Parallel algorithms for fitting Markov arrival processes

Mindaugas Bražėnas, Gábor Horváth, M. Telek

Research output: Contribution to journalArticle

1 Citation (Scopus)

Abstract

The fitting of Markov arrival processes (MAPs) with the expectation–maximization (EM) algorithm is a computationally demanding task. There are attempts in the literature to reduce the computational complexity by introducing special MAP structures instead of the general representation. Another possibility to improve the efficiency of MAP fitting is to reformulate the inherently serial classical EM algorithm to exploit modern, massively parallel hardware architectures. In this paper we present three different EM-based fitting procedures that can take advantage of the parallel hardware (like Graphics Processing Units, GPUs) and apply a special MAP structure, the Erlang distributed-continuous-time hidden Markov chain (ER-CHMM) structure for reducing the computational complexity. All the proposed parallel algorithms have their strengths: the first one traverses the samples only once per iteration, the second one is memory efficient (far more than the classical serial algorithm), and the third one has exceptionally low execution times. These procedures are compared with the standard serial forward–backward procedure for performance comparison. The new algorithms are orders of magnitudes faster than the standard serial procedure, while (depending on the variant) using less memory.

Original languageEnglish
Pages (from-to)50-67
Number of pages18
JournalPerformance Evaluation
Volume123-124
DOIs
Publication statusPublished - Jul 1 2018

Fingerprint

Parallel algorithms
Parallel Algorithms
Expectation-maximization Algorithm
Computational complexity
Computational Complexity
Hidden Markov Chain
Hardware
Data storage equipment
Continuous-time Markov Chain
Expectation Maximization
Hardware Architecture
Parallel Architectures
Performance Comparison
Graphics Processing Unit
Markov processes
Execution Time
Iteration
Standards

Keywords

  • EM algorithm
  • GPU
  • Markov arrival process
  • Parallel computation
  • Traffic model fitting

ASJC Scopus subject areas

  • Software
  • Modelling and Simulation
  • Hardware and Architecture
  • Computer Networks and Communications

Cite this

Parallel algorithms for fitting Markov arrival processes. / Bražėnas, Mindaugas; Horváth, Gábor; Telek, M.

In: Performance Evaluation, Vol. 123-124, 01.07.2018, p. 50-67.

Research output: Contribution to journalArticle

Bražėnas, Mindaugas ; Horváth, Gábor ; Telek, M. / Parallel algorithms for fitting Markov arrival processes. In: Performance Evaluation. 2018 ; Vol. 123-124. pp. 50-67.
@article{806df83525534e728ab41457ad3bcb80,
title = "Parallel algorithms for fitting Markov arrival processes",
abstract = "The fitting of Markov arrival processes (MAPs) with the expectation–maximization (EM) algorithm is a computationally demanding task. There are attempts in the literature to reduce the computational complexity by introducing special MAP structures instead of the general representation. Another possibility to improve the efficiency of MAP fitting is to reformulate the inherently serial classical EM algorithm to exploit modern, massively parallel hardware architectures. In this paper we present three different EM-based fitting procedures that can take advantage of the parallel hardware (like Graphics Processing Units, GPUs) and apply a special MAP structure, the Erlang distributed-continuous-time hidden Markov chain (ER-CHMM) structure for reducing the computational complexity. All the proposed parallel algorithms have their strengths: the first one traverses the samples only once per iteration, the second one is memory efficient (far more than the classical serial algorithm), and the third one has exceptionally low execution times. These procedures are compared with the standard serial forward–backward procedure for performance comparison. The new algorithms are orders of magnitudes faster than the standard serial procedure, while (depending on the variant) using less memory.",
keywords = "EM algorithm, GPU, Markov arrival process, Parallel computation, Traffic model fitting",
author = "Mindaugas Bražėnas and G{\'a}bor Horv{\'a}th and M. Telek",
year = "2018",
month = "7",
day = "1",
doi = "10.1016/j.peva.2018.05.001",
language = "English",
volume = "123-124",
pages = "50--67",
journal = "Performance Evaluation",
issn = "0166-5316",
publisher = "Elsevier",

}

TY - JOUR

T1 - Parallel algorithms for fitting Markov arrival processes

AU - Bražėnas, Mindaugas

AU - Horváth, Gábor

AU - Telek, M.

PY - 2018/7/1

Y1 - 2018/7/1

N2 - The fitting of Markov arrival processes (MAPs) with the expectation–maximization (EM) algorithm is a computationally demanding task. There are attempts in the literature to reduce the computational complexity by introducing special MAP structures instead of the general representation. Another possibility to improve the efficiency of MAP fitting is to reformulate the inherently serial classical EM algorithm to exploit modern, massively parallel hardware architectures. In this paper we present three different EM-based fitting procedures that can take advantage of the parallel hardware (like Graphics Processing Units, GPUs) and apply a special MAP structure, the Erlang distributed-continuous-time hidden Markov chain (ER-CHMM) structure for reducing the computational complexity. All the proposed parallel algorithms have their strengths: the first one traverses the samples only once per iteration, the second one is memory efficient (far more than the classical serial algorithm), and the third one has exceptionally low execution times. These procedures are compared with the standard serial forward–backward procedure for performance comparison. The new algorithms are orders of magnitudes faster than the standard serial procedure, while (depending on the variant) using less memory.

AB - The fitting of Markov arrival processes (MAPs) with the expectation–maximization (EM) algorithm is a computationally demanding task. There are attempts in the literature to reduce the computational complexity by introducing special MAP structures instead of the general representation. Another possibility to improve the efficiency of MAP fitting is to reformulate the inherently serial classical EM algorithm to exploit modern, massively parallel hardware architectures. In this paper we present three different EM-based fitting procedures that can take advantage of the parallel hardware (like Graphics Processing Units, GPUs) and apply a special MAP structure, the Erlang distributed-continuous-time hidden Markov chain (ER-CHMM) structure for reducing the computational complexity. All the proposed parallel algorithms have their strengths: the first one traverses the samples only once per iteration, the second one is memory efficient (far more than the classical serial algorithm), and the third one has exceptionally low execution times. These procedures are compared with the standard serial forward–backward procedure for performance comparison. The new algorithms are orders of magnitudes faster than the standard serial procedure, while (depending on the variant) using less memory.

KW - EM algorithm

KW - GPU

KW - Markov arrival process

KW - Parallel computation

KW - Traffic model fitting

UR - http://www.scopus.com/inward/record.url?scp=85047798849&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85047798849&partnerID=8YFLogxK

U2 - 10.1016/j.peva.2018.05.001

DO - 10.1016/j.peva.2018.05.001

M3 - Article

VL - 123-124

SP - 50

EP - 67

JO - Performance Evaluation

JF - Performance Evaluation

SN - 0166-5316

ER -