unit, logical operation execution body) for data processing, rather than data cache and flow control. Such a structure is very suitable for large-scale data that is highly unified, independent of each other, and a pure computing environment that does not need to be interrupted.

Existing big data computing and deep learning network systems perform continuous data processing by pre-determining the operation functions of each computing executive. Therefore, once the network is started and enters the running period, the executive needs to operate and process the data block according to the predetermined operation. , and the operation of the execution body and the operation of the data block are fixed. This will lead to waste of resources in the process of processing data blocks at some execution bodies, and even cause the operation process to be too slow to vacate data blocks due to upstream and downstream relationships and affect the operation of downstream execution bodies.

Therefore, people hope to have a data processing network that can adjust the relationship between a specific executable and a data block, so as to improve the data processing speed of the overall data processing network and improve the reuse rate of the executable.

Contents of the invention

Due to the highly uniform type of data processed by big data and deep learning, it is possible to provide a method that can eliminate the above-mentioned problems existing in the prior art. The purpose of this disclosure is to provide an executive body creation system, including: a task topology map generation component, which is used to decompose the job to be completed into a series of tasks executed by the executive body in the heterogeneous architecture, and in the job decomposition At the same time, based on the inherent relationship between the decomposed tasks, a task relationship topology diagram is generated, the task nodes in the task topology diagram contain all the node attributes required to perform the corresponding tasks, and the task topology diagram generation component includes specified component nodes A selection unit, configured to select a specific task node with a specified component including an additional node attribute according to a specific task in the process of generating the task relationship topology diagram, and the additional node attribute changes the way the specific task node executes the corresponding task; and the execution body Create components, based on the task relationship topology graph and all node attributes of each task node or all node attributes containing specified components, create a corresponding execution body for each task node in the computing resource, and make it related to specific nodes with additional node attributes The specific execution body corresponding to the task node executes the corresponding task according to the additional node attributes.

According to the execution body creation system of the present disclosure, the additional node attribute contained in the designated component is a data operation node attribute, which is used to make the created specific execution body only perform operations on the header data of the processed data.

According to the execution body creation system of the present disclosure, the additional node attribute contained in the specified component is a data storage node attribute, which is used to make the created specific execution body only store the header data of the processed data.

According to the execution body creation system of the present disclosure, the additional node attribute contained in the designated component is a data modification node attribute, which is used to enable the created specific execution body to modify the header data of the processed data.

According to another aspect of the present disclosure, a method for creating an executive body is provided, including: decomposing the job to be completed into a series of tasks performed by the executive body in the heterogeneous architecture, and while decomposing the job, based on the Decompose the inherent relationship between tasks to generate a task relationship topology diagram, the task nodes of the task topology diagram contain all the node attributes required to perform the corresponding tasks; in the process of generating the task relationship topology diagram, according to the specific task selection has additional The specific task node of the specified component of the node attribute, the additional node attribute changes the way the specific task node executes the corresponding task; and based on the task relationship topology map and all node attributes of each task node or all nodes containing the specified component The attribute creates a corresponding execution body for each task node in the computing resource, and makes the specific execution body corresponding to the specific task node with the additional node attribute execute the corresponding task according to the additional node attribute.

According to the execution body creation method of the present disclosure, the additional node attribute contained in the designated component is a data operation node attribute, which is used to make the created specific execution body only perform operations on the header data of the processed data.

According to the execution body creation method of the present disclosure, the additional node attribute contained in the designated component is a data storage node attribute, which is used to make the created specific execution body only store the header data of the processed data.

According to the execution body creation method of the present disclosure, the additional node attribute contained in the designated component is a data modification node attribute, which is used to enable the created specific execution body to modify the header data of the processed data.

By specifying a component node selection unit to select a specific task node with a specified component that includes additional node attributes, it is possible to make the execution body created based on all node attributes that include node additional attributes do the operation range, operation mode, and storage mode of the data block. Corresponding to the change of node attributes, the executive body has more flexible operation functions, so as to realize the fine management of storage space in the static data processing network, and optimize the execution body hardware of each node. By specifying the component node selection unit to select a specific task node with the node attribute included in the specified component, the corresponding specific execution body can modify the header data of the data block, so that the execution body can exceed the inherent properties of the data block itself More diverse operations are performed on the data blocks, so that the diversified processing requirements of the data processing network for the data blocks can be expanded.

Part of the other advantages, objectives and features of the present disclosure will be realized through the following description, and part of them will be understood by those skilled in the art through the study and practice of the present disclosure.

Description of drawings

FIG. 1 is a schematic structural diagram of a system for creating an executive in a data processing network according to the present disclosure;

FIG. 2 is a schematic diagram showing the influence of a specified component node selection unit on a specific execution body in the execution body creation system in the data processing network according to the present disclosure.

FIG. 3 is a schematic structural diagram of an executive network created by the executive creation system in the data processing network according to the present disclosure.

Detailed ways

The present disclosure will be described in further detail below in conjunction with the embodiments and drawings, so that those skilled in the art can implement it with reference to the description.

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the following exemplary examples do not represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatuses and methods consistent with aspects of the present disclosure as recited in the appended claims.

