KR102300118B1 - Job placement method for gpu application based on machine learning and device for method - Google Patents

Abstract

In the present invention, when information about an application that requires execution is received, the step of checking the properties of the application and the state of the system based on the information about the received application, selecting the input data of the learning model from among the properties of the application, and learning Inputting to the model, applying the learning model to determine the task sequence of the application and the co-location structure of the application and other applications, and executing the application on the GPU according to the determined task order and co-location structure It provides a machine learning-based job placement method for GPU applications.

Description

JOB PLACEMENT METHOD FOR GPU APPLICATION BASED ON MACHINE LEARNING AND DEVICE FOR METHOD

The present invention provides a machine learning-based task placement method for GPU applications.

A GPU (Graphic Processing Unit) virtualization system can be used as a General-Purpose GPU (GPGPU) function by performing parallel operations with multiple processing cores, and is a system provided to users by a data center or cloud company.

In the GPU virtualization system, the service provider provides the GPU in a way that provides dedicated access rights to the user who requests the GPU. can be executed However, when the kernel of several applications is executed as a single process, the resource usage characteristics of the executed applications are diverse, and problems such as degradation of the overall system performance may occur due to the joint execution of several applications. Research on this is in progress.

The present invention aims to improve the performance and resource utilization of the entire system when the GPU executes multiple applications jointly by providing a machine learning-based task placement method for GPU applications.

In the present invention, when information about an application that requires execution is received, the step of checking the properties of the application and the state of the system based on the information about the received application, selecting the input data of the learning model from among the properties of the application, and learning Inputting to the model, applying the learning model to determine the task sequence of the application and the co-location structure of the application and other applications, and executing the application on the GPU according to the determined task order and co-location structure It provides a machine learning-based job placement method for GPU applications.

According to an embodiment, the application properties input to the learning model include at least one of GPU application dictionary information, GPU application profiling information, and cluster environment information, and the GPU application profiling information includes application's GPU utilization, GPU memory usage, It may include at least one of GPU core usage, PCIe throughput, and execution time.

According to an embodiment, at least one of GPU utilization, GPU memory usage, GPU core usage, and PCIe throughput included in the GPU application profiling information may be information collected by periodically monitoring the GPU for a predetermined time.

According to an embodiment, the cluster environment information may include at least one of a node name, a GPU card, a GPU structure, a GPU memory, the number of GPU cores, and a PCIe bandwidth.

According to an embodiment, the learning model observes a current state and determines an action according to a policy, wherein the decision is determined such that an expected discount cumulative reward when performing the action is maximized, wherein the current state of the learning model is It is determined based on the application properties input to the learning model, and the action may be related to the work slot of the co-located structure.

According to an embodiment, the number of actions may be equal to the number of work slots of the co-located structure.

According to an embodiment, the learning model may use gradient descent to reduce the deviation of the distribution of the discount cumulative reward function.

According to an embodiment, the gradient descent method may be a method of predicting a gradient descent value after subtracting an accumulated empirical discount accumulated reward obtained according to a policy from an expected accumulated discount reward.

According to one embodiment, the determining of the co-location structure includes the steps of adding the excess work slots to a back log when a work slot exceeding the number of available work slots is required, and the number of available work slots. The method may include filling the empty job slots with jobs added to the backlog when there are empty job slots because a smaller number of job slots is required.

According to an embodiment, the learning model may be a DQN learning model.

According to an embodiment, after executing the application on the GPU, the method may further include monitoring an execution result of the application and modifying an input of the learning model according to the monitoring result.

The present invention provides a computer-readable storage medium storing a computer program in which the above-described method is performed when the computer program is executed by a processor.

According to an embodiment disclosed in the present invention, since tasks are allocated to the shared resources of the GPU based on the application profiling data, the resource utilization rate of the entire GPU system can be improved without changing according to the type of application.

In addition, according to an embodiment disclosed in the present invention, it is possible to reduce the interference between each application, it is possible to reduce the decrease in the work speed.

