next up previous
Next: Conclusion Up: New Delay Analysis in Previous: New Delay Analysis Algorithm

Evaluation

 

In a suite of simulation experiments we compared the proposed new method for end-to-end delay analysis with that of two commonly used methods ([8, 9]), which we call Algorithm  Decomposed and Algorithm Service Function. These methods were originally proposed by Cruz and adopted in various forms by many others. We compare their performance on a network with FIFO servers arranged in a feedforward topology. These experiments show that our new method generally computes tighter bounds on end-to-end delays than the other approaches.

Experiments

We evaluate the performance of the new approach by comparing Algorithm Integrated to the delay computation methods described by Cruz in [8, 9], which we call Algorithm  Cruz. In this section, we first define the performance metric and then describe the system configuration considered. The performance results will be presented and discussed in the next section.

Topology and Traffic Descriptions.

In our evaluation, we consider a simple tandem network with n tex2html_wrap_inline2181 switches, which are connected in a chain.

  

Figure 3: A Tandem Network.

An example of such a tandem network with 5 switches is illustrated in Figure 3. There are 2n+1 connections in this network. Connection 0 is the longest; it enters the network at the middle input port of the first switch and exits the network from the middle output port of the n-th switch. For k=0 to n-1, the (2k+1)-th session enters the network by the upper input port of the k-th switch and exits the network from the upper output port of the (k+1)-th switch, the (2k+1)-th session enters the network by the lower input port of the k-th switch and exits the network from the upper lower output port of the (k+2)-th switch. The middle output port of each switch, excepted the first one, carries four connections, including Connection 0.

In order to simplify the evaluation, we assume that every source traffic is controlled by a token bucket with a unit bucket size ( tex2html_wrap_inline2203 ) and the token arrival rate tex2html_wrap_inline2205 , where U is the work load of the network. While an increase of the traffic burstiness (larger value for tex2html_wrap_inline2209 ) increases the overall end-to-end delays, our experiments indicate that it does not affect the relative performance of the approaches evaluated in these experiments. In particular, increasing the traffic burstiness has no effect on the relative improvement tex2html_wrap_inline2211 (defined below) for any pairing of methods.

Performance Metric.

We quantify the performance of algorithms using two measures. One is the end-to-end delay tex2html_wrap_inline2213 estimated by Algorithm X for the end-to-end delay suffered by the connection which travels the longest path in the network (Connection 0 in our case) under the work load U. The other is called the relative improvement tex2html_wrap_inline2219 , which is used to compare two algorithms and is expressed as

equation489

Delay Computation

In the Appendices B and C we summarize the formulas used for the delay calculation in the decomposition based and service-curve based approach as described in [8, 9, 10, 5, 31]. We use these formulas to derive closed forms for the worst-case delay for Connection 0 in the topology used in these experiments. We call the resulting delay computation algorithms Algorithm Decomposed for the decomposed approach and Algorithm Service Curve for the service-curve based approach.

Algorithm Decomposed

We derive the worst-case end-to-end delay of Connection 0 by adding the local delays on the servers along its path. For this, we let tex2html_wrap_inline2225 be the local delay suffered by traffic of Connection ) at Server k. In Appendix B we derive the following equations for tex2html_wrap_inline2225 :

displaymath2221

The end-to-end delay tex2html_wrap_inline2231 for Connection 0 using Algorithm  Decomposed is then obtained by adding the local delays:

displaymath2222

Algorithm Service Curve

The delay calculation in this approach is based on the definition for the service curve given in Equation C.26. As we compare the performance of the various approaches for a network with pre-defined servers (FIFO servers in this case), synthesizing scheduling algorithms from pre-defined service curves is not viable. We must use an induced service curve approach, where we derive the service curve from the scheduling policy used in the server. The performance of such a method, however, greatly depends on how tight service curves can be defined for a given service discipline. In Appendix C we derive an upper bound on the service curve for a FIFO server, which in turn give raise to a lower bound for the end-to-end delay tex2html_wrap_inline2235 for Connection 0 with the service curve method. As we derive in Appendix C, the worst case delay tex2html_wrap_inline2235 is lower-bounded by the following expression:

displaymath2233

It is important to emphasize at this point that the following comparisons are between upper bounds on end-to-end delays for both Algorithm Integrated and Algorithm Decomposed, and lower bounds for Algorithm Service Curve. The results for the performance of Algorithm Service Curve, both in terms of end-to-end delays and in terms of relative performance, must therefore be considered as optimistic.

Numerical Results and Observations

The results of our experiments comparing the performance of the three approaches are depicted in Figures (4), (gif), and (6). Figure 4 compares the service-curve based approach to the decomposition-based approach and illustrates how the former is not well suited (as was to be expected) for analyzing non-guaranteed-rate service disciplines, in this case FIFO. As the network load increases, the inadequacy of modeling a FIFO server with a service curve becomes evident. For larger systems, this gets partly offset by the compounding effects of summing convervative local delay bounds in the decomposition-based approach.

From Figure 5 we see that Algorithm Integrated always outperforms Algorithm Decomposed. Furthermore, for loads up to 80%, the performance improvement increases with growing network size. This is expected as Algorithm Integrated takes delay dependences within server pairs into account.

While the performance improvement of Algorithm Integrated over Algorithm Service Curve can be inferred by transitivity, we show a comparison in Figure 6 for illustrative purposes. The results of this experiment show that the performance gains are significant, eccept for large systems under high load.

 

  

Figure 4: Comparison between Decomposed Method and Service Curve Method.

 

 

  

Figure 5: Comparison between Integrated Method and Decomposed Method.

 

 

  

Figure 6: Comparison between Integrated Method and Service Curve Method.

 


next up previous
Next: Conclusion Up: New Delay Analysis in Previous: New Delay Analysis Algorithm

Riccardo Bettati
Wed Jul 14 18:25:49 CDT 1999