The terminology used in the present disclosure is for the purpose of describing particular embodiments only and is not intended to limit the present disclosure. As used in this disclosure and the appended claims, the singular forms "a", "the", and "the" are intended to include the plural forms as well, unless the context clearly dictates otherwise. It should also be understood that the term "and/or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

It should be understood that although the terms first, second, third, etc. may be used in the present disclosure to describe various information, the information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of the present disclosure, hereinafter, one of the two possible devices may be referred to as a first executive or as a second executive, and similarly, the other of the two possible devices may be What is called the second executive can also be called the first executive. Depending on the context, the word "if" as used herein may be interpreted as "at" or "when" or "in response to a determination."

In order to enable those skilled in the art to better understand the present disclosure, the present disclosure will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

FIG. 1 is a schematic structural diagram of an executive creation system used in a data processing network according to the present disclosure. As shown in Figure 1, the executive body creation system is arranged in a heterogeneous architecture, which is composed of CPU00 and CPU01 respectively connected to CPU00, GPU00, GPU01, and GPU2, and GPU10, GPU11, and GPU12 connected to CPU01. Although only two CPUs and six GPUs are shown here, the heterogeneous architecture may include more CPUs, and more or fewer GPUs connected to each CPU may be determined based on actual needs.

The executable creation system 100 according to the present disclosure is deployed in the heterogeneous architecture shown in FIG. 1 . Although the components of the executive body creation system shown in FIG. 1 are shown separately from each CPU and GPU, this is for highlighting and convenient description of the processing. In fact, the components of the executable creation system are all distributed in the CPU and/or GPU.

As shown in FIG. 1 , the execution body creation system 100 includes a task topology graph generation component 120 and an execution body creation component 130 . The executive web component 140 is a creation result web component of the executive creation component 130 .

As shown in FIG. 1 , the task topology graph generation component 120 is used to decompose the required job into a series of tasks executed by the executives in the heterogeneous architecture, and when the job is decomposed, based on the decomposed tasks Intrinsic relationship between tasks to generate a topological map of task relationships. The executable creation system 100 is set up to process predetermined types of job data. In order to continuously process the same type of data, jobs need to be decomposed into simple tasks suitable for CPU or GPU computing units to perform calculations or other operations. Specifically, a job is broken down into tasks that are related to each other. The job decomposition by the task topology graph generation component 120 includes the decomposition of the data block and the data processing model, that is, the decomposition of the job is to set the decomposition mode of the job data to be processed. Specifically, according to the description of the job task requirements, the job is decomposed into a multi-layer neural network structure according to the process to be processed. A job (Job) is decomposed into a series of interdependent tasks (Task), this dependency is usually described by a directed acyclic graph (Directedacyclic graph, DAG), each node represents a task, and the connecting lines between nodes Represents a data dependency (producer and consumer relationship). The situation of the task relationship graph after job decomposition is not specifically described here.

While decomposing the job step by step, the task topology graph generation component 120 also forms the task relationship topology layer by layer. Due to the inherent logical relationship among the various tasks formed by the job decomposition, as the job is decomposed into different tasks and on different task layers, the task topology map generation component 120 also forms a task relationship topology map. These task-relationship topologies form neural networks among the decomposed tasks. In the case of complex tasks, the task relationship topology graph contains multiple layers, so a multi-layer task neural network is also formed. Each layer of neural network contains not only the neuron nodes corresponding to specific tasks, but also the relationship between each neuron, and it contains both the data parallel network for future tasks that will be used to process sliced data, as well as the data parallel network that will be used for A model-parallel network for the task of sharding models. Optionally, these neural networks may only include data parallel networks. Whether to include the data parallel network and the model parallel network at the same time can be configured according to actual needs.

In order to create an executable body for any node in the task topology map at one time in the subsequent executable body creation component, the task topology map generation component 120 according to the present disclosure, while generating each node in the task topology map, assigns each node the necessary information to execute the corresponding task. All required node attributes. The attributes of all nodes include, for example, resource attributes specifying the resources required by the task corresponding to the node, condition attributes that trigger the trigger conditions for task execution, and so on. The above-mentioned all node attributes also include the additional node attributes mentioned later. It is precisely because each node in the task topology diagram of this disclosure contains all node attributes, so it automatically has all the resources and all attributes of the task when it is created subsequently, and is in a fully configured state. When specific data executes specific tasks, such as dynamic allocation of environmental resources and dynamic configuration of trigger conditions and so on. For each executive created based on the task topology graph of the present disclosure and nodes containing all node attributes, it is in a static state when processing specific data, and the only change is the difference in input data. The nodes of the neural network of the existing executive body creation system for deep learning contain very few or no node attributes, so the execution of the corresponding task of the node needs to temporarily derive the required attributes during the execution of the specific task to dynamically obtain the corresponding attributes. attribute to complete the corresponding task. And this kind of temporary derivation attribute requires temporary derivation every time for the same task, so it will require a lot of computing overhead.

It should be pointed out that, while the task topology diagram generation component 120 forms the task relationship topology diagram layer by layer, it needs to optimize the already formed task relationship topology diagram. Therefore, the task topology map generation component 120 according to the present disclosure also includes a topology map optimization component 121 . The topology graph optimization component 121 includes various optimization units, such as redundant node elimination unit 1211, blocked node elimination unit 1212, and equivalent subgraph transformation unit 1213 such as a designated component node selection unit, and other components for optimizing the topology graph Unit 1214. Although it is shown that the above three units are included in FIG. 1 of the present disclosure, it does not mean that the present disclosure must include all of these units. The implementation of the present disclosure does not necessarily include the above-mentioned topology optimization component 121 . The existence of the topology optimization component 121 will make the task topology generated by the task topology generation component 120 more reasonable, and will run more smoothly in the subsequent data processing process with higher processing efficiency.