1 is a diagram for explaining the relationship between performance and shared resources during joint execution of applications.
2 is a flowchart of a machine learning-based task arrangement method for GPU application according to an embodiment of the present invention.
3 is a diagram for explaining the structure of a learning model according to an embodiment of the present invention.
4 is a diagram for explaining an algorithm related to a learning model according to another embodiment of the present invention.
5 is a diagram for explaining a result of resource allocation according to time when an application is jointly executed according to the prior art.
6 is a diagram for explaining a resource allocation result according to time when applications are jointly executed according to an embodiment of the present invention.
7 is a diagram for explaining the structure of a work arrangement service model according to an embodiment of the present invention.
8 to 10 are diagrams for explaining experimental results for each workload of a machine learning-based job arrangement method for GPU application according to an embodiment of the present invention.
11 is a view for explaining a result of an experiment for analysis of job speed degradation for a machine learning-based job arrangement method for GPU application according to an embodiment of the present invention.
12 is a diagram for explaining a training overhead comparison experiment result for a machine learning-based task arrangement method for GPU application according to an embodiment of the present invention.
13 is a diagram for explaining a price comparison between a machine learning-based job arrangement device for GPU application and a conventional device according to an embodiment of the present invention.

Since the present invention can have various changes and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and it should be understood to include all modifications, equivalents and substitutes included in the spirit and scope of the present invention. In describing each figure, like reference numerals have been used for like elements.

Terms such as first, second, A, and B may be used to describe various elements, but the elements should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component. and/or includes a combination of a plurality of related listed items or any of a plurality of related listed items.

When a component is referred to as being “connected” or “connected” to another component, it is understood that the other component may be directly connected or connected to the other component, but other components may exist in between. it should be On the other hand, when it is said that a certain element is "directly connected" or "directly connected" to another element, it should be understood that no other element is present in the middle.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as “comprise” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It should be understood that this does not preclude the existence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Unless defined otherwise, all terms used herein, including technical and scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram for explaining the relationship between performance and shared resources during joint execution of applications.

In the conventional machine learning-based MPS, when the target resource is the GPU, the target application is mainly limited to the machine learning application. there was. In addition, since detailed resources of the application are not taken into account, interference between shared resources occurs, thereby deteriorating the performance of the entire system.

1 shows when two of the five applications (Lammps, Gromacs, CNN, vgg16, googlenet) share GPU resources, GPU, GPU memory, SM (core), load on input/output bandwidth (I/O bandwidth) It is a drawing for explaining the degree. Among the five applications in FIG. 1, Lammps and Gromacs are HPC applications, and CNN, vgg16, and googlenet are machine learning applications.

Specifically, FIG. 1(a) is a diagram showing the degree of application execution time when two applications share GPU resources, FIG. 1(b) is a diagram showing the load degree of GPU memory, FIG. 1 (c) is a diagram showing the degree of load on the GPU core, and (d) of FIG. 1 is a diagram showing the degree of load on the input/output bandwidth. Each result of (a) to (d) of FIG. 1 indicates that the darker the color, the greater the load degree of each item.

1 (a) to (d), when the CNN application and the Gromacs application share GPU resources, the execution time and the load level of the GPU core are 3, but the load level of the GPU memory is 1, and , it can be seen that the load degree of the input/output bandwidth is very severe at 5. Also, if the googlenet application and the vgg16 application share GPU resources, it takes about 4 execution time, but the GPU memory load is 3, the GPU core load is 2, and the I/O bandwidth load is 1. can

Based on the results of FIG. 1 , it can be seen that the execution time becomes longer when the load of the GPU memory increases, but in the case of other factors, it is difficult to see that the increase in the load of each element affects the execution time. This is because interference that may occur is different depending on the characteristics of each shared application, and it is difficult to predict system performance due to such interference.

2 is a flowchart of a machine learning-based task arrangement method for GPU application according to an embodiment of the present invention.

In step 210, when information on an application to be executed is received, the properties of the application and the state of the system may be checked based on the received information on the application.

Meanwhile, in the machine learning-based task arrangement method for GPU application according to an embodiment of the present invention, data input to the learning model may include at least one of the application properties of Table 1. In other words, the application properties input to the learning model may include at least one of GPU application dictionary information, GPU application profiling information, and cluster environment information.

category application properties GPU Application Preliminary Information application name Application type input file size input variable About GPU Application Profiling GPU Utilization GPU memory usage PCIe throughput execution time Cluster Environment Information node name GPU card GPU architecture GPU memory number of GPU cores PCIe bandwidth

Among them, the GPU application dictionary information may include at least one of an application name, an application type, an input file size, and an input variable, and the GPU application profiling information includes an application's GPU utilization, GPU memory usage, GPU core usage, PCIe throughput, and It may include at least one of the execution time. In addition, the cluster environment information may include at least one of a node name, a GPU card, a GPU structure, a GPU memory, the number of GPU cores, and a PCIe bandwidth. On the other hand, Table 1 is an example of the items of the resource use history of the application collected for the use of the learning model, it is apparent to those skilled in the art that other information can be input to the learning model.

