The development of a discrete-event simulation tool, called GMSim, based on the generalized semi-Markov process (GSMP) formalism is described. The GSMP representation comprises both analysis and simulation in a unied framework. This paper focuses on the simulation aspect and how to deal with a combinatorially exploding state space. A compositional GSMP modeling methodology is proposed, which in turn combined with an object-oriented programming approach. A key feature of the resulting tool is the close resemblance with the underlying mathematical structure. This facilitates coherent modeling and also an efficient implementation. The tool is completely generic and extendible by Tcl script programming. Application specific components are developed by C++ programming in combination with M4 macro processing.

This paper compares unbuffered wormhole-switched networks and buffered packet-switched networks with respect to estimation of expected packet latency. Simulation experiments show that the variance of the natural sample-mean estimator is significantly larger in the wormhole-switched case. An increased estimator variance is an evidence of a harder estimation problem. In particular, it leads to larger confidence intervals. The increased variance is attributed to the in-line blocking property of wormhole-switched networks. This property results from lack of smoothing buffers at the switching points. The overall effect is to increase both the variance of individual latency recordings and also the amount of serial correlation in the sequence of latency samples. The performance of two frequently used variance estimators is assessed in this paper. Estimating variance by assuming independent samples fails for unbuffered networks. A batched-means method do work provided the ba...

This paper discusses variance reduction for estimation of mean packet latency in wormhole-switched networks. Our interest is motivated by the fact that the variance of the natural sample-mean estimator is much larger for unbuffered systems than for traditional buffered packetswitched networks. One way to achieve variance reduction is to modify the natural estimator in terms of a control variate. The effect of two such strategies are studied in this paper. Simulation experiments are used to quantify the effect. The first approach uses an internal control variate built from latency inter-sample times. The other strategy is to employ a M /M /s queuing model which is simulated in parallel. This model represents an approximation of the true network, hence subsequent queuing times can be used to construct an external control variate. Our conclusion is that a variance reduction of about 20% can be obtained. On the balance we recommend the method using an internal control var...

In this paper we consider how to estimate the variance of a mean network latency estimator. The estimator variance is needed for computation of confidence intervals. We focus on the case where latency samples are highly correlated and argue that a spectral-based estimation method is appropriate. A spectral representation of the latency samples shows that the requested variance is related to the value at frequency zero. We try to estimate the required variance from a least-square fit of spectral data. Polynomial splines are used as target functions for the fit. They have sufficient flexibility to adapt to various spectral shapes. The adaptability is inherent and does not require user-specified parameters. Hence, the method is well suited in a sequential procedure. The success of the approach depends on putting constraints on the spline-fit in order to control the flexibility. This is based on the anticipated spectral shape. In particular, we assume that the spectrum of a latency time-...

In this paper we compare the average performance of one class of low-discrepancy quasi-Monte Carlo sequences for global optimization. Weiner measure is assumed as the probability prior on all optimized functions. We show how to construct van der Corput sequences and we prove their consistency. Numerical experimentation shows that the van der Corput sequence in base 2 has a better average performance.