KR102266324B1 - Worker node management method of managing execution platform and platform system for the same - Google Patents

Abstract

The present invention relates to a method for managing worker nodes in a machine learning execution management platform and a platform system therefor, which improve reconfiguration performance of the worker nodes by increasing operation speed for addition or deletion of the worker nodes, and can flexibly respond to situations occurring within the platform. The method includes the steps of: registering the node to be deleted to a deletion list among the worker nodes recorded in the node list in a session node; searching for the highest priority worker node to perform machine learning in the session node which has been requested to assign a task from a master node; comparing the priority ID indicating the highest priority worker node in the session node and the deletion registration ID registered in the deletion list; deleting the highest priority worker node from the node list in the session node, when the priority ID and deletion registration ID are the same; allocating tasks so that machine learning is performed in the worker node corresponding to the priority ID in the session node, when the priority ID and deletion registration ID are different; and performing machine learning on the worker node to which the task is assigned according to a machine learning execution command.

Description

WORKER NODE MANAGEMENT METHOD OF MANAGING EXECUTION PLATFORM AND PLATFORM SYSTEM FOR THE SAME}

The present invention provides a method for managing worker nodes in a machine learning execution management platform that can improve the reconfiguration performance of worker nodes by increasing the operation speed for addition or deletion of worker nodes and flexibly respond to situations occurring within the platform, and a method for managing the same It is about the platform system.

Machine learning is a field of artificial intelligence (AI) that has evolved from the study of pattern recognition and computer learning theory. It is a system and algorithm that performs learning and prediction based on empirical data and improves its own performance. build

The integrated remote execution platform for machine learning consists of a master node, a session node, and a worker node from the upper layer. It also has a separate external network storage, and users access the platform through an external program.

Therefore, when a user accessing the platform registers a task that is a target to perform machine learning, the registered task is first allocated to a session node in accordance with a framework to perform machine learning.

A session node assigns a task to a worker node that can perform the task and has the best computational processing power, and when a user issues a machine learning-related command to be performed in the worker node, machine learning is performed according to the user's command.

However, the conventional integrated remote execution platform for machine learning has a problem in that it is difficult to process when reconfiguration is necessary because the worker nodes are arranged and allocated only based on flops, which are CPU/GPU computational processing power.

That is, the situation requiring reorganization of the worker node list, such as additional insertion, arbitrary removal, or deletion of worker nodes, is not considered, so there is a problem in that it is difficult to handle the case where the number of worker nodes is changed, such as expansion or reduction of the platform.

US registered patent US 16/248,560 Republic of Korea Patent Publication No. 10-2012-0001688

The present invention is to solve the problems described above, and by increasing the operation speed for adding or deleting worker nodes, it improves the reconfiguration performance of the worker nodes and executes machine learning that can flexibly respond to situations occurring within the platform An object of the present invention is to provide a method for managing worker nodes in a management platform and a platform system for the same.

To this end, a method for managing a worker node in a machine learning execution management platform according to the present invention includes: a deletion node registration step of registering a node to be deleted from among the worker nodes recorded in the node list in the session node to the deletion list; a worker node search step of searching for a worker node with the highest priority to perform machine learning in the session node requested for task assignment from the master node; a node validity determination step of comparing a priority ID indicating the highest priority worker node with a deletion registration ID registered in the deletion list in the session node; a node list cleaning step of deleting the highest priority worker node from the node list in the session node when the priority ID and the deletion registration ID are the same; a task assignment step of allocating a task so that machine learning is performed in a worker node corresponding to the priority ID in the session node when the priority ID and the deletion registration ID are different; and a machine learning step of performing machine learning in a worker node to which the task is assigned according to a machine learning execution command.

At this time, a worker node analysis step of analyzing the performance of the worker nodes in the session node; and a node list generating step of generating the node list including a heap data structure according to the performance order of the analyzed worker nodes.

In addition, the session node classifies the performance of the worker node into flops, which are computational processing capabilities of the CPU and GPU, and it is preferable to sort the node list by using gigaflops (GFlops) as a unit.

