GPU cluster deep learning task parallelization method, device and electronic equipment

Abstract

The GPU cluster deep learning task parallelization method, device, and electronic equipment provided in the embodiments of the present application relate to the field of Internet technology. By analyzing the similarity between the pending deep learning task and each computing node of the GPU cluster, it is determined that the pending deep learning task is The target computing node in the GPU cluster to reduce the possibility of computing node resource contention, thereby improving the utilization rate and execution efficiency of deep learning task system resources, and then according to the number of GPUs required by the deep learning task to be processed, the The processing deep learning task is divided into multiple target subtasks, and the interference level and communication cost of the target subtasks are analyzed, so as to determine the target GPU of the target subtask in the target computing node, so as to avoid unbalanced resource allocation on the GPU in the computing node, It achieves a high degree of parallelization of deep learning tasks, improves the resource utilization of GPU clusters, and improves the execution efficiency of deep learning tasks.

Description

technical field [0001] The present application relates to the field of Internet technology, in particular to a parallelization method, device and electronic equipment for GPU cluster deep learning tasks. Background technique [0002] With the deepening of deep learning research, deep learning technology has achieved fruitful results in computer vision, speech recognition, text processing and other fields, bringing great convenience to people's lives. However, the complex structure of the neural network model and the huge amount of data put forward higher requirements for computing power. GPU (Graphic Processing Unit, Image Processor) cluster integrates multiple GPU computing resources, provides powerful and efficient parallel computing capabilities for computing-intensive deep learning tasks, and effectively solves the computing needs of multiple deep learning tasks. [0003] However, when a deep learning task runs on a resource-sharing GPU cloud platform, its execution eff...

AI Technical Summary

Problems solved by technology

[0005] Although this method realizes the parallelization of deep learning tasks to a certain extent, this method mainly considers the usage of resources and does not consider the physical characteristics of resources and the characteristics of the task itself. It cannot achieve efficient parallelization of deep learning tasks, which will reduce the Execution Efficiency of Deep Learning Workloads

At the same time, this method does not support the fine-grained multi-task allocation of the GPU, and cannot make full use of the GPU resources on the node, which will affect the efficient execution of deep learning tasks, reduce the GPU utilization of the node, and thus affect the resource utilization of the GPU cluster.

Images