In the machine learning-based job placement method for GPU application according to an embodiment of the present invention, the application attribute may be information submitted by a user to the system to perform the application, or may be information collected while the application is executed. , the method by which this application attribute is obtained is not limited thereto. In addition, the usage of each item included in the GPU application profiling information of Table 1 may be periodically monitored and collected for a certain period of time. For example, at least one of GPU utilization, GPU memory usage, GPU core usage, and PCIe throughput included in the GPU application profiling information may be information collected by periodically monitoring the GPU for a predetermined time.

In step 220, input data of the learning model may be selected from the properties of the application and input to the learning model. Here, the learning model may be a DQN learning model.

In step 230, the learning model may be applied to determine which tasks are collocated with the task sequence of the application.

Here, the learning model observes the current state and determines the action according to the policy, wherein the decision is determined such that the expected discount cumulative reward when performing the action is maximized, and the current state of the learning model is input to the learning model It is determined based on the application attribute, and the action may be related to the work slot of the co-located structure. Also, the number of actions may be equal to the number of work slots of the co-located structure.

Also, the learning model may use gradient descent to reduce the deviation of the distribution of the discount cumulative reward function. In this case, the gradient descent method may be to predict the gradient descent value after subtracting the accumulated empirical discount accumulated rewards obtained according to the policy from the expected discount accumulated rewards.

In step 240, according to the determined task order, it may be co-located and executed on the GPU.

On the other hand, in step 240, if a work slot exceeding the number of available work slots is required, adding the excess work slots to a back log and a number of work slots smaller than the number of available work slots are required and filling the empty job slots with jobs added to the backlog when there is an empty job slot.

In addition, the method according to an embodiment of the present invention may further include, after executing the application on the GPU, monitoring the execution result of the application and modifying the input of the learning model according to the monitoring result.

Therefore, the machine learning-based task placement method for GPU applications according to an embodiment of the present invention predicts the interference between applications sharing GPU resources based on the profiling data of the application and performs task placement accordingly, so the system has the effect of improving the overall performance of

Hereinafter, the method according to an embodiment of the present invention will be described on the assumption that the DQN learning model is used.

3 is a diagram for explaining the structure of a learning model according to an embodiment of the present invention.

3 is a diagram showing the overall DQN structure in which an agent interacts with the environment and repeats learning, reward, and observation. Referring to FIG. 3 , at time t, the agent observes the current state ( s _t ) and is requested to perform a specific action a. After performing a specific action, the state transitions to the next state ( s _t+1 ) and the agent receives a reward ( r _t ). At this time, state transition and reward contents have Markov properties because they are affected only by the environment state and the action performed by the agent. The agent iterates through the learning process and aims to maximize the expected Cumulative Discounted Reward.

Here, the action is selected based on the agent's policy, where the policy may be expressed as in Equation 1.

[Equation 1]

로 나타낼 수 있다. 즉, DQN 학습 모델의 최종 목표는 정책(policy)

에 따라 상태 s에서 행동 a를 수행할 때 예상되는 할인 누적 보상을 최대화하는 것이며, 이때 예상되는 할인 누적 보상은

로 표현할 수 있다.Here, π(s,a) represents the probability that action a will be executed in state s. In the method according to an embodiment of the present invention, instead of storing a large number of {state, action} values in a table, a policy parameter θ is calculated and obtained using a deep neural network (DNN), which is a function approximator. Therefore, the policy is

can be expressed as In other words, the final goal of the DQN learning model is policy

is to maximize the expected discounted cumulative reward when performing action a in state s according to

can be expressed as

Meanwhile, the method according to an embodiment of the present invention may use the properties of the application defined in Table 1 above as an input of the DQN learning model for collaboration arrangement. At this time, the learning model input state s=(j,R) includes a working vector j and a cluster resource vector R. Through the resource environment history item of the cluster, the cluster resource can be expressed as in Equation 2 below, and each element of R means the amount of available resources of each GPU shared resource.

