HPC Autotuning Using Median Filtering Under Noisy Performance Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing autotuning methods for high-performance computing (HPC) systems are inefficient in noisy environments due to stochastic interference, leading to suboptimal performance and prolonged optimization times, especially when dealing with Cauchy noise and resource-intensive resampling.

Innovation Solution

An autotuning method for HPC systems using the median of performance evaluations instead of the mean, combined with a filtering threshold and a decreasing confidence interval, to filter out unsuitable parametrizations and optimize resource usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If resampling is performed to reduce noise impact, then measurement precision is improved, but resource consumption increases

Engineering Contradiction:
Improveperformance measurement accuracyVSAvoidcomputational resource consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent changes the statistical parameter from mean to median for performance evaluation. This parameter change makes the evaluation robust to Cauchy noise without requiring multiple resampling evaluations, thus improving measurement precision while avoiding increased resource consumption. The median inherently filters out extreme values that would otherwise require repeated sampling to eliminate.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If optimization iterations are increased to achieve better tuning, then productivity is improved, but loss of time increases

Engineering Contradiction:
Improveoptimization effectivenessVSAvoidoptimization time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent performs preliminary filtering of parametrizations using the median-based evaluation before conducting full optimization iterations. By discarding unpromising parametrizations early through the filtering threshold mechanism, the system avoids wasting time on ineffective iterations, thus improving optimization effectiveness while reducing total optimization time.

Inventive Principle:
Principle #10Preliminary action

3Use of energy by moving object

If filtering threshold is lowered to discard more parametrizations, then resource consumption is reduced, but measurement precision deteriorates

Engineering Contradiction:
Improvecomputational resource consumptionVSAvoidperformance evaluation accuracy
Core Design Contradiction:
Use of energy by moving objectVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the filtering threshold is dynamically adjusted based on the distribution of median performance values. The system continuously monitors performance evaluations and adapts the threshold to maintain optimal filtering precision, ensuring that resource consumption is reduced without sacrificing measurement accuracy. This feedback loop allows the system to automatically calibrate the filtering stringency.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3982271B1Method for autotuning noisy HPC systems
Publication Date: 2026.03.04 BULL SA
  • EP3982271B1 patent drawingFigure 1~2
  • EP3982271B1 patent drawingFigure 3~4
  • EP3982271B1 patent drawingFigure 5~6

AI summary

The invention relates to a method for autotuning a high-performance computing system comprising the steps of: - If the performance has not been evaluated at least twice with the current value of the set of parameters: ∘ Evaluating the performance of the HPC system with the current value of the set of parameters, - If the performance has been evaluated at least twice: ∘ Computing the median of the evaluated performances, ∘ If the median does not verify a rule based on a filtering threshold, discard the current value of the set of parameters and restart the method with at least one other value generated by the optimization module, ∘ If the median verifies the rule, performing a resampling method based on the median of the evaluated performances and on a confidence interval decreasing with the number of steps of the optimization method.