Task Queue Scheduling for Cloud Deep Learning SLA Prioritization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current parallel allocation solutions for deep learning models in cloud computing services struggle to efficiently schedule tasks across multiple users with different service-level allocation (SLA) levels, leading to difficulties in load balancing and resource allocation.
Innovation Solution
A task processing method that creates a to-be-scheduled task queue based on user priorities and SLA levels, allowing tasks to be scheduled in the same round while allocating computing resources effectively, ensuring that tasks from users with higher SLA levels are prioritized and that idle resources are balanced across special-purpose processing units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If parallel allocation solution schedules tasks based on resource requirements and available resources, then task parallelism is achieved, but load balancing across multiple users with different SLA levels deteriorates
Solution Approach 1:
The patent segments the task queue into multiple priority levels corresponding to different SLA levels. Tasks are divided into first task sets (higher priority) and second task sets (lower priority), allowing differential scheduling strategies for each segment while maintaining overall system parallelism
Solution Approach 2:
The patent applies local quality by implementing different scheduling policies for different SLA levels. High SLA tasks receive preferential treatment with higher scheduling priority, while low SLA tasks are scheduled when resources are available, creating localized optimization for each user tier
2Adaptability or versatility
If service-level allocation (SLA) levels are implemented for multiple users, then user priority differentiation is achieved, but resource allocation efficiency deteriorates
Solution Approach 1:
The patent implements dynamic scheduling where the scheduling priority of tasks changes based on SLA levels and current resource availability. The system dynamically adjusts task execution order without sacrificing overall resource utilization efficiency
Solution Approach 2:
The patent changes the scheduling parameter (priority) based on SLA levels. By introducing SLA-based priority as a scheduling parameter, the system achieves user differentiation while maintaining efficient resource allocation through parameter-driven scheduling decisions
3Productivity
If tasks from multiple users are scheduled in the same round, then scheduling efficiency is improved, but task dependence relationships deteriorate
Solution Approach 1:
The patent segments tasks into user-specific task sets while maintaining round-based scheduling. Each user's tasks are processed as a coherent unit within the round, preserving internal dependencies while enabling parallel scheduling across users
Solution Approach 2:
The patent performs preliminary organization of tasks into task sets before scheduling. By pre-grouping tasks from the same user and establishing their internal dependency relationships, the system can schedule multiple user task sets in parallel while maintaining reliability of individual task dependencies
Data Source
AI summary
A task scheduling method comprises the steps of: in response to the reception of a request for processing a plurality of task sets, creating a current to-be-scheduled task queue in a task processing system based on priorities of the plurality of task sets and tasks in the plurality of task sets, where a plurality of to-be-scheduled tasks in the current to-be-scheduled task queue are scheduled in the same round of scheduling; allocating computing resources used for scheduling the plurality of to-be-scheduled tasks; and enabling the plurality of to-be-scheduled tasks in the current to-be-scheduled task queue to be scheduled by using the computing resources. In this manner, a plurality of tasks with different priorities and quotas can be scheduled according to SLA levels of users, and the efficiency and flexibility of parallel services of cloud computing deep learning models are improved by using a run-time load-balancing scheduling solution.