A method and device for gpu topology partitioning

Abstract

The invention discloses a GPU topology partitioning method and device, including: determining the interconnection bandwidth between multiple GPUs according to the physical topology information of multiple GPUs, and generating a GPU topology map including multiple GPUs; GPUs are randomly divided into two partitions; the migration gain of all GPUs in the GPU topology map is calculated, and the GPU with the highest migration gain in the partition including more GPUs is migrated to the partition including fewer GPUs, and the crossover of the current partitioning scheme is calculated. The number of partition connections, and remove the migrated GPU from the GPU topology map; repeat the previous step until all GPUs in the GPU topology map are removed, and select the partition scheme with the smallest number of cross-partition connections as the partition result. The present invention can specifically optimize the topological partition of the GPU from the bottom layer according to different connection relationships between the GPUs, reduce the time-consuming transmission between the GPUs, and improve the calculation speed of artificial intelligence.

Description

technical field [0001] The present invention relates to the field of computers, and more particularly, to a method and device for GPU topology partitioning. Background technique [0002] In the fields of high-performance computing and artificial intelligence, GPUs are often used for computing acceleration. GPU is used on a large scale due to its powerful computing power and low power consumption, especially in the hot artificial intelligence field in recent years, most model training is run on GPU, which can save a lot of computing time, thus Speed ​​up model iteration. Due to the high cost of GPUs, more and more artificial intelligence developers hope to fully improve the resource utilization of GPUs and maximize the value of GPUs under limited GPU resources. However, most artificial intelligence developers lack the underlying knowledge of GPUs. The communication between GPUs in the existing technology is inefficient due to the lack of underlying optimization, and the lac...

AI Technical Summary

Problems solved by technology

However, most AI R&D personnel lack the underlying knowledge of GPU. The communication between GPUs in the existing technology is inefficient due to the lack of underlying optimization, and the lack of organization of GPU partitions slows down the calculation speed of AI.

[0003] There is no effective solution to the problem of slowing down the calculation speed of artificial intelligence caused by the lack of organization of GPU partition in the existing technology

Images