In addition, a deletion table generation step of generating a deletion list table in which the deletion registration ID is recorded in the session node; and an exclusion node analysis step of detecting a deletion target worker node registered in the deletion list table in the master node.

In addition, it is preferable that the delete list table is a table of a hash map structure in which mapping is performed according to a key value generated by a hash function.

In addition, the method further comprises a node assignment request step of receiving a request from the master node to allocate a task on which machine learning is performed to a worker node, wherein the master node is composed of a framework in which the machine learning is performed. It is preferable to request assignment of the task to the session node that manages the worker node.

In addition, in the deletion node registration step, the master node extracts a randomly excluded or faulty worker node and requests deletion from the session node, and the session node registers the deletion registration ID in the deletion list of the hash map structure. it is preferable

In addition, in the worker node search step, it is preferable that the session node searches for a root node of the highest priority among the node list including the heap data structure as the highest priority worker node.

Preferably, in the node validity determination step, the priority ID of the worker node corresponding to the root node in the session node is compared with the deletion registration ID registered in the deletion list.

In addition, in the node list cleaning step, when the priority ID and the deletion registration ID are the same, it is preferable to perform a pop operation to delete the root node on the node list and then reconstruct the node list in a heap data structure method. .

Preferably, the method further includes a deletion list update step of excluding the deletion registration ID of the root node deleted in the pop operation from the deletion list.

Preferably, in the task assignment step, the session node assigns the task to a worker node corresponding to the root node.

Preferably, the session node allocates a task to a worker node corresponding to the root node, and after performing a pop operation to delete the root node, the node list is reconstructed in a heap data structure method.

On the other hand, the machine learning execution management platform on which the worker node management method according to the present invention is performed comprises: an external input module for receiving a task on which machine learning is performed; a worker node that performs machine learning on the input task and is registered in a plurality of node lists; a session node for registering a node to be deleted among the worker nodes recorded in the node list in the deletion list; and a master node that receives a request for task assignment from a user and instructs the session node to search for the highest priority worker node to perform the machine learning. The ID and the deletion registration ID registered in the deletion list are compared, and when the priority ID and the deletion registration ID are the same, the session node deletes the ID of the highest priority worker node from the node list, and the priority ID and When the deletion registration ID is different, it is preferable that the session node allocates the task so that machine learning is performed in the worker node corresponding to the priority ID.

The present invention as described above improves the internal speed of the previously implemented machine learning platform through the node list recording the worker nodes and the deletion list recording the nodes that need to be excluded from among the worker nodes.

In addition, the worker node list is implemented as a heap data structure to improve the reorganization, insertion, and selection speed of the node list, and to compensate for the slow speed through the delete list for search and delete operations that are slow in the heap data structure.

1 is an overall configuration diagram of a machine learning execution management platform system according to the present invention.
2 is a block diagram of a node list and a delete list according to the present invention.
3 is a flowchart illustrating a method for managing worker nodes in a machine learning execution management platform according to the present invention.
4 is a detailed flowchart illustrating a deletion node registration step according to the present invention.

Hereinafter, a method for managing a worker node in a machine learning execution management platform according to a preferred embodiment of the present invention and a platform system therefor will be described in detail with reference to the accompanying drawings.

1 , the machine learning execution management platform system according to the present invention includes an external input module 10 , a worker node 20 , a session node 30 , and a master node 40 . In a preferred embodiment, it further includes an external network storage (DB-O) and an internal platform management DB (DB-I).

According to this configuration, the master node 40 registers a task input from the user through the external input module 10 , and the session node 30 allocates the registered task to the worker node 20 to the worker node. In (20), machine learning is performed on the task.

In particular, as shown in FIG. 2 , the present invention is a machine learning platform previously implemented through the node list 31 recording the worker node 20 and the deletion list 32 recording the nodes that need to be excluded among the worker nodes 20 . improve speed.

That is, instead of directly searching for the worker nodes 20 recorded in the node list 31, at least one worker node to be excluded by referring to IDs of the worker nodes 20 registered in advance in the deletion list 32 ( 20) should be checked quickly.

