Return to Index
Operations Research Models and Methods
Computation Section
Queuing Add-in
 - Non-Markovian Queues

The add-in computes approximate values for the steady-state measures when the arrival or service processes are not required to have exponential distributions.

To add a non-Markovian Queuing model on the worksheet, place the cursor at a cell where the model is to be described and select Add NonMarkov from the menu. The dialog box below allows entry of the parameters of the model. This model does not require the interarrival and service distributions to have exponential distributions. Rather, the model requires the specification of the coefficients of variation (COV) of the arrival and service processes. The COV for arrivals is the standard deviation for time between arrivals divided by the mean time between arrivals . For an exponential distribution the COV is 1. Distributions with less variability than the exponential have COV < 1, while distributions with more variability have COV > 1. The service process COV is the standard deviation of the service time divided by the mean service time.


Example: An order picking process


An order picking process in a warehouse gets calls for service at an average rate of 8.5 per hour. The average time to fill the order is 0.1 hours. For analysis purposes assume both times are exponentially distributed. Analyzing the system as an M/M/1 queue, the average time in the queue is 0.5667 hours. An opportunity arises to reduce the variability of the process for filling orders. The inventory manager wonders if the change is worth the cost.

To analyze this problem we select the NonMarkov option from the menu. The dialog below defines the parameters of the model. The name, arrival rate, service rate and number of channels are entered in the appropriate boxes. The approximation does not apply to finite queues or finite populations, so these options are not present. If the replication entry is greater than 1, the number entered determines the number of Queuing models put on the worksheet. We illustrate the presentation of three replications below.


The check boxes on the dialog determine optional presentation of the steady state results. Checking the Show Titles box causes titles to be added to the worksheet. When defining a series of Queuing models in sequential cells, it is useful to show the titles for the first model, and then skip the titles for the remaining models.

The set of results for the Non-Markovian case is smaller than those available for Poisson queues. This is partially due to the restriction against finite queues and finite populations, making some results not relevant. The approximations used do not allow the computation of state probabilities.

The results for three replications of the Queuing model are shown below. When the models were created on the worksheet all had the default parameters shown for PP1_1. To illustrate the effect of reducing the variability in the service times, we set the COV to 0.5 in the second model and 0 in the third (no variability). In general the numbers and time in the queue decrease as variability decreases. Note that with COV of 1, the distribution is actually Markovian (Poisson process). Thus we see an M in the type designation when the COV is 1.

It is clear from the results that reducing variability in the service process causes decreased time in the queue. The mean time in service does not change because that is fixed by the data and unaffected by the variability.



Using Reduced Variability to Obtain more Throughput

We wonder whether the benefits of the reduced variability could be used to obtain more throughput, rather than reduced queues. That is, we want to increase the arrival rate in the second two systems to obtain a time in the queue equal to 0.567. This is indeed possible as shown in the analysis below. Note that for the systems in columns I and J, the arrival rates have been increased substantially while the time in the queue remains at 0.567.

This analysis was performed using the Excel Solver, called with the Solver command on the Tools menu of Excel. The Solver dialog shown below is set up to obtain the solution in column I above. The target cell, I13, holds the mean time in the queue. The changing cell, the cell I3, holds the arrival rate. We have asked Solver to find the value of the arrival rate that gives a mean queue time of 0.5667. When performing the analysis it was necessary to give an initial value larger than the ultimate solution of 9.006. With a smaller initial value, Solver tried arrival rates greater than 10 and the procedure stopped because the system is unstable for arrival rates greater than 10.


Reducing Variability of Both Arrival and Service Processes

  Reducing the variability of both the interarrival times and service times further reduces the time in the queue as shown below. The results of column N show that by cutting the COV to 0.5 reduces the queue time to 1/4 of its value with Poisson processes. Reducing the COV's for both processes to 1/10 the original value reduces the queue time to 1/100 of its former value.
  Again we investigate the possibility of increased throughput while keeping queue time as a constant. We used Solver as previously described to obtain the results below. The arrival rate for the third case is at almost the theoretical maximum of 10.
  When both COV are equal to 1 the system can be analyzed with no approximation using the Poisson queuing formulas. The formulas used by the Non-Markovian analysis are also accurate for M/G/1 systems. For all other systems, the results are approximate.


Return to Top

tree roots

Operations Research Models and Methods
by Paul A. Jensen
Copyright 2004 - All rights reserved

Next Page