[Equation 2]

On the other hand, task j represents profiling history information of resources collected for each predetermined time slot (T) for each task, and can be expressed as in Equation 3.

[Equation 3]

은 수학식 4와 같이 정의할 수 있다.Accordingly, j _i may include information on resource usage that changes for each _{time slot (T j} ) of the corresponding application (APP), and is included in _{j i}

can be defined as in Equation (4).

[Equation 4]

The machine learning-based task arrangement method for GPU application according to an embodiment of the present invention may use task information and the value of the GPU cluster environment as input values of the DNN, and in order to input changing resource usage, the input state space is can be modified.

4 is a diagram for explaining an algorithm related to a learning model according to another embodiment of the present invention.

The first part 410 of FIG. 4 means a part in which, when the current state is s _i , an action a _i is selected, and a reward _i and a policy q _i to be received by the agent are determined accordingly. And the second part 420 is a part for modifying the policy variable θ in order to reduce the distribution deviation of the discount cumulative compensation function Q value.

Referring to the first part 410 , it is requested to observe the current state s _t and perform a specific action a. After performing a specific action, the state transitions to the next state ( s _t+1 ) and the agent receives a reward ( r _t ). Also, referring to the first region 415 of the first portion 410 , the method according to an embodiment of the present invention performs action a in state s to maximize the discount cumulative reward function Q value.

에 따라 상태 s에서 행동 a를 수행할 때 예상되는 할인 누적 보상

을 최대화한다. 참고로 강화형 기계 학습은 Q값의 분포의 편차를 줄이기 위해 경사 하강법을 사용하는데, 종래의 경사 하강법의 경우, 정책에 따라 얻은 경험적 할인 누적 보상만을 사용한다.A method according to an embodiment of the present invention includes a policy

The expected discount cumulative reward when performing action a in state s according to

to maximize For reference, the reinforcement-type machine learning uses gradient descent to reduce the variance in the distribution of Q values. In the case of the conventional gradient descent method, only the empirical discount cumulative reward obtained according to the policy is used.

에서 정책에 따라 얻은 경험적 할인 누적 보상

를 차감하여 경사 하강 값을 예측한다. However, in the method according to an embodiment of the present invention, as shown in the second region 425 of the second part 420 , a prediction is performed for a policy parameter having a lower distribution than in the prior art. Discount Accumulated Reward

Accumulated rewards from empirical discounts earned in accordance with the policy

The gradient descent value is predicted by subtracting

5 is a diagram for explaining a result of resource allocation according to time when an application is jointly executed according to the prior art.

Figure 5 shows the result of performing the GPU shared resource arrangement using the DeepRM learning model, it can be seen that the resources considered in the DeepRM learning model are CPU and memory. On the other hand, in the case of the DeepRM learning model, since a virtual task is created, task placement is not performed for real applications. In addition, as described above, it is characterized in that only a certain amount of resource usage is considered during the execution time for the CPU and memory.

6 is a diagram for explaining a resource allocation result according to time when applications are jointly executed according to an embodiment of the present invention.

6 is a diagram composed of a cluster image for four resources (GPU, GPU memory, GPU core, and PCIe) and three job images (job slot 1 to job slot 3). And, the determinant at the bottom of FIG. 6 indicates the amount of resource used by each task for each time slot T, and can be expressed as {T, (R1, R2, R3, R4)}. For example, the purple job (job slot 1) of FIG. 6 is {0, (1, 1, 0, 1)}, {1, (2, 3, 2, 1)}, { 2, (3, 0, 2, 1)}, {3, (0, 0, 3, 1)}, {4, (0, 0, 0, 1)}. Also, the blue job (job slot 2) is {0, (2, 2, 1, 3)}, {1, (1, 1, 1, 2)}, {2, (1, 1) for 5 time slots. , 0, 2)}, {3, (1, 2, 0, 0)}, {4, (0, 0, 0, 0)}. Finally, the green job (job slot 3) is {0, (1, 1, 2, 0)}, {1, (1, 0, 1, 0)}, {2, (0, 1, 1, 0)}, {3, (3, 0, 0, 2)}, {4, (3, 4, 0, 2)}.

