Simulation Results

Simulation Results

A simulator of the iPOINT switch has been enhanced with ABR functionality. For the simulations, two standard fair rate allocation algorithms have been implemented. The first one, called "Dynamic Max Rate Control Algorithm" (DMRCA) is an approximate fair rate calculation algorithm. The second algorithm is the enhanced Explicit Rate Indication for Congestion Avoidance (ERICA+) algorithm, is an exact fair rate calculation algorithm. To evaluate their performance with the Virtual Output Queue, this paper simulates the standard configuration given below.

Parking Lot Configuration

Parking_Lot.bmp (29550 bytes)

Figure 1 Parking lot configuration

In the parking lot configuration, one connection (VC 3) passes across fewer switches than the others. The bottleneck is at the output of switch 2. If the ABR algorithm is fair, each ABR connection gets one third of the available ABR bandwidth. Algorithms with fairness problems will allocate a higher rate to the connection that travels fewer switches (VC 3). VC 4 transmits VBR traffic. Its effect is to modulate the available ABR bandwidth.

Influence of Virtual Output Queue

The length of the virtual output queue is an estimation of the number of ABR cells stored in the switch for a given link. Since the differential queue length is written into the RM-cell before they enter the per-VC queue, the information transmitted by the RM-cell is delayed as a function of the per-VC queue lengths. Thus, it takes more time until the algorithm can react. The goal of this section is to determine the penalty of this delay.

In order to determine the magnitude of this performance penalty, simulations with virtual output queue are compared to simulations where the queue length can be determined without delay.

voq_DMRCA.gif (33027 bytes)

Figure 2 DMRCA : influence of virtual output queue (queue length: switch 2)

The effect of the virtual output queue is quite minor in the simulation of the DMRCA (fig. 2). The queue swing is generally a bit larger during transient phases. This effect is caused by delay in obtaining the queue length information.

When using the virtual output queue, the second peak of the queue length is approximately 1.1 (3600 / 3300) times larger than in the simulation with non-delayed queue information.

voq_ERICA+.gif (28683 bytes)

Figure 3 ERICA+ : influence of virtual output queue (queue length: switch 2)

ERICA+ is more sensitive to the delayed queue length information. When using the virtual output queue, the second peak of the queue length is about 1.6 (3400 / 2200) (compared to 1.1 for DMRCA) times larger than in the simulation with non delayed queue information. Observe that the cell rate allocation is only slightly delayed and that the undershots are a bit more pronounced (fig. 3).

Figure 4 and 5 show the maximum ABR queue length as a function of the VBR max-min bit rate ratio. These graphs show how the virtual queue length is affected by increasing changes in the Max-to-Min Bit Rate Ratio.

Figure 4 DMRCA : maximum queue length as a function of the VBR Max-to-Min Bit Rate Ratio

The performance of the DMRCA is not significantly influenced by the delayed queue length information caused by the virtual output queue (fig. 10). Even when the max-to-min bit rate ratio increases to a factor of six, the maximum ABR queue length does not become noticeably larger than it does without the delay.

Figure 5 ERICA+ : maximum queue length as a function of the VBR Max-to-Min Bit Rate Ratio

ERICA+ shows to be more sensitive to the delay caused by the virtual output queue. We observe that in figure 11 the maximum ABR queue length grows faster with the increasing max-to-min bit rate ratio, when using the virtual output queue.

Last updated 03/12/98 by Matthias Bossardt