Specifically, during the process of generating the task topology graph by the task topology graph generating component 120, there may be cases where corresponding nodes are repeatedly generated for a certain task. For example, in a neural network subgraph, there may be two juxtaposed nodes, both of which have the same upstream node and the same downstream node, and correspond to the same task. Such a node is a redundant node situation. The existence of such redundant nodes will repeatedly consume computing resources in heterogeneous architectures, complicating the neural network. Therefore, this redundant node is a node that needs to be removed. If such a duplicate node is found in the process of generating the task topology map by the task topology map generation component 120, the redundant node elimination unit 1211 will know the existence of this node and directly delete the redundant node, so that the redundant node The upstream and downstream nodes are only associated to the upstream and downstream nodes of the same node as the redundant node to be deleted (a node performing the same task as the redundant node). In addition, in the process of generating the task topology map by the task topology map generation component 120, there may be interactions between certain tasks, and downstream node congestion may occur due to untimely task processing, which will result in blocked nodes. Blockages are propagated forward. For this reason, if such blocked nodes are found during the process of generating the task topology map by the task topology map generation component 120, the blocked node eliminating unit 1212 will eliminate the nodes that cause operation blockage in the task topology map. Specifically, it is to change the connection edge between the blocked node and the upstream node, add one or more nodes, and eliminate the conduction of the blocked node to the upstream.

In the process of generating the task topology graph by the task topology graph generation component 120, in order to allow the execution body to have more flexible operation functions, the refined management of the storage space in the static data processing network can be realized, and the execution body hardware of each node can be obtained. Optimization, the specified component node selection unit 1213 selects a specific task node with a specified component containing an additional node attribute according to the specific task in the process of generating the task relationship topology diagram, and the additional node attribute changes the way the specific task node executes the corresponding task For example, change the scope of task data processing by the executive corresponding to the task node, change the scope of output data storage, eliminate an unnecessary backward operation, change the modifiable attributes of output data, and change the frequency of execution of tasks by the executive, etc.

Although only three topological graph optimization units such as redundant node elimination unit 1211, blocked node elimination unit 1212, and specified component node selection unit 1213 are listed in detail above, there are many topology graph optimization units used in the present disclosure, Not described one by one here. In addition, during the process of generating the task topology graph by the task topology graph generating component 120, there may be situations where the generated network subgraphs for some associated tasks are relatively complex or inefficient. In order to obtain a more efficient task topology map, the task topology map generation component 120 will generate multiple network subgraphs for some associated tasks, thus, it is necessary to perform equivalent transformations on various subgraphs in the topology map optimization component 121 , so as to select the subgraph network with the highest computing efficiency from multiple subgraph networks capable of completing the same operation function to replace the current subgraph network. Although various optimization units of the above topology graph optimization component 121 are disclosed, any other optimization units may also be included, such as the other unit 1214 shown in FIG. 1 .

After the task topology graph generation component 120 generates the task neural network topology graph of each layer, the executive body creation component 130 creates a corresponding executive body for each task in the computing resources included in the heterogeneous architecture based on the task relationship topology graph. Specifically, according to the hardware resources required in the task description, and based on all node attributes of each node, a corresponding number of computing units and corresponding storage units are designated for each task in the heterogeneous architecture to form an execution body to execute the corresponding task. The created executable includes various resources in the computing resources in the heterogeneous architecture, such as storage unit, message sending or receiving unit, computing unit and so on. The computing unit of the executive body may be one or may contain multiple, as long as it can complete the specified task. After the execution body is created, it will always execute the specified task and will not change, unless the task to be executed disappears, for example, the heterogeneous architecture to which the execution body belongs is applied to the processing of other types of jobs again. The created network relationship between the executives corresponds to the relationship between each neural network node in the task topology diagram, thus forming the executive network component 140 shown in FIG. 1 . Each execution body constituting the execution body network component 140 is distributed in one or more CPUs and coprocessors connected to the CPUs constituting a heterogeneous architecture. Coprocessors such as GPUs, TPUs, etc. As schematically shown in FIG. 1 , each execution body of the execution body network component 140 is replaced by various tiny circles. Some small circles are connected in series with each other through dotted lines to form a data processing path. There are some branches in a data processing path. There are data processing paths that intersect each other to form a more complex relationship between two or more data processing paths. These data processing paths will always remain the same on this heterogeneous architecture.

When receiving the actual job data, the executive network component 140 slices the actual job data into task data, and the task data is continuously input into the data processing path, so as to complete the processing of the task data. Visually speaking, the same type of data fragments in the continuously input data will be fixedly input into the same data processing path, and the input data fragments are input like pipelines, and enter the data entry of the same data processing path in turn, and after processing The generated data is also automatically sent to the next downstream execution body in the data processing path until it flows through the entire data processing path.