On the other hand, since the input of the neural network is preferably expressed in a fixed state, as many as M of the number of work slots can be used as the input. The result of learning is used to determine the next action a, and the action space is expressed as a={0,1,...,M}. When a=1, it means to execute the task assigned to the first task slot, a=0 means to hold the action, and indicates that there is no space for the task image in the current time slot T in the cluster image. Accordingly, an appropriate task a is determined and executed while proceeding according to the time slot T, and the resource cluster image is updated accordingly.

If more than M job slots are required, they can be added to the back log component of the state space, and when empty job slots exist, the jobs that were added to the backlog are empty jobs. fill the slot. However, the above method cannot co-locate and execute tasks because only one task can be executed according to result a.

Meanwhile, in the method according to an embodiment of the present invention, resources may be allocated in consideration of the number of simultaneously usable work slots of one resource. For example, when the use of one resource requires a total of 4 work slots, the present invention can allocate resources to as many work slots as possible within a range where one resource does not exceed 4 resources per one time slot. Therefore, in the first time slot (T=0) of FIG. 6, the number of resources allocated to each work slot can be expressed as {4, 4, 3, 4}, which means that in the first time slot, the GPU core has three resources. is allocated, GPU, GPU memory and PCIe means that 4 resources are allocated.

Meanwhile, referring to FIG. 6 compared to FIG. 5, the method according to an embodiment of the present invention shares resources while reducing performance degradation compared to the conventional DeepRM of FIG. 5 because there are more variables to consider when allocating shared resources. The advantage is that it is possible to In addition, according to the conventional DeepRM, a virtual task is performed and only a single task arrangement is possible, but according to an embodiment of the present invention, an actual task can be performed and a joint arrangement can be performed. That is, according to an embodiment of the present invention, it is possible to reduce the execution time and increase resource utilization through co-location.

7 is a diagram for explaining the structure of a work arrangement service model according to an embodiment of the present invention.

Referring to (a) of Figure 7, the infrastructure layer of the method according to an embodiment of the present invention is a target GPU computing container (computing container), the orchestration platform (orchestration platform) is the premise that using kubernetes. The proposed service model refers to a service model reflecting the method according to an embodiment of the present invention described based on FIGS. 2 to 6 . The task arrangement service can actually arrange and execute the action values selected according to the characteristics of each application and the system environment, that is, the tasks to be co-located, and monitor the status of the system and tasks. Accordingly, it is possible to provide a service that detects failure, termination, etc. of a task and arranges the following tasks to a resource.

7B is an algorithm showing the overall flow of arranging tasks to be jointly executed while reflecting the profiling data of the application submitted by the user and the status of cluster resources based on reinforcement machine learning.

Referring to (b) of FIG. 7 , first, the algorithm submits an application to be executed by a user, and checks profiling data of the submitted application and cluster status of the system. Thereafter, input data to be used as an input of the learning model is determined. _{In this case, j i} of Equation 4 and R of Equation 2 may be input as profiling data of the application and state information of the system.

After that, reinforcement machine learning is applied to determine the order of tasks in the entire workload queue and the tasks to be co-located in that order. After co-locating and executing the tasks on the GPU according to the order of the tasks causing the interference, the state of the tasks is monitored. The GPU co-location and execution and monitoring operations described above are repeatedly performed until the entire job is deployed.

8 to 10 are diagrams for explaining experimental results for each workload of a machine learning-based job arrangement method for GPU application according to an embodiment of the present invention.

A total of 5 workloads using 4 HPC applications and 7 Tensorflow benchmark ML (machine learning) applications provided by NGC in Table 2 below for performance comparison between the method according to an embodiment of the present invention and the prior art Performance analysis was performed for (GPU heavy, GPU light, GPU memory heavy, GPU memory light, random).

HPC ML Lammps googlenet Resnet50 Gromacs alexnet Resnet101 Qmcpack Vgg16 Resnet152 Hoomd Vgg11

In addition, as a prior art that is compared with the method according to an embodiment of the present invention, a total of four work arrangement methods in Table 3 were utilized.

