Parallel Efficiency Calculation for Heterogeneous Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for evaluating parallel computer system performance are limited by the need for load balance and are not applicable to heterogeneous systems, making it difficult to quantify parallel efficiency and identify performance impediments, especially in grid or cluster environments where processor performance varies.
Innovation Solution
A method for calculating parallel efficiency that includes calculating load balance contribution, virtual parallelization, and performance impediment factor contributions, allowing for quantitative evaluation and optimization of parallel processing systems without requiring uniform processor performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional parallel efficiency calculation methods are used, then performance evaluation can be performed under ideal load balance conditions, but the method cannot be applied to heterogeneous systems with varying processor performances
Solution Approach 1:
The patent applies local quality by introducing processor-specific performance weights (wi) that reflect individual processor capabilities within the heterogeneous system. Instead of treating all processors uniformly, each processor's contribution to parallel efficiency calculation is adjusted according to its local performance characteristics, enabling accurate evaluation across diverse hardware configurations while maintaining measurement precision.
Solution Approach 2:
The patent changes the evaluation parameters from uniform processing time measurements to weighted performance metrics. By transforming the efficiency calculation to use processor-specific weights and normalized performance indicators, the method adapts to heterogeneous systems while preserving the ability to accurately measure parallel efficiency through modified parameter relationships.
2Measurement precision
If actual measurement of processing times is performed, then quantitative parallel efficiency can be calculated, but the method requires identical processor performances which is not satisfied in grid or cluster environments
Solution Approach 1:
The patent introduces performance weights (wi) as intermediary variables that mediate between actual processor performance variations and the efficiency calculation framework. These weights act as a translating layer that converts heterogeneous processor capabilities into a unified measurement scale, enabling quantitative parallel efficiency calculation across grid and cluster systems with diverse hardware configurations.
Solution Approach 2:
The patent transforms the measurement parameters from direct processing time comparisons to weighted performance ratios. By changing the calculation parameters to include processor-specific normalization factors and performance weights, the method maintains quantitative precision while becoming compatible with heterogeneous grid and cluster environments where processor performances naturally vary.
3Reliability
If scalability evaluation with multiple measurements is performed, then qualitative parallel performance can be assessed, but the evaluation process becomes complex and is not made in general
Solution Approach 1:
The patent segments the parallel efficiency calculation into distinct components: processor-specific performance measurements, weight assignments, and aggregated efficiency computation. This segmentation allows the complex evaluation to be broken down into manageable steps that can be executed systematically, reducing overall process complexity while maintaining reliable quality assessment through structured decomposition of the evaluation methodology.
Data Source
AI summary
This parallel efficiency calculation method can be applied, even in a case where a load balance is not kept, to many parallel processing including a heterogeneous computer system environment, and quantitatively correlates a parallel efficiency with a load balance contribution ratio and a virtual parallelization ratio, as parallel performance evaluation indexes, and parallel performance impediment factor contribution ratios. A parallel efficiency Ep(p) is calculated by using a load balance contribution ratio Rb(p) representing a load balance degree between respective processors included in a parallel computer system, a virtual parallelization ratio Rp(p) representing a ratio, with respect to time, of a portion calculated in parallel by the respective processors to processing executed in the parallel computer system, and a parallel performance impediment factor contribution ratio Rj(p) representing a ratio of a processing time of a portion of each parallel performance impediment factor to a total processing time of all the processors.