FIG. 2 is a schematic diagram showing the influence of a specified component node selection unit on a specific execution body in the execution body creation system in the data processing network according to the present disclosure. As shown in Figure 2, in the system for big data computing and deep learning, the data processing network is composed of various executive bodies. For the convenience of description, only twelve are shown in Figure 2, marked as 21 and 22 respectively , ... 32. In actual application scenarios, there can be any number of executives based on needs. These executables perform predetermined operations on input data blocks. Although three data blocks are shown in Figure 2, in actual scenarios, the number of data blocks is massive.

As shown in FIG. 2, in the elliptical execution body 21-29, predetermined operation processing is performed on the externally input data blocks 1-3. The executable is usually a computing unit in a data processing device, such as a computing unit in a GPU, or a computing component. One of the executive bodies 21-29 may receive only one data input, or may receive multiple data inputs. Some executables do not need to input any content data.

In the deep learning data processing network, the executives are given predetermined processing tasks based on the determined data processing type in the initial stage. With the inflow of data blocks, each execution body is fixedly receiving the data block output from the upstream execution body and outputting or outputting the generated data block to its downstream execution body.

In the data processing network of the present disclosure, a data block includes header data and content data, and Table 1 below provides a data block structure table.

Table 1

As shown in Table 1 above, the header data determines that the metadata contained in the header data describes the specific content of the content data of the data block, and guides the specific location of the content data in the given storage space. By default, the execution body's operation on the data block is the operation on the header data and content data in the data block as a whole. However, in the specific data processing process, in order to achieve various purposes and eliminate some problems in the static data processing network, it is necessary to selectively operate on the data block, change the operation mode of the data block or modify the data block itself .

In the specific data processing process, since the specified component node selection unit 1213 will select the task node containing the specified component from the node library for the specific task node in the topology map corresponding to the predetermined data block or the predetermined execution body 20, Therefore, these task nodes have some specified node attributes. Hereinafter, unless otherwise specified, a task node generally corresponds to an executive body. Moreover, if it is not necessary to specify which one of the execution bodies 21-29 the execution body is, the symbol 20 is uniformly used instead, unless it is specifically specified.

In one scenario, the specified component node selection unit 1213 can select a task node with a specified component or additional node attribute for a task node of a certain execution body, so that the execution body corresponding to the task node specifies that the execution body uses data block range. For example, a data block as used in this disclosure includes header data or content data. Specify the task node selected by the component node selection unit 1213 to have header data that enables the predetermined execution body 25 to only use the data block 1-2 generated by the execution body 22 . In the data processing network, since the specified component node selection unit 1213 selects the task node corresponding to the execution body 25 to have a specified component that only uses the header data, the execution body 25 will change from the predetermined conventional operation mode to obtain The new operation mode after specifying the component, for example, only uses the header data of the data block 1-2 generated by the executable 22. In other words, since the corresponding task node has the designated component, the executive body 25 consumes the header data of the data block 1-2. Therefore, the executable body 25 only needs the specific value of the header data of the data block 1-2, and there is no need to obtain the content data of the data block 1-2 saved by the executable body 22. This will obviously speed up the flow of data from executive 22 to executive 25 .

In a static data processing network, an executive is a consumer of data, and may also be a producer of data at the same time. Therefore, in a data processing network composed of multiple executives, data blocks flow between different executives, and become a new data block after being processed or consumed by the executives, so as to be reserved by the next executive. As described below for Figure 3, after the downstream execution body consumes or uses the data block output by the upstream execution body, it will feed back the use completion message to the upstream execution body, so that the upstream execution body will empty the storage occupied by the output data block space in order to receive new data blocks that need to be processed. In actual scenarios, some operations performed by an executive body may require multiple input data blocks to complete the scheduled operation. It is very likely that after receiving the first input data block, the executable cannot immediately execute the operation because the second or third data block has not been received, so the output execution body of the first input data block The feedback message of the multi-input executive body will not be obtained quickly, so that the output executive body of the first input data block cannot be vacant for storing the storage space of the first input data block (such as the output data buffer of the executive body), thereby also No further feedback messages can be sent to the upper-level executive body. This causes a congestion situation in the data flow to propagate all the way to the previous stage. Specifically, as shown in FIG. 2 , for the executable body 29 , it needs to obtain the data block 2-2 output from the executable body 26 and the data block 3-3 output via the executable body 27 and the executable body 28 when performing operations. . Therefore, in the process of executing the operation, there may be that the execution body 29 has received the data block 3-3 output by the execution body 28 at the first time, but has not received the data block 2-3 output by the upstream execution body 26, so it cannot Execute a predetermined action. Therefore, even if the execution body 29 has obtained the data block 3-3, because the execution body 29 has not executed the predetermined operation and consumed the data block 3-3, the upstream execution body 28 cannot obtain the feedback message from the execution body 29, and cannot empty the data block 3-3. The storage space occupied by the owned data block 3-3 (state that cannot be rewritten). In this way, the execution body 28 cannot always send a confirmation message that the execution is completed to its upstream execution body 27, which also causes the storage space occupied by the data block 3 in the execution body 27 to be unable to be vacant, so eventually the execution body 27 is bound. The fixed storage space cannot receive new data blocks for continuous data processing. In this way, even if the downstream execution bodies 30 , 31 and 32 of the execution body 27 have completed predetermined operations, new data blocks cannot be obtained from the execution body 27 for subsequent processing. As a result, the executive bodies 30, 31, and 32 are also in a stalled waiting state. Therefore, because the execution body 29 is in a waiting state due to the delay in receiving the output data block from the execution body 26, the data processing network has data congestion on some task nodes related to the execution body 29. In order to eliminate this data flow blocking situation existing in the data processing process. For this purpose, the specified component node selection unit 1213 or the blocked task node elimination unit will select a task node with a specified component that holds the header data before the corresponding task node of the execution body 29 that is expected to have data flow blocking, such as the execution body The task node corresponding to 28 enables the execution body 28 to change the operation relationship between the data blocks to be processed in the case of having the specified component, thereby changing the predetermined operation mode of the execution body 28 on the data blocks, for example, The data storage mode of the processed data block is changed to a new data storage mode, for example, only the head data of the obtained data block is kept. Similarly, in order to cooperate with the selected task node corresponding to the execution body 28 to have the specified component that only holds the header data, it is required that the task node corresponding to the execution body 29 selected by the specified component node selection unit 1213 has only the execution body 28 The specified element of the header data of the generated data. In this way, because the executable body 28 only saves the header data, and because there is a specified component of "only use the header data" when the executable body 29 does not perform an operation, the output data cache of the executable body 28 is used to store Part of the content data is in an empty state, so that the execution body 28 can feed back a message to the execution body 27, so that the execution body 27 can keep its own output data cache in an empty state so as to be in a state where it can perform the next round of operations to obtain new data blocks . Therefore, the downstream execution bodies 30 - 33 of the execution body 27 can timely enter the processing of the next data block, thereby eliminating the impact of data flow congestion at the execution body 29 on the data processing operations of the relevant execution bodies.