Conventional work placement method SJF single batch Place tasks in increasing order of execution time Tetris multiple batches Arranged in a packing method according to the application's work usage and resource availability DeepRM Max Placement by applying reinforcement learning with the maximum value of resource usage of the task as input DeepRM Mean Placement by applying reinforcement learning with the average value of resource usage of the task as input

The GPU container cluster utilized for the experiment is a node cluster using Kubernetes, a container orchestration platform, and k8s, a kubernetes default device plugin, was modified for multi-task deployment. In addition, the tasks were executed using the NVIDIA docker container, and the specifications of the computing node used in the experiment are shown in Table 4 below.

Node Details CPU (Master) GPU (node) architecture Intel® Core ^™ i7-5820K Nvidia GeForce Titan Xp D5x core clock 3.30GHz 1.58GHz number of cores 6 3840 memory size 32 GB 12 GB Threading API - Nvidia CUDA 10.0 PCIe bandwidth - 32 GB/s operating system Ubuntu 16.04.6 LTS Ubuntu 16.04.6 LTS

In the case of Figure 8, the prior art for GPU heavy (a) and GPU light (b), which are workloads classified based on average GPU utilization 40%, and the method according to an embodiment of the present invention (execution time), It is a diagram showing GPU utilization and GPU memory utilization.

The GPU heavy workload of Figure 8 (a) was composed of vgg11, vgg16, resnet50, resnet101, resnet152, qmcpack, hoomd, and lammps applications showing an average GPU utilization of 40% or more. When deployed in the SJF method, it has the shortest execution time of 4807 seconds, but the average GPU utilization is 70.50% and the average GPU memory is 36.44%, which is the lowest. Among the methods for executing multiple tasks, the method (Dqn GPU) according to an embodiment of the present invention shows the highest performance at 4891 seconds, and shows high resource utilization with GPU utilization of 82.09% and 59.13%. In the case of DeepRM Mean, the execution time is the longest at 5530 seconds, because when the GPU memory is used to the maximum, a memory problem (OOM: Out Of Memory) occurred and the failed application was re-executed.

The GPU light workload of FIG. 8 (b) was composed of alexnet and gromacs applications showing an average GPU utilization of less than 40%. When applications with low GPU utilization are executed, the SJF method has the lowest performance and resource utilization at 2295 seconds. This seems to have increased the execution time compared to other techniques that can be jointly executed by single-deploying applications in the SJF method. In the case of application gromacs, the execution time is longer than that of alexnet, but it uses less GPU memory, because up to three tasks can be performed at one time. The method (Dqn GPU) according to an embodiment of the present invention has the best performance at 1214 seconds, GPU utilization 89.31%, and GPU memory utilization 88.42% .

In the case of Figure 9, GPU memory heavy (a) and GPU memory light (b), which are workloads classified based on an average of 50% of GPU memory utilization, the prior art and the execution time of the method according to an embodiment of the present invention, GPU It is a diagram showing utilization and GPU memory utilization.

The GPU memory heavy workload of Fig. 9 (a) was composed of vgg11, vgg16, resnet50, resnet101, resnet152 applications showing an average of 50% or more GPU memory utilization. In the case of SJF, the execution time is the shortest, followed by the method (DqnGPU) according to an embodiment of the present invention. The reason that the execution time of the method according to an embodiment of the present invention is increased is likely to have caused resource competition when the executed applications have high GPU utilization and are co-located. However, it can be seen that the shortest execution time of 1214 seconds and the highest resource utilization of 89.31% and 88.42% when compared with other joint execution arrangement techniques.

The GPU memory light workload of FIG. 9 (b) was composed of alexnet, googlenet, qmcpack, hoomd, and gromacs applications showing an average GPU memory utilization of less than 50%. It can be seen that the method (DqnGPU) according to an embodiment of the present invention is the most excellent in terms of execution time. It has the shortest execution time at 3928 seconds, and at this time, based on the learning results, it is because gromacs and qmcpack applications with low GPU memory utilization but long execution time were co-located and performed.

In the case of FIG. 10, it is a view showing the execution time, GPU utilization, and GPU memory utilization of a method according to an embodiment of the present invention and the prior art for an arbitrarily configured workload.

The random workload of FIG. 10 was composed of all applications used in the experiment. When co-location is performed by the method (DqnGPU) according to an embodiment of the present invention, the performance of the task execution time is 2481 seconds, the GPU utilization is 62.16%, and the GPU memory utilization is 20.45%. In terms of execution time, it shows better performance than other co-execution batch techniques. However, GPU utilization is lower than that of Tetris and DeepRM Max techniques when compared to other workload experiments, and GPU memory utilization is similar when compared to SJF. Referring to FIG. 10 , it can be seen that in the case of the method according to an embodiment of the present invention, since various resource use characteristics exist when several applications exist, the influence of resource competition can be reflected in various ways when performing task assignment.