In addition, the list of worker nodes 20 is implemented as a heap data structure to improve the reorganization, insertion, and selection speed of the node list 31, and search and deletion operations, which are slow in the heap data structure, are described above. Implement a delete list 32 to compensate for speed.

Accordingly, the present invention improves the reconfiguration performance of the worker node 20 by increasing the operation processing speed for addition or deletion of the worker node 20, and enables a flexible response to situations occurring in the platform.

Prior to a detailed description of the characteristic functions of the present invention, first looking at the integrated remote performance management technology applicable to the present invention, each node 20, 30, 40 is implemented in a hierarchical structure to transmit commands and allocate resources. do.

Each node including the worker node 20, the session node 30, and the master node 40 is a kind of computing module capable of processing a process for machine learning, and these nodes are external network storage (DB). -O) to perform data processing.

In addition, the worker node 20, the session node 30, and the master node 40 may be integrally implemented in one computing module, or some resources may be implemented in an external third module, and the functions of each node are different from each other. It can be implemented together in Node.

The external network storage (DB-O) provides a dataset of machine learning, and the worker node 20, which sets the machine learning execution environment, downloads the dataset from the external network storage (DB-O) to learn machine learning. carry out

A user accesses the platform of the present invention through the external input module 10 . The external input module 10 is implemented as an external program such as an API (Application Programming Interface), for example, and thus easily accesses the platform.

In addition, the user inputs a task, which is a target to perform machine learning, by using the external input module 10 . The input task is assigned to the session node 30 according to the framework to perform machine learning. The session node 30 assigns the corresponding task to the worker node 20 .

At this time, when the user inputs a machine learning related command to be performed in the worker node 20 through the external input module 10 , the corresponding command is delivered to the worker node 20 and executed. For example, when you issue a 'running machine command', machine learning starts.

The master node 40 monitors the resource of the worker nodes 20 in the platform, checks the machine learning progress of the task, uploads the task to the platform, transfers the user's command input from the external input module 10, and the task and dataset has the ability to manage

The session node 30 manages the worker nodes 20 composed of frameworks classified into the same or the same group. Each session node 30 monitors the resources of the worker nodes 20 and transmits a command received from the master node 40 to the worker node 20 .

The worker node 20 performs machine learning corresponding to the task received from the session node 30 . Also, it periodically reports the resource status of the worker node 20 to the session node 30 .

Therefore, the worker node 20 sets the machine learning execution environment through the 'task executor', downloads the machine learning dataset from the external network storage (DB-O), and when the machine learning execution command is delivered, the machine in the prepared execution environment run a run

Also, the resource status of the worker node 20 is reported to the session node 30 through the 'resource management module'. As the resource status is reported to the session node 30 in real time or periodically, it is used to select the worker node 20 that performs the task assigned to the platform.

On the other hand, as described above, the present invention improves the reconfiguration performance of the worker node 20 by increasing the operation processing speed for addition or deletion of the worker node 20 and flexibly responds to situations occurring within the platform. do.

As shown in FIG. 2 , the present invention is a machine learning platform previously implemented through a node list 31 recording a worker node 20 and a deletion list 32 recording a node that needs to be excluded from among the worker nodes 20 . Increases internal speed.

In addition, the list of worker nodes 20 is implemented as a heap data structure to improve the reconstruction, insertion, and selection speed of the node list 31, and to compensate for the slow search and delete operation speed in the heap data structure, the delete list ( 32) is implemented.

Therefore, the node list 31 and the deletion list 32 have meanings as information storage units or data structures specially constructed or newly added in the present invention.

In this case, the node list 31 is a data structure for managing the worker node 20 in the session node 30 . In the node list 31 , worker nodes 20 to which machine learning tasks are assigned are recorded. As will be described later, the node list 31 uses a heap data structure as an embodiment.

In addition, the sorting of the node list 31 is based on the flop, which is the computational processing power of the CPU/GPU, and the unit uses giga-flops (GFlops), and the flops of each CPU core and the flops of the GPU are added to the CPU and GPU. Consider the computational processing power of both hardware.

