This chapter presents the simulation results concerning the packet loss probability and average recirculation time. We consider input-queued switch in evaluating the performance of URODL. Only the packets at the head of the queues (i.e. the packet who’s POS is zero) can content for their destined outputs. In window-base lookahead URODL, the first w packets in each input queue sequentially content for access to the switch outputs. In our simulation, packets arrive at each input port following Bernoulli process with different offered load. The destination of each incoming packet is uniform distributed. That is, for a given offered load ρ in an NxN switch, the packet arrives at each input port with probability ρ and each packet destines to a particular output port with probability 1/N. The multilevel priority URODL serves packets on a “straight priority” basis, i.e., if there exists a higher priority packet in multilevel priority URODL, and then no lower priority packets can be served. When multilevel priority URODL is full and an incoming packet whose priority is higher than one of the packets in URODL, the packet at end of URODL is dropped and the incoming packet is inserted to a position determined by the control algorithm detailed in Chapter 4. URODL is hereafter assumed to support three priority levels – high, medium and low.
(a) buffer size = 16
(b) buffer size = 64
Figure 10. Simulation results concerning packet loss probability of 32-port input-queued switch with different buffer size.
Figure 10 presents the simulation results of traveling-type ODL with and without recirculation in a 32x32 switch. Fig. 10.a and 10.b plot the distribution of packet loss probability with buffer sizes of 16 and 64 respectively. In Fig. 10, the gap between the two curves is caused by the packet traveling out of ODL, while contention happens.
The traveling-type ODL cannot hold the packets propagate out of ODL, but URODL can do so by recirculation. The simulation result shows URODL can reduce the packet
loss probability significantly.
Figure 11. The throughput of different window-based lookahead architecture in 32-port switch with buffer size 16.
In Fig. 11, we show the simulation result for the proposed window-based lookahead URODL architecture with window size w increases from 1 to 8. Notice that when a window size w = 1 corresponds to the original URODL. The maximum throughput can be achieved are 0.56, 0.69, 0.75, 0.79, 0.82, 0.84, 0.85, and 0.86 respectively for various window size w = 1, 2, 3, 4, 5, 6, 7 and 8. The throughput improves significantly (about 23%) by increasing window size w to 2. The window size of 4 improves the maximum throughput by about 5% compared with the window size of 3. As shown in the simulation results, the improvement decreases to about 1%, as the window size increases from 6 to 7, and 7 to 8. Therefore, in order to reduce the HOL blocking people can choose the window-based lookahead URODL with window size of 2, 3 or 4 instead of large window size which needs more additional optical components and slower scheduling time.
(a) (H, M, L) = (33.3%, 33.3%, 33.3%)
(b) (H, M, L) = (20%, 30%, 50%)
(c) (H, M, L) = (70%, 20%, 10%)
Figure 12. Simulation results of packet loss probability of 32-port input-queued switch with different distribution of packet priority.
(a) (H, M, L) = (33.3%, 33.3%, 33.3%)
(b) (H, M, L) = (20%, 30%, 50%)
(c) (H, M, L) = (70%, 20%, 10%)
Figure 13. Simulation results of packet loss probability of 32-port input-queued switch with different distribution of packet priority.
Figure 12 and Figure 13 depicts the simulation results of URODL with QoS support, and compares it to the URODL without QoS support, using a 32x32 switch and a buffer size of 16. In Fig. 12 the packet loss probability is computed by number of packet loss in each class divided by total number of incoming packet. In Fig. 13 the relative packet loss probability is computed by dividing the number of lost packets in each class by the total number of incoming packets in each class. Three priority levels, high, medium and low, are assumed. The URODL without QoS support serve packets without distinguishing the packet priority. The control strategy we proposed in this paper for the QoS enable URODL is used to simulate the packet loss probability among three priority levels. Three different combinations of incoming packet priority are used to observe the packet loss probability. In uniform priority distribution as shown in Fig 12.a, the low and medium priority packets of multilevel priority URODL inherit the packet loss probability of the high priority packets, so the loss probability of high priority packets almost reaches zero. Even in non-uniform priority distribution shown in Fig. 12.b and Fig. 12.c, the high priority packet loss probability is still close to zero. In Fig 12.c, the high priority packets take 70% of overall incoming packets, the high priority packet loss probability remains zero under 80%
offered load. However, in URODL without QoS support, the loss probability associated with each priority level is proportional to the distribution of the incoming packet priority; the higher priority packet may be dropped even when the lower priority packet is in the queue. Fig 12.a shows the packet loss rates are same among high, medium and low priority packets while URODL without QoS support is used.
Furthermore, in Fig 12.c and Fig 13.c, we can observe when the loss rate of high priority packets reaches zero under 80% traffic load, the loss rate of medium priority begins decreasing. It happens again when the loss rate of medium priority packet gets down to zero, the loss rate of low priority packets starts decreasing. The results show
that the higher priority packets are given the preemptive authority to prevent loss in multilevel priority URODL. In Fig. 12.c the packet loss probability of medium priority packets in multilevel priority URODL is higher than the low priority packets, this is because the medium priority packets take 20% of overall incoming packets and the low priority packets only take 10% of overall incoming packets. Figure 13 shows that the relative packet loss of each priority class. In Figure 13 we see more clearly about packet loss probability among three different priority classes. They show that the relative packet loss probability of higher priority packets is always less then the lower priority packets in multilevel priority URODL. However, in URODL without QoS support the relative packet loss probability among three priority classes are the same. This shows that our multilevel priority URODL can guarantee QoS by distinguishing packets among the three different priorities.
(a) average number of recirculation
(b) maximal number of recirculation
Figure 14. Average and maximal times of recirculation of 32-port switch with buffer size 16.
In simulating packet loss probability of URODL and multilevel priority URODL, statistics are also gathered on the average and maximal times of recirculation in multilevel priority URODL, as shown in Fig 14. Because of insertion loss appearing invariably in AWG and ODL, the times of recirculation while a packet is waiting for scheduling may cause the intensity of light to decline. An optical amplification can be
used to URODL to boost the optical signal. However, ASE noise limits the number of times that a packet can travel through the optical amplifier. In URODL, the average times of recirculation in Fig 14.a indicate that the high and medium priority packets require only few recirculations on the average. The high priority packets experience average less than two recirculations. The maximal number of recirculations in Fig 14.b shows that in the worst case a low priority packet could be recirculated in a large number of times. Even though the lower priority packets may be dropped due to large number of recirculation, the optical signal integrity of higher priority packets can be guaranteed.