11 is a diagram for explaining a result of an experiment for analysis of job speed degradation for a machine learning-based job arrangement method for GPU application according to an embodiment of the present invention.

For the analysis of work speed degradation, the speed degradation of a total of 30 jobs for 4 jointly performed batch techniques included in the random workload of FIG. 10 was analyzed. Based on the execution time of a single executed application, the evaluation was divided into no speed degradation (error 2%), speed degradation less than 10%, 10% or more and less than 20%, and 20% or more.

Referring to FIG. 11 and Table 5, it shows the slowdown of each operation for the four techniques when running a random workload. In the method according to an embodiment of the present invention, there was no slowdown in a total of 20 tasks. Because it learns according to the usage of resources that change according to execution time, it seems that the decrease in speed caused by resource competition is small by placing it within a range that does not exceed the actual resources. However, the slowdown that occurs means that, in addition to resources not taken into account when performing jointly, the characteristics of the running application, various conditions or environments may be affected. In the case of Tetris and DeepRM Max techniques, there was no slowdown in 14 tasks. In the case of DeepRM Mean, a total of 12 is the most common at 10% or more of a decrease in speed. In the case of DeepRM Mean, because it is arranged according to the average resource value, the OOM problem that occurs during joint execution caused a lot of slowdown in operation.

no degradation Less than 10% degradation 10% or more degradation 20% or more degradation Tetris 14 10 4 2 DqnGPU 20 5 3 2 DeepRM Max 14 6 7 3 DeepRM Mean 13 5 8 4

12 is a diagram for explaining a training overhead comparison experiment result for a machine learning-based task arrangement method for GPU application according to an embodiment of the present invention.

12 is a diagram illustrating a result of comparing overhead defined as a ratio of training time to total execution time using a random workload. The total execution time was compared by normalizing it based on the SJF with the shortest execution time. Assuming that the resource usage history of the entire application exists, the total execution time including the training time for the initial task assignment is compared. In the case of the method (DqnGPU), DeepRM Max, and DeepRM Mean according to an embodiment of the present invention, the training time is included because the arrangement is based on the results obtained through reinforcement machine learning.

Referring to FIG. 12 , the results of training overhead of each batch technique are shown. The execution time normalization values are about 1.294, 1.081, 1.303, and 1.346 for Tetris, a method according to an embodiment of the present invention (DqnGPU), DeepRM Max, and DeepRM Mean, respectively. In the method according to an embodiment of the present invention, the training time is 123 seconds, which is about 0.054, and the execution time minus the training time is 1.027. When comparing the results of SJF and execution time, it can be seen that the performance difference is small. This means that there is almost no difference in performance when placing the next task after excluding the first training time of the task set. Not only does the resource utilization increase, but if you continue to train online while performing the task, the accuracy will increase, which means that the performance of the next task set can be improved. The training time of DeepRM Max is 0.048, and when only the execution time of Tetris and DeepRM Max is compared, it is 1.294 and 1.254, which is better in the case of DeepRM Max. Therefore, it can be seen that the task placement technique applying reinforcement machine learning is meaningful except for the initial training overhead.

13 is a view for explaining a price comparison between a machine learning-based job arrangement device for GPU application and a conventional device according to an embodiment of the present invention.

13 is a comparison of the price of the SJF and the device according to an embodiment of the present invention based on the GPU instance provided by the cloud company.

Table 6 shows the GPU instance prices provided by Amazon EC2. G3 instatnce offers NVIDIA Tesla M60, G4 instatnce offers NVIDIA T4 GPU cards. If it is assumed that g4dn.4xlarge is equal to two g3.xlarge in terms of GPU memory resources, up to $0.602 can be saved when g4dn.4xlarge is shared (hereafter referred to as cost 1). Also, assuming that g4dn.4xlarge is equal to 4 g4dn.xlarge in terms of vCPU and memory resources, it is possible to reduce the price by up to $0.301 when using g4dn.4xlarge (hereafter referred to as cost 2).