Flop: This indicates the arithmetic processing speed of the unit (Flops floating operations per second) is a computer, gigaflops GFLOPs of the image is the execution frequency of the first floating-point operations per second of the computer units billion (= 10 ^9).

In this way, the task input from the user is sequentially assigned from the worker node 20 having the best computational processing power, and machine learning is performed on the task input from the worker node to which the task is finally assigned.

The deletion list 32 is a list proposed by the present invention to improve deletion performance in the node list 31. When a heap data structure is applied to the node list 31, the node search and deletion operation speed is reduced. This is to make up for the very poor shortcomings.

In order to supplement the performance of the delete operation, when the session node 30 requests the deletion of the worker node 20, the ID of the worker node 20 to be deleted (hereinafter, 'deletion registration ID') instead of directly deleted from the node list 31 ') is stored in the deletion list 32 and used.

The deletion list 32 enables, for example, to perform ID insertion, search, and deletion operations of the worker node 20 at high speed using a hash map structure. Thereafter, when a request for assignment of a new machine learning task is transmitted to the session node 30 , selection and assignment of the worker node 20 is performed.

Selection and allocation of worker nodes is selected from available resources capable of performing machine learning, and the highest-priority worker node according to computational processing power is selected from the node list 31 . The highest priority is, by way of example, a root node, which is the highest node in the heap data structure.

In addition, when referring to the root node, if the ID of the worker node 20 is a worker node 20 registered in the deletion list 32, a pop operation is performed on the node list 31 without assigning a machine learning task to the corresponding worker. The node 20 is deleted.

This process proceeds until a machine learning task is assigned as the worker node 20 capable of properly performing machine learning becomes a root node, or until all elements of the node list 31 are exhausted.

To this end, as shown in FIG. 3 , the worker node management method in the machine learning execution management platform according to the present invention includes a deletion node registration step (S10), a worker node search step (S20), and a node validity determination step (S30).

In addition, the present invention includes a node list cleaning step (S40), a task assignment step (S50) and a machine learning step (S60), and as a preferred embodiment, the node allocation request step (S20a) and the deletion list update step (S41) include more

The present invention with these technical configurations improves the structure of the node list 31 and also introduces the deletion list 32 to more efficiently manage the worker node 20 compared to the existing machine learning execution management platform.

In the case of the node list 31, the speed of sorting and searching is increased by changing from the existing list-type data structure to a heap data structure. Furthermore, a delete list 32 is introduced to compensate for the shortcomings of searching and deleting the heap data structure.

That is, instead of searching and deleting the machine learning availability of the worker nodes 20 recorded in the node list 31 in the session node 30 in real time, when assigning the machine learning task, the deletion list 32 is allocated by querying. The availability of the desired worker node 20 is immediately checked and the best worker node 30 can be selected.

In addition, when a deletion request is made, the ID of the worker node 20 to be deleted is stored in the deletion list 32 instead of being deleted immediately from the node list 31 . Thereafter, when a new task is input, if the worker node 20 to which the machine learning task is to be assigned is the worker node 20 included in the deletion list 32, a method of deleting the task without assigning it is used.

More specifically, in the deletion node registration step S10 , a node to be deleted among the worker nodes 20 recorded in the node list 31 in the session node 30 is registered in the deletion list 32 . The worker node 20 registered in the deletion list 32 is a node that cannot perform machine learning due to random removal or node failure.

To this end, the master node 40 extracts the randomly excluded or faulty worker node 20 and requests the session node 30 to delete it. As a preferred embodiment, the session node 30 receiving the deletion request from the master node 40 registers the deletion registration ID in the deletion list 32 of a hash map structure as described below.

As shown in FIG. 2 , in the deletion node registration step S10 , the worker node 20 of the node list 31 is referred to, and the exclusion target worker node 20 is recorded in the deletion list 32 . Therefore, it is necessary to generate the node list 31 and the deletion list 32 prior to this.

The generation of the node list 31 includes a worker node analysis step S1 and a node list generation step S2, among which, in the worker node analysis step S1, the performance of the worker nodes 20 in the session node 30 Analyze the list order in the list.

