A hybrid load scheduling optimization method for deep learning of heterogeneous GPU clusters

Abstract

A hybrid load scheduling optimization method for deep learning of heterogeneous GPU clusters, comprising: statically adding node type labels to multiple lower-layer computing nodes of a heterogeneous GPU cluster; classifying applications for upper-layer applications of a distributed cluster; For multiple applications, multiple different types of lower-level computing nodes are evenly distributed to multiple applications through the scheduling module; calculate the time required for multiple different types of lower-level computing nodes to run on multiple applications; time, the performance differences of heterogeneous GPU applications are found; and the performance differences of heterogeneous GPU applications are traded by the second price trading method. In the present invention, in the heterogeneous GPU cluster, the scheduling optimization model method is more excellent than the traditional distributed processing framework in processing the mixed load of deep learning, especially when the cluster environment is complex and the isomerization is serious, the cluster can be fully utilized resources and significantly improve the execution efficiency of the system.

Description

Technical field: [0001] The invention relates to the technical field of GPU clusters, in particular to a hybrid load scheduling optimization method for deep learning of heterogeneous GPU clusters. Background technique: [0002] With the development of information technology and the gradual expansion of the cluster scale, the upper-layer applications of distributed clusters are gradually becoming more complex, such as common network search, voice assistant, etc. These applications are obtained through deep learning task training. The lower-level nodes of a distributed cluster are composed of a large number of GPU servers that provide computing resources for deep learning training tasks. However, with the continuous optimization and updating of GPU servers, the lower-level nodes also gradually show the characteristics of heterogeneity. Therefore, how to allocate reasonable and efficient computing resources for deep learning mixed workloads in heterogeneous GPU clusters has bec...

AI Technical Summary

Problems solved by technology

This allocation method does not consider the heterogeneous characteristics between nodes, and cannot take advantage of the performance advantages of heterogeneous computing nodes. In a heterogeneous environment, it has poor performance in processing mixed loads of deep learning.

Images