Patent · US Active

Hybrid scheduling method for deep learning workloads, and computing apparatus with hybrid scheduling

US12056525B2 · kind B2 · utility

2Cited by
2References
21Claims
0Family size

Assignee

Inventors

Key dates

Filing dateMar 10, 2021
Grant dateAug 6, 2024
Priority date
Expiry dateMar 25, 2041

Classification

  • Technology area (CPC G)Physics
  • CPC primaryG06N20/00
  • WIPO fieldComputer technology
  • WIPO sectorElectrical engineering

Abstract

A scheduling method performed by a computing apparatus includes: generating an input vector including a resource status and metadata of each of tasks for parallel execution; determining an action for the input vector by executing an actor network based on the input vector; performing first resource scheduling for each of the tasks based on the determined action; performing second resource scheduling for each of the tasks based on the input vector; evaluating performance of first resource scheduling results of the first resource scheduling and second resource scheduling results of the second resource scheduling, for each of the tasks, using a critic network; selecting one of the first and second resource scheduling results for each of the tasks based on a result of the evaluating; and allocating resources to each of the tasks based on a resource scheduling result selected for each of the tasks.

Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.