GPU vCPUs Memory (GB) GPU memory (GB) GPU Core Cost (USD) G3.xlarge One 4 30.5 8 2048 0.75 g4dn.xlarge One 4 16 16 2560 0.526 g4dn.4xlarge One 16 64 16 2560 1.204

In addition, the graph of FIG. 13 shows a comparison result according to the cost assumed above based on the execution of the previous random workload. In the case of cost 1, when a shared instance (DqnGPU) is used according to an embodiment of the present invention, it is possible to reduce costs by about 29.86% compared to when a single instance is used. In the case of cost 2, it can be seen that the cost is reduced by about 49.98% when an instance (DqnGPU) that is shared according to an embodiment of the present invention is used rather than when a single instance is used.

The above description is merely illustrative of the technical idea of the present invention, and a person of ordinary skill in the art to which the present invention pertains may make various modifications and variations without departing from the essential characteristics of the present invention. Accordingly, the embodiments implemented in the present invention are not intended to limit the technical spirit of the present invention, but to explain, and the scope of the technical spirit of the present invention is not limited by these embodiments. The protection scope of the present invention should be construed by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

Claims (12)

In a machine learning-based task placement method for GPU applications performed by a machine learning-based task placement device for GPU applications,
when information on an application for which execution is requested is received, checking properties of the application and a state of the system based on the received information on the application;
selecting input data of a learning model from among the properties of the application and inputting it into the learning model;
determining a work order of the application and a co-location structure of the application and other applications by applying the learning model; and
executing the application on the GPU according to the determined task sequence and co-location structure;
wherein the learning model observes a current state and determines an action according to a policy, wherein the determination is determined such that an expected discounted cumulative reward when performing the action is maximized,
A machine learning-based job placement method for GPU applications. According to claim 1,
The application properties input to the learning model include at least one of GPU application dictionary information, GPU application profiling information, and cluster environment information, and the GPU application profiling information includes GPU utilization, GPU memory usage, and GPU core usage of the application. , PCIe throughput and execution time. 3. The method of claim 2,
At least one of the GPU utilization, the GPU memory usage, the GPU core usage, and the PCIe throughput included in the GPU application profiling information is information collected by periodically monitoring the GPU for a certain period of time, GPU application A machine learning-based task placement method for 3. The method of claim 2,
The cluster environment information is
A machine learning-based job placement method for a GPU application, comprising at least one of a node name, GPU card, GPU structure, GPU memory, number of GPU cores and PCIe bandwidth. According to claim 1,
The current state of the learning model is determined based on the application property input to the learning model, and the action relates to a task slot of the co-placement structure. 6. The method of claim 5,
The machine learning-based task placement method for GPU application, characterized in that the number of actions is equal to the number of task slots of the co-location structure. 6. The method of claim 5,
The learning model is a machine learning-based task arrangement method for GPU applications, characterized in that it uses gradient descent to reduce the deviation of the distribution of the discount cumulative reward function. 8. The method of claim 7,
The gradient descent method is
A machine learning-based task arrangement method for GPU application, characterized in that the gradient descent value is predicted after subtracting the empirical discount accumulation reward obtained according to the policy from the expected discount accumulation reward. In a machine learning-based task placement method for GPU applications performed by a machine learning-based task placement device for GPU applications,
when information on an application for which execution is requested is received, checking properties of the application and a state of the system based on the received information on the application;
selecting input data of a learning model from among the properties of the application and inputting it into the learning model;
determining a work order of the application and a co-location structure of the application and other applications by applying the learning model; and
executing the application on the GPU according to the determined task sequence and co-location structure;
Determining the co-location structure comprises:
If you need a job slot that exceeds the number of available job slots,
adding excess work slots to a back log; and
filling the empty job slots with jobs added to the backlog when there are empty job slots because a number of job slots smaller than the number of available job slots is requested;
A machine learning-based job placement method for GPU applications, comprising: According to claim 1,
The learning model is a machine learning-based task placement method for GPU applications, characterized in that the DQN learning model. According to claim 1,
after executing the application on the GPU, monitoring the execution result of the application; and
modifying the input of the learning model according to the monitoring result;
Machine learning-based job placement method for GPU applications, characterized in that it further comprises. A computer-readable storage medium storing a computer program in which the method according to any one of claims 1 to 11 is performed when the computer program is executed by a processor.

Images