To this end, the session node 30 may monitor the worker nodes 20 or transmit computing resources that the worker node 20 can provide to the session node 30 in real time or periodically. The performance of the worker node 20 may be preferably classified into the computational processing power of the CPU and the GPU.

In the node list generation step S2 , the node list 31 composed of a heap data structure is generated according to the performance order of the worker nodes 20 . Preferably, the node list 31 is classified into flops, which are computational processing capabilities, and is sorted using giga-flops (GFlops) as a unit during classification.

The heap in the heap data structure satisfies the heap property as a data structure of a complete binary tree designed to speed up the operation to find the maximum and minimum values.

The heap property is that, for example, if A is a parent node of B, a case-to-order relationship is established between the key value of A and the key value of B. A heap in which the key value of the parent node is always larger than the key value of the child node is referred to as the 'maximum heap'. and the opposite is called 'minimal heap'.

At this time, since the present invention determines the stack in the order in which the computational processing power of the worker nodes 20 is good, the worker node 20 with the good computational processing capability follows the maximum heap with a larger key value, so the root node is the highest priority. become a node.

Also, in FIG. 2 , the deletion list 32 is generated through the deletion table creation step S3 and the exclusion node analysis step S4 . In this case, in the deletion table creation step ( S3 ), the session node 30 creates a deletion list table in which the deletion registration ID is recorded.

The delete list table preferably has a hash map structure in which mapping is performed according to a key value generated by a hash function. A hash map is a data structure used to add an associative array, a structure that can map a key to a value, and calculates an index into an array such as a bucket or slot using a hash function.

In the exclusion node analysis step S4, the master node 40 monitors each worker node 20 to detect the deletion target worker node 20 to be registered in the deletion list table. In the present invention, deletion is a concept including exclusion and means excluded from machine learning.

The deletion target worker node 20 detected by the master node 40 is notified to the session node 30 in the above-described deletion node registration step S10 , and the session node 30 is deleted from among the worker nodes 20 . A target node is registered in the deletion list 32 .

When the session node 30 receives the information of the worker node 20 to be deleted and inserts the ID into the deletion list 32, it can ignore the duplicate worker node 20 deletion request using a hash map structure, A plurality of worker nodes 20 deletion requests may be included in the deletion list 32 .

On the other hand, as above, after registering the node list 31 of the worker node 20 and the worker node 20 to be deleted among them in the deletion list 32 and recording the deletion changes, the newly input machine learning task It is assigned to the worker node (20).

Accordingly, in the node assignment request step ( S20a ), the session node 30 is requested from the master node 40 to allocate a task on which machine learning is performed to the worker node. The master node 40 requests assignment of a task to the session node 30 that manages the worker node 20 configured as a framework in which machine learning is performed.

Thereafter, in the worker node search step ( S20 ), the highest priority worker node 20 to perform machine learning is searched for in the session node 30 , which has been requested to be assigned a task. Here, the search is a process of finding and determining the order of the worker nodes 20 having the best computational processing power regardless of the deletion list 32 .

As described above, when the node list 31 is constructed in the heap data structure, the session node 30 searches for the highest root node among the node list 31 formed in the heap data structure. The root node corresponds to the highest priority worker node 20 .

In the case of the maximum heap, where the key value of the parent node always has the heap property than the key value of the child node in the heap property, the stack of the nodes is determined in the order of CPU/GPU processing power, so the root node is the highest-priority worker node (20 ) is specified, so when the worker node 20 is allocated, it is first searched for.

Next, the node validity determination step (S30) is a step of determining whether the worker node 20 retrieved with the highest priority is registered in the deletion list 32. A priority ID and a deletion list for designating the highest priority worker node 20 ( 32) and compare the registered deletion registration ID.

To this end, the session node 30 extracts the highest priority worker node 20 through a search when a new machine learning task is input, and the extracted highest priority worker node 20 is combined with the nodes registered in the deletion list 32 . Search for a match.

When the node list 31 in which the worker nodes 20 are recorded is formed of a heap data structure, the session node 30 includes the priority ID and the deletion list 32 of the worker node 20 corresponding to the root node among the heap data structures. Compare the deletion registration ID registered in .