As shown in FIG. 2 , in order to make the data block conform to the operation requirements of some executives during the operation, for example, the specified component node selection unit 1213 sends the specified component to the data block 2, thereby modifying its data attribute. For example, this attribute shows that the header data of the data block 2 informs the executable that operates on it that it can be modified by the subsequently received executable, for example, it can be modified by the executable 24, so as to meet the operational requirements of the executable 24. In summary, this is a change to the header data of the input data block.

The specified component node selection unit 1213 can select the task node with the specified component, so as to change the operation scope, operation mode and storage mode of the data block of the executive body, which eliminates the need to temporarily change the operation mode of the execution body during operation. That is, in the task node selection stage, a priori method is used to select the complete task node (no need to modify it at runtime), so as to eliminate the extra overhead at runtime. On the other hand, as shown in FIG. 2 , in order to make the output data block meet the operation requirements of the downstream execution body, the specified component node selection unit 1213 can select a task corresponding to the execution body with a specified component with modifiable header data. The node makes the corresponding execution body and data block affected by the specified component, for example, makes the data block 2-3 output by the execution body 26 change its head data content to meet the operation needs of the downstream execution body 29.

In addition, as shown in FIG. 2 , in order to eliminate unnecessary backward operations, the designated component node selection unit 1213 may select a task node corresponding to the executive body 22 that has designated constituent elements that eliminate backward operations, so that the executive body 22 Under the influence of the designated component, the backward operation on the executive 22 at the executive 23 is eliminated. In the data processing network involving deep learning systems, there are mostly forward operations and reverse operations, and these operations require execution by default. For the forward operation that is determined not to require the reverse operation, for example, because the specified component node selection unit 1213 of the present disclosure selects the task node corresponding to the executive body 22 with the specified component that eliminates the backward operation, the reverse operation performed by the executive body 23 is eliminated. operate. This results in significant savings in processing and overhead.

According to the specified component node selection unit 1213 of this disclosure, by selecting a specific task node, the executive body has more flexible operation functions, so as to realize the refined management of the storage space in the static data processing network, and make the execution of each task node Body hardware is optimized. By specifying the component node selection unit 1213 to select a task node with a specified component capable of modifying the header data of the data block for the corresponding executable body, so that the execution body can modify the header data of the data block, so that the execution body It is possible to perform more diverse operations on the data block beyond the inherent properties of the data block itself, so as to expand the diverse processing requirements of the data processing network for the data block. Due to the existence of the specified component node selection unit 1213, more convenience is brought to the writing of the program running in the data processing network.

Although in the description of the present disclosure, some executives created by the executive creation system in the data processing network are described as two independent entities, it is optional, but it does not mean that the separate existence of the two is necessary to realize the present disclosure arrangement, it is possible to combine the two.

When the executive creation system in the data processing network according to the present disclosure is used in the field of big data technology and deep learning and constitutes a distributed system, the fluency of data processing is very important. When the data of a task node is in the blocked state, it will cause the data processing of other parts to be suspended, which will cause the entire system to pause in data processing, so that the data flow processing is in a waiting state. In big data computing and deep learning, the execution body creation system in the data processing network according to the present disclosure is adopted. Since the data block contains content data and header data, the operation of the execution body with the specified component node selection unit 1213 is selected. relationship and the operation relationship between the execution body and the data block, and the task node receiving the specified component can change the predetermined operation relationship between the execution body and the data block, so that the storage space associated with the execution body can be finely managed, so that Some storage space can be efficiently utilized, and the hardware performance of the execution body is optimized to improve the efficiency of the execution body to continuously process data blocks.

