Systems and Means of Informatics

2024, Volume 34, Issue 3, pp 109-122

MODELING OF THE INPUT FLOW OF LANL MUSTANG COMPUTING CLUSTER WORKLOADS

  • M. P. Krivenko

Abstract

Statistical analysis is an indispensable element in the construction of a mathematical model of the object under study. Queuing systems as an object of research have specific features that make it necessary to go beyond the general theory of stochastic processes. The article discusses the construction of the models of the input flow of multiprocessor systems based on the trace of the real workload of the Mustang cluster obtained as a part of the Atlas project (www.project-atlas.org). Mustang data features include a long observation period, an impressive amount of data collected, a wide field of research due to the simplified nature of previous studies and fuzzy conclusions already made, the combination of fragments with different flow intensities, the presence of stationary and nonstationary areas, and the inapplicability of the simple Poisson flow model. As a solution to the problems that arise for stationary data fragments, it is proposed to use the branching Poisson process model. The well-known methods of estimating the model parameters are supplemented by the procedure for refining estimates and formalized methods of confirmatory analysis. Given the large amounts of data being processed, it is important to build effective algorithms for calculating the characteristics of streams and smoothing out sample indicators.

[+] References (6)

[+] About this article