Next, in the node list cleaning step (S40), if the worker node 20 corresponding to the searched highest priority (root node) is registered in the deletion list 32, the highest priority worker node 20 is removed from the node list 31 Remove.

If the priority ID and the deletion registration ID are the same, the worker node 20 searched for task assignment is unsuitable or impossible to perform machine learning, so the session node 30 returns the searched highest priority worker node 20 to the node list ( 31) will be deleted.

At this time, the session node 30 deletes the root node from the node list 31 of the heap data structure by performing a pop operation. In addition, after the existing root node is deleted according to the pop operation, the node list 31 is reconstructed in the heap data structure method.

Since the pop operation removes the data at the top of the current stack and deletes it, the node list 31 is reconstructed with the data below it after the current root node is deleted through the pop operation.

After reconstructing the node list 31 , the deletion list 32 may also be reconstructed. The reconfiguration of the deletion list 32 is implemented by the deletion list update step S41 of excluding the deletion registration ID of the root node deleted in the pop operation from the deletion list 32 .

Next, in the task assignment step ( S50 ), if it is determined that the retrieved highest priority worker node 20 is not registered in the deletion list 32 , the machine learning task newly input to the retrieved highest priority worker node 20 is added allocate

Specifically, when the priority ID and the deletion registration ID are different, the session node 30 assigns a task to the corresponding worker node 20 so that machine learning is performed in the worker node 20 corresponding to the priority ID.

When the highest priority worker node 20 is the root node of the node list 31 , a task is assigned to the worker node 20 corresponding to the root node in the session node 30 . That is, the task is assigned to the best worker node 20 among the worker nodes 20 not registered in the deletion list 32 .

After allocating the task, it is preferable to update the node list 31 as well. This is to allow the next worker node 20 to be used by excluding the worker node 20 currently used for machine learning.

Accordingly, the session node 30 allocates a task to the worker node 20 corresponding to the root node of the highest priority, and performs a pop operation for deleting the allocated root node (S51). Also, after the pop operation is performed (S51), the node list 31 is reconstructed in a heap data structure method.

Thereafter, in the machine learning step (S60), machine learning is performed in the worker node 20 to which the task is assigned according to the machine learning execution command. The machine learning execution command is transmitted after receiving the user's command input through the external input module 10 from the master node 40 .

At this time, the worker node 20 sets the machine learning execution environment through the 'task executor', downloads the machine learning dataset from the external network storage (DB-O), and when the machine learning execution command is delivered, machine learning in the prepared execution environment carry out

In addition, the resource status of the worker node 20 is reported to the session node 30 through the resource management module. As the resource status is reported to the session node 30 in real time or periodically, it is subsequently used to select the worker node 20 to perform the task assigned to the platform.

Hereinafter, a machine learning execution management platform system in which the method for managing worker nodes according to the present invention is performed will be described with reference to the accompanying drawings.

However, the machine learning execution management platform system according to the present invention is applied to the worker node management method described above. Therefore, redundant descriptions will be omitted below as much as possible.

1 and 2 above, the machine learning execution management platform system according to the present invention includes an external input module 10 , a worker node 20 , a session node 30 and a master node 40 . do. Preferably, it further includes an external network storage (DB-O) and an internal platform management DB (DB-I).

The machine learning execution management platform system according to the present invention may be built in a machine learning server or PC as an embodiment. In addition, it may be built as a database server including a DBMS (DataBase Management System), and some of the computing resources or databases may be built in conjunction with an external third server.

In this case, the external input module 10 receives a task on which machine learning is performed, and the user accesses the platform of the present invention through the external input module 10 . The external input module 10 is implemented as an external program, such as an API (Application Programming Interface), for example, so that the user easily accesses the platform.

Each of the nodes 20 , 30 , 40 including the worker node 20 , the session node 30 , and the master node 40 is implemented in a hierarchical structure to transmit commands to each other and allocate resources.