FIG. 3 is a schematic structural diagram of an executive network created by the executive creation system in the data processing network according to the present disclosure. As shown in Figure 3, the large dotted box represents an executive body. In the executive body network component 140 shown in FIG. 3 , for convenience of illustration, only five executive bodies are shown. In fact, corresponding to the task topology diagram, as many task nodes as there are in the neural network, there are as many executive bodies in the executive body network component 140. Therefore, continuous small solid squares are used on the lower left side of FIG. 3 to represent other execution bodies not shown. body. FIG. 3 schematically shows the composition of each execution body constituting the present disclosure, which includes a message warehouse, a finite state machine, a processing component, and an output data cache. As can be seen from Figure 3, each executive body seems to contain an input data cache, but it is marked with a dotted line. Actually this is an imaginary building block, which will be explained in detail later. Each executive in the data processing path, such as the second executive in Figure 3, is established based on a task node in the neural network of the task topology graph, and based on the complete node attributes, forms the second executive and its upstream and downstream The topological relationship of the execution body, the message warehouse, the finite state machine and the processing method (processing component) and the cache location of the generated data (output data cache). Specifically, when the second execution body performs data processing, its task needs the output data of the third execution body upstream of it. When the third executor generates data to be output to the second executor, for example, when generating third data, the third executor will send a data ready message to the second executor to the second executor's The third data of the second execution body is already in the output data cache of the third execution body and is in an available state, so that the second execution body can read the third data at any time. The finite state machine of the second executor modifies its state after the message warehouse obtains the message from the third executor. If the processing component of the second executor generates data, such as the second data, after the last execution of the calculation task, and caches it in its output data buffer, and sends the data to the downstream executors of the second executor, such as the two first The executables (the first executable A and the first executable B) send out a message that the second data can be read.

After the first executor A and the first executor B read the second data and use them, they will respectively feed back messages to the second executor, informing the second executor that the second data has been used. Therefore, the second executor The body's output data cache is empty. At this time, the finite state machine of the second executive will also modify its state.

In this way, when the state change of the finite state machine reaches a predetermined state, for example, when the input data (such as the third data) required for the execution of the operation of the second executive body is in the obtainable state and its output data cache is in the empty state, then notify The processing component reads the third data in the output data cache of the third executive, and executes the specified computing task, thereby generating the output data of the executive, such as new second data, and storing it in the output of the second executive data cache.

Similarly, when the second executive body completes the specified calculation task, the finite state machine returns to its initial state, waiting for the next state change cycle, and at the same time, the second executive body feeds back the message that the use of the third data is completed to the third executive body. The message warehouse of the third executor sends a message that the second data has been generated to the first executor A and the first executor B, and informs the first executor A and the first executor B that the second data is ready to be read fetched state.

After the third execution body obtains the message that the second execution body has used up the third data, the output data buffer of the third execution body is in an empty state. Similarly, after the second executor obtains the message that the first executor A and the first executor B have finished using the second data, the output data cache of the second executor is in an empty state.

The above-mentioned process of executing tasks of the second executable body also occurs in other executable bodies. Therefore, under the control of the finite state machine in each executive, based on the output of the upstream executive, the same kind of tasks are processed cyclically. Therefore, each executive body is like a fixed task post personnel on a data processing path, thereby forming a pipeline processing of data without any other external instructions.

It should be pointed out that, as shown in FIG. 3 , the specified group in the task node selected by the specified component node selection unit 1213 in the first executive body A and the first executive body B, the second executive body, and the third executive body Elements (such as the first specified component, the second specified component, and the third specified component), so as to perform specified operations on specified data. For example, the first executive body A and the first executive body B (although two first executive bodies are exemplified here, it may only have one first executive body or more first executive bodies). A specified component is a specified component that enables the first executable body to use only the header data of some upstream executable body's output data, and a third specified component in the third executable body is a specified component that enables the third executable body to modify the data The specified element of the header data and the fourth specified element in the fourth executable are the specified elements such that the fourth executable does not require a backward operation. It should be pointed out that although the second specified component is used here to explain the characteristics of the second executable, the specified function of the second executable here is not necessarily directly determined by the specified component node selection unit 1213 , but is determined by the task node with the specified component selected by the specified component node selection unit 1213, that is, a special task node formed directly in the process of forming a task node, the attribute of the task node itself includes its Only the function of saving the header data, so that the executive body corresponding to the task node executes the function of only saving the header data. The same goes for other executive bodies. In other words, the execution body contains one of the node attributes described by the specified component contained in the corresponding task node. However, for the convenience of description, the specified component is still used in each executive body to describe the specific attributes or functions contained in this special executive body itself.

In the data storage method of the present disclosure, header data and content data are stored separately. As described above with respect to FIG. 3 , since the first execution body A and the first execution body B have specified components that only use the header data of the data, the upstream second execution body only holds the head data of the data. Because the second execution body has only carried the header data of the data in the output data cache of the third execution body (the header data is not stored in the output data cache but is stored in the CPU connected to the coprocessor, that is The header data is stored separately from the content data in the output data cache), so the second execution body feeds back to the third execution body a message that the content data of the third data has been used up, so the third execution body outputs the data in the data cache for The space for content data is vacated. The output data cache of the third executable only needs to keep the header data of the second data for the second executable. Therefore, the upstream execution body of the third execution body will also obtain the feedback message of the third execution body, thereby eliminating the phenomenon that the upstream execution body of the third execution body is blocked.

