ML-Based VM Rescheduling for Cloud Workload Stability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Cloud service providers face challenges in balancing the needs of diverse workloads by evicting interruptible workloads to allocate resources to higher-priority ones, leading to frequent interruptions and prolonged downtime due to unnoticed notifications and immediate re-eviction of redeployed virtual machines.
Innovation Solution
A machine-learning model predicts computing resource availability to adaptively schedule the resumption of evicted virtual machines, reducing the likelihood of immediate re-eviction by deferring redeployment until resource availability is stable.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If VMs hosting interruptible workloads are evicted and immediately redeployed, then resource allocation responsiveness is improved, but the VMs are likely to be evicted again shortly after resumption
Solution Approach 1:
The system performs preliminary actions by sending notifications to customers before evicting VMs, allowing them to prepare for the eviction and potentially mitigate its impact. This advance warning mechanism helps customers save state or prepare for resumption, reducing the negative effects of eviction while maintaining resource allocation flexibility.
Solution Approach 2:
The system dynamically adjusts VM scheduling based on priority levels and resource availability. High-priority VMs are protected from eviction while low-priority interruptible VMs can be evicted when resources are scarce. The scheduler continuously adapts to changing conditions, balancing resource allocation responsiveness with VM execution stability through dynamic priority-based management.
2Productivity
If cloud service providers allocate additional computing resources to high priority workloads by evicting lower-priority VMs, then resource allocation efficiency is improved, but customer service quality deteriorates due to unnoticed notifications
Solution Approach 1:
The system implements feedback mechanisms by sending notifications to customers when their VMs are evicted. This feedback loop ensures customers are aware of the eviction and can take appropriate actions, such as saving state or preparing for resumption. The notification system provides continuous feedback about VM status changes, improving service continuity while maintaining efficient resource allocation.
Solution Approach 2:
The system enables customers to self-manage their interruptible workloads by providing them with notifications and control over their VM configurations. Customers can identify which workloads are interruptible and configure them accordingly, allowing the system to automatically evict and redeploy these VMs without manual intervention, thus improving both resource allocation efficiency and service continuity.
3Loss of information
If notifications of VM interruption are sent to customers, then information transparency is improved, but service disruption increases due to manual redeployment delays
Solution Approach 1:
The system implements automatic redeployment of evicted VMs without requiring manual customer intervention. When a VM is evicted, the system automatically schedules and executes its redeployment, maintaining full information transparency through notifications while eliminating service disruption delays. This automation allows the system to simultaneously provide complete information about VM status changes and minimize service interruption time.
Data Source
AI summary
Systems and methods are provided for scheduling a virtual machine (VM) to host a workload in a cloud system. In particular, the disclosed technology schedules an evicted VM for redeploying an interruptible workload. The scheduling is based on capacity prediction and inference data associated with a type of the evicted VM. Capacity signal predictor generates training data for training a machine learning model using capacity signal history data of the cloud system. The machine-learning model, once trained, predicts capacity including a rate of evictions for the types of the evicted VM. The predicted data is based on at least the current status of available computing resources. Upon receiving a notice associated with a workload interruption, the intelligent scheduler prioritizes the evicted VM for scheduling and determines whether to defer redeploying the evicted VM based on the rate of eviction for the type of the evicted VM.