These worker nodes 20, session nodes 30, and master nodes 40 are a kind of computing module capable of processing processes for machine learning, and access the external network storage (DB-O) to process downloaded data. can be done

The external network storage (DB-O) provides a dataset of machine learning, and the worker node 20, which has set the machine learning execution environment, downloads the dataset from the external network storage (DB-O) and run a run

Specifically, the worker node 20 performs machine learning on a task input from the user, and IDs designating each of the plurality of worker nodes 20 are recorded in the node list 31 . Accordingly, the worker node 20 is managed by the session node 30 in which the node list 31 is built.

As the node list 31 has a heap data structure, the highest priority worker node 20 with the highest current computational processing capacity among the worker nodes 20 becomes the root node, and the priority ID corresponding to the root node is deleted. It is compared to that registered in the list 32 . Through this comparison, task assignment is determined.

The session node 30 manages the worker nodes 20 composed of frameworks classified into the same or the same group. In addition, the session node 30 monitors the resources of the worker nodes 20 and transmits the command received from the master node 40 to the worker node 20 .

In particular, the session node 30 registers a node to be deleted among the worker nodes 20 recorded in the node list 31 in the deletion list 32 . In addition, the worker node 20 to which the machine learning task is assigned is determined through a process of determining whether the highest priority (root node) worker node 20 is registered in the deletion list 32 .

The master node 40 receives a request for task assignment from a user and transmits a command to the session node 30 to search for the highest priority worker node 20 to perform machine learning, and a node to be excluded from among the worker nodes 20 . (except in the execution of machine learning) is detected and notified to the session node 30 .

In the above technical configuration, the session node 30 compares the priority ID designating the highest priority worker node 20 with the deletion registration ID registered in the deletion list 32, and if the comparison result is the same, the corresponding worker node 20 exclude On the other hand, if the comparison results are different, a machine learning task is assigned.

When the priority ID and the deletion registration ID are the same, the worker node 20 searched for task assignment is randomly excluded or a node failure has occurred, so the session node 30 is the highest priority worker node 20 in the node list 31 ) so that the machine learning task is not assigned to the worker node 20 by deleting the ID.

At this time, when the node list 31 is a heap data structure, the highest priority worker node 20 is the worker node 20 designated by the root node, and after deleting the root node from the deletion list 32 through the pop operation, The node list 31 is reconstructed into a heap data structure using only the remaining nodes.

On the other hand, when the priority ID and the deletion registration ID are different, the session node 30 allocates a task so that machine learning is performed in the worker node 20 corresponding to the priority ID, and the task is assigned to the worker node 20 ), machine learning is performed according to the machine learning execution instruction.

In the above, specific embodiments of the present invention have been described above. However, the spirit and scope of the present invention is not limited to these specific embodiments, but various modifications and variations are possible within the scope that does not change the gist of the present invention. You will understand when you grow up.

Therefore, since the embodiments described above are provided to fully inform those of ordinary skill in the art to which the present invention pertains the scope of the invention, it should be understood that they are exemplary in all respects and not limiting, The invention is only defined by the scope of the claims.

10: external input module
20: worker node
30: Session node
31: node list
32: delete list
40: master node
DB-I: Platform DB
DB-O: External network storage

Claims (14)

