NVMe Cache Scheduling With Workload Intensity Fairness

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing NVMe SSDs suffer from interference and unfairness between concurrent workloads due to simplistic threshold-based prioritization techniques that fail to account for the degree of workload intensity, leading to significant slowdowns and potential security vulnerabilities.

Innovation Solution

Implementing a machine learning-based prediction mechanism using Venn predictors to intelligently categorize workloads into priority levels based on their request generation rates, ensuring well-calibrated probabilistic predictions to reduce interference and achieve fairness by scheduling requests accordingly.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If threshold-based prioritization techniques are used in NVMe SSDs, then device throughput is improved, but fairness between concurrent workloads deteriorates due to interference at different levels within the SSD

Engineering Contradiction:

Improvedevice throughputVSAvoidfairness between workloads

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent changes the parameter of workload prioritization from simple threshold-based classification to intensity-based dynamic prioritization. By monitoring request generation rates and calculating workload intensity metrics, the system dynamically adjusts priority levels based on actual workload characteristics rather than fixed thresholds, thereby improving fairness while maintaining throughput

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent replaces the mechanical threshold-based prioritization mechanism with a machine learning-based prediction mechanism. Venn predictors are used to intelligently categorize workloads into priority levels based on their request generation rates, substituting the rigid mechanical threshold system with an adaptive intelligent system that reduces interference and achieves fairness

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Productivity

If multiple submission queues are supported in NVMe SSDs, then device utilization is improved, but interference between concurrent workloads increases leading to unfairness

Engineering Contradiction:

Improvedevice utilizationVSAvoidinterference between workloads

Core Design Contradiction:

ProductivityVSObject-generated harmful factors

Solution Approach 1:

The patent introduces an intermediary workload classification and prioritization mechanism between the multiple submission queues and the SSD processing layers. This intermediary system categorizes workloads by intensity and manages queue scheduling to minimize interference between concurrent workloads, allowing multiple queues to operate efficiently without causing unfairness

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies different prioritization strategies to different workloads based on their local characteristics (request generation rates, intensity levels). By tailoring the prioritization treatment to each workload's specific properties rather than applying a uniform approach across all queues, the system reduces interference while maintaining high utilization

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS12535968B2Workload prioritization fairness techniques in non-volatile memory express (NVMe) systems

Publication Date: 2026.01.27 DELL PROD LP

AI summary

Systems and methods for implementing a workload fairness policy at the data cache level within a non-volatile memory express solid-state drive utilizes a prioritization technique that predicts a degree of intensity for workloads based on a request generation rate. The degree of intensity is used to prioritize and rank workloads, thereby, avoiding interference between requests from each other.