Software Execution Time-Series Normalization for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing profiling tools for software applications on high-performance computing systems fail to accurately detect aberrant behavior due to variations in execution context, such as resource availability and data volume, leading to inefficiencies and suboptimal execution times.
Innovation Solution
A method to normalize and adjust time series data from multiple executions of the same application to account for resource and data variations, followed by calculating a metric to identify aberrant behavior, using techniques like dynamic time warping and cumulative sum distances, allowing for the detection of abnormal executions and optimization of execution parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If profiling tools are used to monitor application executions, then execution behavior can be analyzed, but false detection of aberrant behavior occurs due to context variations
Solution Approach 1:
The patent introduces a reference execution as an intermediary standard against which current executions are compared. This reference serves as a mediator that filters out normal context variations while highlighting true aberrations. The comparison mechanism uses the reference execution to normalize and contextualize performance metrics, preventing false positives caused by legitimate execution environment differences.
Solution Approach 2:
The patent transforms raw performance metrics into normalized parameters by comparing them against reference execution data. This parameter transformation process adjusts for context variations by expressing metrics as deviations from the reference rather than absolute values, thereby improving detection accuracy while reducing false alarms caused by environmental fluctuations.
2Productivity
If execution parameters are optimized based on profiling data, then application performance improves, but computation time increases due to extensive data processing
Solution Approach 1:
The patent extracts only the essential comparison metrics needed for detection and optimization, separating them from the complete profiling data set. By focusing on key performance indicators and comparing only these extracted parameters against reference values, the system achieves effective optimization with reduced computational overhead, avoiding the need to process entire data sets.
Solution Approach 2:
The patent applies partial action by performing detection and optimization on a subset of critical execution parameters rather than analyzing all possible metrics. This selective approach processes only the most influential parameters that impact performance, achieving effective optimization while minimizing computation time through focused rather than exhaustive analysis.
Data Source
Figure 1~2(c)
AI summary
A method for detecting aberrant behavior in a set of executions of one or more applications on an information processing device, implemented by a computer and comprising the following steps: - triggering (S1) said set of executions in conjunction with a profiling tool to collect, for each execution, at least one time series of measurement points associating, for each measurement point, a value with a measured parameter; - automatic formatting (S2) of the time series obtained for said set, by adjusting, for each time series, its length, its values, and its number of measurement points; - calculation (S3) of a metric between two time series from among the time series collected for said set of executions; - detection (S4) of an aberration from said distance