A Two-Stage Interference-Aware Offline Task Scheduling Method and System for Mixed Clusters

By employing a task scheduling method that combines Gaussian mixture models with causal inference, reinforcement learning, and attention mechanisms in large-scale data centers, the interference problem between online services and offline tasks is resolved, thereby improving service quality and resource utilization.

CN117519939BActive Publication Date: 2026-05-26SHANGHAI JIAOTONG UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
SHANGHAI JIAOTONG UNIV
Filing Date
2023-11-22
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

In large-scale data centers, when online services and offline tasks are deployed in a mixed manner, there is a problem of degraded quality of online services. This is mainly due to the interference caused by offline tasks competing for CPU and memory resources. Existing technologies are difficult to effectively avoid or mitigate this interference. Especially when resource utilization is high, it is often necessary to pause or kill offline tasks to ensure the quality of online services.

Method used

We employ a task feature clustering prediction model based on Gaussian mixture model and causal inference, combined with an interference quantification model based on reinforcement learning and attention mechanism. By profiling the resources of offline tasks and quantifying interference, we select the optimal container for scheduling to avoid interference.

Benefits of technology

It improves the service quality and overall resource utilization of online services, reduces performance overhead and waste caused by interference, is suitable for distributed hybrid deployment scenarios of large-scale clusters, and has good generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117519939B_ABST
    Figure CN117519939B_ABST
Patent Text Reader

Abstract

This invention provides a two-stage interference-aware offline task scheduling method and system for hybrid clusters, comprising: establishing an offline task feature clustering prediction model based on Gaussian mixture model based on historical task resource data analysis; constructing a task feature analysis based on causal inference based on the analysis of different tasks' varying sensitivity to resources, inferring whether a task is CPU-sensitive or memory-sensitive; establishing an online interference quantification model based on reinforcement learning based on the inferred task resource features, scoring task scheduling, and selecting a batch of containers that do not violate SLAs as candidate containers; analyzing and predicting the combination relationships of different tasks, and selecting the best container from the candidate containers as the scheduling strategy. This invention is suitable for microservice frameworks with distributed hybrid deployment in large-scale clusters.
Need to check novelty before this filing date? Find Prior Art