The first specified component in the first execution body A and the first execution body B is the specified component of the header data that only uses the output data of some upstream execution bodies, so that it is specific to the output data of the second execution body The data in the cache forms a consumption relationship that only uses header data. After both the first executive body A and the first executive body B have read the header data in the output buffer of the second executive body, they will feed back a message to the second executive body, so that the second executive body will send another header data in the next cycle. Part of the data is moved to its output data cache.

The third specified component in the third executable is a specified component that can modify the header data of the data. In order to make the output data block meet the operation requirements of the downstream executable (such as the second executable) of the third executable, The specified component node selection unit 1213 can make the third specified component between the third executable body and the data block, so that the header data of the data block contains "usable" metadata, so that the third executable body can The data output by the upstream executive is modified.

Although the basic composition of the executive body and the operational relationship between directly upstream and downstream executive bodies are described above for Figure 3, the processing components of some executive bodies may not perform actual operations at all, but only move data and change the data structure. Position, that is, a simple handling executive. For example, the processing component of the second executor only transfers the header data of the data acquired from the third executor to its output data cache without performing any transformation on the data acquired by the third executor. The existence of this kind of carrying executive body can eliminate the blockage of the entire upstream of the data processing path and the processing suspension of other branches caused by some execution bodies causing the blockage to be transmitted upstream.

By setting the second executor that only saves header data before the first executor A and the first executor B, the first executor A and the first executor B can know based on the message from the second executor, When operating on the data, only the header data transported by the second executable and stored in its output data buffer is read. When one of the first executor A and the first executor B is waiting for other input data, the second executor only needs to keep its header data without actually storing content data. Therefore, the third executor can put the output data buffer of the third executor in an empty state after receiving the feedback message from the second executor, so that the third executor can communicate with the first executor A and the first executor downstream of the third executor. When B is in the waiting blocking state, it can still process new data. If there is no second executor, and the first executor A and the first executor B are directly connected to the third executor, the third executor will not be able to It makes its output data cache empty (the content data is very large and will take up a lot of memory), which will cause its upstream execution body to be in a stagnant state and unable to perform operations. Therefore, the specified component node selection unit 1213 selects the task node corresponding to the first execution body A and the first execution body B that contains the first specified component that only uses the header data, and in the first execution body A Before the task node corresponding to the first executive body B, add the task node corresponding to the second executive body containing the second specified component that only holds the header data, so that the first executive body A and the first executive body B When performing actual data processing, only the head data of the output data of the direct upstream execution body is obtained, so that all the upstream data of the second execution body can make the output data cache empty in time and be in a state where data processing can be performed. Through the components for accelerating data processing according to the present disclosure, the blockage in the data processing path is eliminated, so that when there is an execution body task node in the waiting state in the path, the normal operation of its upstream execution body is not affected, thereby speeding up data flow.

The first executable A and the first executable B can obtain the content data similar to the data generated by the third executable through other executables to complete their final calculation. As shown in Figure 3, the fourth executable, as the upstream executable of the first executable A and the first executable B, can output content data similar to the data generated by the third executable, so that the first executable Provide a solution when A and the first executive body B need content data. In the deep learning neural network, there are often some teammate-type task nodes. The execution bodies created and described based on the task nodes are exactly the same. The third execution body and the fourth execution body shown in Figure 3 belong to the deep learning neural network. Teammate executive task nodes in the network. This belongs to the prior art in this field and will not be described in detail here.

As mentioned earlier, each executive in Figure 3 shows that it contains an input data cache, but it does not actually contain it, because each executive does not need any cache to store the data to be used, but only It is only necessary to obtain the required data in a state that can be read. Therefore, when the data to be used by each execution body is not in a specific execution state, the data is still stored in the output data cache of its upstream execution body. Therefore, for visual display, the input data cache in each executive body is represented by a dotted line, which does not actually exist in the executive body. In other words, the output data cache of the upstream execution body is the virtual input data cache of the downstream execution body. Therefore, in FIG. 3, a dotted line is used for the input data cache.

Refer back to Figure 1. As shown in FIG. 1 , the execution body creation system for heterogeneous architectures according to the present disclosure also includes a job description component 110, which is used to describe the job neural network model, the number of neural network layers, and the neural network in each layer of neural network based on the job type. the number of dollars. Specifically, the job description component 110 describes the computing resources required by the job and what kind of computing needs to be performed. For example, the job description is used to explain whether the job is used for image classification or speech recognition classification, the number of layers of the neural network required, the number of task nodes for each layer, the connection between layers, and the process of performing data processing Enter the storage location of the data. Describing jobs is a prior art. The job description component 110 of the present disclosure adopts a separation description method, splits the object to be described into several related dimensions, differentiates and describes from several aspects or dimensions, and describes the orthogonal relationship between several dimensions. Since the described dimensions describe jobs from different dimensions in a way of distinguishing and separating each other, and are in an orthogonal relationship with each other, each dimension does not interfere with each other, and the description of tasks does not need to consider the correlation between dimensions, so it can be significantly reduced. The complexity of the program codes running in the executable creation system of the heterogeneous architecture of the present disclosure also significantly reduces the intellectual burden of programmers who write these program codes. Although the job description component 110 is shown in FIG. 1 . However, the object of the present disclosure can also be achieved by using an existing job description component.

