Storage Latency Evaluation via Workload Fingerprinting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional latency-based performance evaluation methods for data storage networks are inaccurate due to their inability to consider various factors such as application workload characteristics, CPU saturation, and cache misses, leading to false indications of performance degradation and failure to identify root causes or predict future latency issues.

Innovation Solution

A method and system that use machine learning to capture and classify workload metrics, creating a workload fingerprint model to identify latency contributors and predict future latency patterns by analyzing I/O operations, CPU usage, cache performance, and other relevant parameters, providing a comprehensive view of data storage network health and performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional latency-based performance evaluation methods are used, then performance degradation can be detected, but the evaluation results are highly inaccurate and prone to false indications

Engineering Contradiction:
Improveperformance evaluation accuracyVSAvoidfalse indication rate
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent segments the performance evaluation into multiple independent components: workload characterization (separating application behavior from system performance), latency measurement, and root cause analysis. By dividing the evaluation process into distinct segments that can be analyzed separately, the system avoids the false indications that arise from treating latency as a single undifferentiated metric.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces workload fingerprints as an intermediary layer between raw latency measurements and performance conclusions. These fingerprints capture application-specific I/O patterns and serve as a mediator that contextualizes latency data, allowing the system to distinguish between expected latency under certain workloads and actual performance degradation.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Difficulty of detecting and measuring

If conventional latency-based methods are used, then performance degradation can be identified, but the root cause cannot be determined

Engineering Contradiction:
Improvelatency issue detectionVSAvoidroot cause information
Core Design Contradiction:
Difficulty of detecting and measuringVSLoss of information

Solution Approach 1:

The patent adds multiple dimensions to the traditional latency measurement by incorporating workload characteristics (I/O size, I/O rate, mix of reads and writes), system state (CPU saturation, cache misses, queue depth), and temporal patterns. This dimensional expansion transforms latency from a single scalar value into a multi-dimensional profile that preserves information about the causes of latency variations.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The patent performs preliminary characterization of normal workload patterns and system behavior before attempting to identify performance issues. By establishing baseline fingerprints of application I/O patterns and system responses under various conditions, the system prepares reference data that enables rapid root cause identification when latency anomalies occur, rather than attempting to diagnose causes in real-time during incidents.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If comprehensive workload analysis is performed to improve accuracy, then the system complexity increases significantly

Engineering Contradiction:
Improveperformance evaluation accuracyVSAvoidevaluation system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent creates simplified copies or representations of complex workload patterns through workload fingerprints. Instead of analyzing every individual I/O operation and system state variable in detail, the system generates condensed fingerprint profiles that capture the essential characteristics of workload behavior. These fingerprints serve as lightweight proxies that maintain analytical accuracy while dramatically reducing computational complexity.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3598306B1Storage system latency evaluation based on I/O patterns
Publication Date: 2022.08.24 HEWLETT PACKARD ENTERPRISE DEV LP
  • EP3598306B1 patent drawingFigure 1
  • EP3598306B1 patent drawingFigure 2
  • EP3598306B1 patent drawingFigure 3

AI summary

A system or method for identifying latency contributors in a data storage network, that may include creating a historical workload fingerprint model for a data storage network from training data, along with monitoring and classifying a current sample data from the data storage network into a cluster, current workload fingerprint, and current workload type. The method may further include assigning a score to the current sample data based on the historical workload fingerprint model and correlating measured latency values from the current sample data to historically measured latency related factors to create a latency score chart that identifies factors causing latency in the data storage network for the current sample data.