2nd International ICST Conference on Performance Evaluation Methodologies and Tools

Research Article

USING PARALLEL REPLICATIONS FOR SEQUENTIAL ESTIMATION OF MULTIPLE STEADY STATE QUANTILES

Download230 downloads
  • @INPROCEEDINGS{10.4108/valuetools.2007.1936,
        author={Mirko Eickhoff and Don McNickle and Krzysztof Pawlikows},
        title={USING PARALLEL REPLICATIONS FOR SEQUENTIAL ESTIMATION OF MULTIPLE STEADY STATE QUANTILES},
        proceedings={2nd International ICST Conference on Performance Evaluation Methodologies and Tools},
        proceedings_a={VALUETOOLS},
        year={2010},
        month={5},
        keywords={Discrete-event simulation steady state quantiles sequential estimation},
        doi={10.4108/valuetools.2007.1936}
    }
    
  • Mirko Eickhoff
    Don McNickle
    Krzysztof Pawlikows
    Year: 2010
    USING PARALLEL REPLICATIONS FOR SEQUENTIAL ESTIMATION OF MULTIPLE STEADY STATE QUANTILES
    VALUETOOLS
    ICST
    DOI: 10.4108/valuetools.2007.1936
Mirko Eickhoff1,*, Don McNickle2,*, Krzysztof Pawlikows2,*
  • 1: Computer Science University of Canterbury Christchurch, New Zealand
  • 2: Management University of Canterbury Christchurch, New Zealand
*Contact email: m.eickhoff@cosc.canterbury.ac.nz, don.mcnickle@canterbury.ac.nz, krys.pawlikowski@canterbury.ac.nz

Abstract

Simulation output data analysis in performance evaluation studies of complex stochastic systems such as the Internet is typically limited to mean values, even though it provides very limited information about the analysed system's performance. Quantile analysis is not as common, even though it can provide much deeper insights into the system of interest. A set of quantiles can be used to approximate a cumulative distribution function, providing full information about a given performance characteristic of the simulated system. In this paper, we will present two new methods for estimating steady state quantiles and distribution functions. The quantiles are estimated using simulation output data from concurrently executed independent replications. They are calculated sequentially and on-line to guarantee that their final statistical errors do not exceed a permitted threshold.