Although the heterogeneous architecture according to the present disclosure as shown in FIG. 1 includes one or more central processing units and at least one coprocessor device connected thereto, the system shown in FIG. 1 may include The gateway component may also include a direct communication component between coprocessors, such as GPUs, such as the larger circles connected between two GPUs by dashed lines in FIG. 1 .

Furthermore, according to another aspect of the present disclosure, the above components also perform a method for accelerating data flow in a data processing path of a coprocessor. Among the multiple execution bodies in the data processing path, the second execution body with the second designated component only uses the header corresponding to the output data block stored in the output data cache of the third execution body as its upstream execution body The data is moved to the local output data cache, and a message that the data can be read is sent to one or more first executable bodies having the first designated component among the plurality of executable bodies. One or more of the plurality of execution bodies in the data processing path, one or more first execution bodies having the first designated component, based on the message from the second execution body, output data from the second execution body The header data is read from the cache, and a predetermined operation is performed. The second executor sends a message to the third executor of the second executor while moving the header data to the local output data cache, so that the third executor puts its own output data cache in an empty state . The first execution body acquires data similar to the content data of the data block generated by the third execution body from the output data cache of the fourth execution body located upstream to perform a predetermined operation. The second executive body puts its own output data buffer in an empty state after receiving all the feedback messages from the first executive body.

The basic principles of the present disclosure have been described above in conjunction with specific embodiments. However, it should be pointed out that those of ordinary skill in the art can understand that all or any steps or components of the methods and devices of the present disclosure can be implemented on any computing device (including processors, storage media, etc.) or a network of computing devices, implemented with hardware, firmware, software, or a combination thereof, this is a person of ordinary skill in the art who uses their basic knowledge after reading the descriptions of the present disclosure. programming skills will do.

Therefore, the object of the present disclosure can also be achieved by running a program or a group of programs on any computing device. The computing device may be a known general-purpose device. Therefore, the object of the present disclosure can also be achieved only by providing a program product including program codes for realizing the method or device. That is, such a program product also constitutes the present disclosure, and a storage medium storing such a program product also constitutes the present disclosure. Obviously, the storage medium may be any known storage medium or any storage medium developed in the future.

It should also be pointed out that, in the apparatus and method of the present disclosure, obviously, each component or each step can be decomposed and/or reassembled. These decompositions and/or recombinations should be considered equivalents of the present disclosure. Also, the steps for performing the above series of processes may naturally be performed in chronological order in the order described, but need not necessarily be performed in chronological order. Certain steps may be performed in parallel or independently of each other.

The specific implementation manners described above do not limit the protection scope of the present disclosure. It should be apparent to those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present disclosure shall be included within the protection scope of the present disclosure.

Claims (8)

1. An executive creation system, comprising: The task topology map generation component is used to decompose the required job into a series of tasks executed by the executive in the heterogeneous architecture, and generate tasks based on the inherent relationship between the decomposed tasks while the job is decomposed A relational topology diagram, the task nodes of the task topology diagram contain all the node attributes required to perform corresponding tasks, and the task topology diagram generation component includes a specified component node selection unit, which is used in the process of generating the task relationship topology diagram according to the specific a task selects a task-specific node having a specified component that includes an additional node attribute that changes the manner in which the task-specific node performs the corresponding task; and Executor creation component, based on the task relationship topology graph and all node attributes of each task node or all node attributes containing specified components, creates a corresponding executor for each task node in the computing resource, and makes it associated with additional node attributes The specific execution body corresponding to the specific task node executes the corresponding task according to the additional node attributes. 2. The execution body creation system according to claim 1, wherein the additional node attribute contained in the designated component is a data operation node attribute, which is used to make the created specific execution body execute only on the header data of the processed data operate. 3. The execution body creation system according to claim 1, wherein the additional node attribute contained in the specified component is a data storage node attribute, which is used to make the created specific execution body only store the header data of the processed data. 4. The execution body creation system according to claim 1, wherein the additional node attribute included in the specified component is a data modification node attribute, which is used to enable the created specific execution body to modify the header data of the processed data. 5. A method for creating an executive body, comprising: Decompose the job to be completed into a series of tasks performed by the executive in the heterogeneous architecture, and generate a task relationship topology map based on the inherent relationship between the decomposed tasks while the job is decomposed. The task topology The task node of the graph contains all the node attributes required to execute the corresponding task; In the process of generating the task relationship topology diagram, selecting a specific task node with a specified component including an additional node attribute according to the specific task, and the additional node attribute changes the way the specific task node executes the corresponding task; and Based on the task relationship topology graph and all node attributes of each task node or all node attributes containing specified components, create a corresponding execution body for each task node in computing resources, and make it correspond to specific task nodes with additional node attributes The specific executor executes the corresponding task according to the additional node attributes. 6. The execution body creation method according to claim 5, wherein the additional node attribute contained in the designated component is a data operation node attribute, which is used to make the created specific execution body execute only on the header data of the processed data operate. 7. The execution body creation method according to claim 5, wherein the additional node attribute contained in the designated component is a data storage node attribute, which is used to make the created specific execution body only store the header data of the processed data. 8. The execution body creation method according to claim 5, wherein the additional node attribute contained in the designated component is a data modification node attribute, which is used to enable the created specific execution body to modify the header data of the processed data.