Parallel Task Scheduling via Hardware Execution Cost Analysis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current deep learning model processing technologies face inefficiencies in parallel processing due to inadequate consideration of hardware execution costs across nodes in clusters, leading to suboptimal execution times and resource utilization.

Innovation Solution

A method that determines parallel computing graphs and operator scheduling schemes based on hardware execution costs, using a combination of actual measurement and predictive modeling to optimize task execution across clusters, thereby improving parallel efficiency and adapting to different environmental conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If parallel processing is adopted to shorten training and applying time, then processing speed is improved, but hardware execution cost increases and resource utilization becomes suboptimal

Engineering Contradiction:
Improveprocessing speedVSAvoidresource utilization efficiency
Core Design Contradiction:
SpeedVSProductivity

Solution Approach 1:

The system dynamically changes scheduling parameters based on hardware execution costs. Different operator tasks are assigned to different nodes according to their specific hardware cost characteristics, enabling optimal resource utilization while maintaining high processing speed through parallel execution.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent applies local quality by determining hardware execution costs for different operator tasks on different nodes individually. Each node's characteristics are evaluated separately, and tasks are scheduled to specific nodes based on their local hardware cost profiles, rather than using a uniform scheduling approach across the entire cluster.

Inventive Principle:
Principle #3Local quality

2Productivity

If hardware execution cost is considered for optimizing task scheduling, then resource utilization efficiency is improved, but system complexity increases

Engineering Contradiction:
Improveresource utilization efficiencyVSAvoidscheduling system complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system performs preliminary action by pre-determining hardware execution costs for different operator tasks on different nodes before actual task scheduling. This advance calculation of execution costs creates a reference basis that simplifies the subsequent scheduling process, as tasks can be directly assigned based on pre-computed cost information rather than requiring complex real-time optimization.

Inventive Principle:
Principle #10Preliminary action

3Loss of time

If parallel computing graphs are determined based on hardware execution cost, then execution time is reduced, but measurement and calculation difficulty increases

Engineering Contradiction:
Improveexecution timeVSAvoidhardware execution cost measurement difficulty
Core Design Contradiction:
Loss of timeVSDifficulty of detecting and measuring

Solution Approach 1:

The patent employs a practical measurement approach where hardware execution costs are determined through actual execution measurements on the target hardware platform. Rather than using complex theoretical models, the system performs direct measurements of execution costs for different operator tasks on different nodes, creating simple, actionable data that directly guides scheduling decisions.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Data Source

PatentEP3866008B1Method for processing tasks in parallel, device and storage medium
Publication Date: 2024.12.11 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP3866008B1 patent drawingFigure 1~2A
  • EP3866008B1 patent drawingFigure 2B~3
  • EP3866008B1 patent drawingFigure 4A

AI summary

Embodiments of the present disclosure disclose a method for processing tasks in parallel, a device and a storage medium, and relate to a field of artificial intelligent technologies. The method includes: determining at least one parallel computing graph of a target task; determining a parallel computing graph and an operator scheduling scheme based on a hardware execution cost of each operator task of each of the at least one parallel computing graph in a cluster, in which the cluster includes a plurality of nodes for executing the plurality of operator tasks, and each parallel computing graph corresponds to at least one operator scheduling scheme; and scheduling and executing the plurality of operator tasks of the determined parallel computing graph in the cluster based on the determined parallel computing graph and the determined operator scheduling scheme.