A method for managing worker nodes in a machine learning execution management platform including a worker node (20), a session node (30) and a master node (40), the method comprising:
a deletion node registration step (S10) of registering a node to be deleted among the worker nodes 20 recorded in the node list 31 in the session node 30 in the deletion list 32;
a worker node search step (S20) of searching for the highest priority worker node (20) to perform machine learning in the session node (30) that has been requested to allocate a task from the master node (40);
a node validity determination step (S30) of comparing the priority ID indicating the highest priority worker node (20) with the deletion registration ID registered in the deletion list (32) in the session node (30);
a node list cleaning step (S40) of deleting the highest priority worker node (20) from the node list (31) in the session node (30) when the priority ID and the deletion registration ID are the same;
task assignment step (S50) of allocating a task so that machine learning is performed in the worker node 20 corresponding to the priority ID in the session node 30 when the priority ID and the deletion registration ID are different; and
A worker node management method in a machine learning execution management platform comprising a; machine learning step (S60) of performing machine learning in the worker node 20 to which the task is assigned according to a machine learning execution command. According to claim 1,
a worker node analysis step (S1) of analyzing the performance of the worker nodes 20 in the session node 30; and
Machine learning further comprising a node list generation step (S2) of generating the node list 31 composed of a heap data structure according to the performance order of the analyzed worker node 20; How to manage worker nodes in an execution management platform. 3. The method of claim 2,
The session node 30 classifies the performance of the worker node 20 into flops, which are the computational processing capabilities of the CPU and GPU, and sorts the node list 31 by using gigaflops (GFlops) as a unit. How to manage worker nodes in a machine learning execution management platform. According to claim 1,
a deletion table creation step (S3) of generating a deletion list table in which the deletion registration ID is recorded in the session node (30); and
Worker node management in the machine learning execution management platform, characterized in that it further comprises; an exclusion node analysis step (S4) of detecting the deletion target worker node 20 registered in the deletion list table in the master node 40 Way. 5. The method of claim 4,
The delete list table is a worker node management method in a machine learning execution management platform, characterized in that it is a table of a hash map structure in which mapping is made according to a key value generated by a hash function. According to claim 1,
Further comprising a node assignment request step (S20a) receiving a request from the master node 40 to assign a task (task) on which machine learning is performed to a worker node,
Machine learning execution management, characterized in that the master node 40 requests the assignment of the task to the session node 30 that manages the worker node 20 configured as a framework in which the machine learning is performed. How to manage worker nodes on the platform. 6. The method of claim 5,
In the deletion node registration step (S10),
The master node 40 extracts the worker node 20 that is randomly excluded or has a failure and requests the session node 30 to delete it,
The method of managing a worker node in a machine learning execution management platform, characterized in that the session node (30) registers the deletion registration ID in the deletion list (32) of the hashmap structure. 3. The method of claim 2,
In the worker node search step (S20),
In the machine learning execution management platform, characterized in that the session node 30 searches for the highest root node among the node list 31 consisting of the heap data structure as the highest priority worker node 20 . How to manage worker nodes. 9. The method of claim 8,
In the node validity determination step (S30),
A worker in a machine learning execution management platform, characterized in that the session node (30) compares the priority ID of the worker node (20) corresponding to the root node and the deletion registration ID registered in the deletion list (32) How to manage nodes. 9. The method of claim 8,
The node list arrangement step (S40) is,
When the priority ID and the deletion registration ID are the same, a pop operation for deleting the root node is performed on the node list 31, and then the node list 31 is reconstructed in a heap data structure method. How to manage worker nodes in a learning execution management platform. 11. The method of claim 10,
The method for managing a worker node in a machine learning execution management platform, characterized in that it further comprises a deletion list update step (S41) of excluding the deletion registration ID of the root node deleted in the pop operation from the deletion list (32). 9. The method of claim 8,
In the task assignment step (S50),
The session node (30) is a worker node management method in a machine learning execution management platform, characterized in that allocating the task to the worker node (20) corresponding to the root node. 13. The method of claim 12,
The session node 30 assigns a task to the worker node 20 corresponding to the root node,
A worker node management method in a machine learning execution management platform, characterized in that after performing a pop operation to delete the root node, the node list (31) is reconstructed in a heap data structure method. In the machine learning execution management platform system in which the worker node management method as in any one of claims 1 to 13 is performed,
an external input module 10 for receiving a task on which machine learning is performed;
a worker node 20 that performs machine learning on the input task and is registered in a plurality of node lists 31;
a session node (30) for registering a node to be deleted among the worker nodes (20) recorded in the node list (31) in the deletion list (32); and
The master node 40 receives a request for assignment of a task from a user and instructs the session node 30 to search for the highest priority worker node 20 to perform the machine learning;
The session node 30 compares the priority ID designating the highest priority worker node 20 with the deletion registration ID registered in the deletion list 32,
When the priority ID and the deletion registration ID are the same, the session node 30 deletes the ID of the highest priority worker node 20 from the node list 31,
When the priority ID and the deletion registration ID are different, the session node 30 allocates the task so that machine learning is performed in the worker node 20 corresponding to the priority ID. platform system.

Images