A task scheduling method and device based on a multi-AI acceleration card and a medium

Abstract

The application discloses a kind of based on the task scheduling method, device and medium of multiple AI acceleration card;By obtaining and according to the task to be scheduled, a plurality of subtasks are parsed, and subtask stage information is obtained;Through task node portrait method processing, the current performance characteristic portrait is generated;The uplink and downlink bandwidth size corresponding to the current PCIe bus is obtained, and the current performance characteristic portrait is combined, and the target task scheduling result is obtained by dynamic scheduling optimization method and multiple type allocation model calculation processing;According to the target task scheduling result, each host memory to acceleration card task, multiple AI acceleration card computing task and acceleration card to host memory task scheduling operation is carried out.Solve the transmission congestion caused by the inference process of the prior art multiple acceleration card multitask, unable to achieve cross-card global collaboration and cannot flexibly schedule tasks, improve the efficiency, orderliness and flexibility of task scheduling.

Images