CHAPTER 5. SIMULATION RESULTS
5.3 P ERFORMANCE C OMPARISONS
Taking into account the assumptions made in the previous section, it will draw comparisons between the DMGWR scheme and two other existing schemes (WDMA and DQBR), and demonstrate its performance with respect to throughput, access delay, and fairness under various system settings, via simulation.
We first compare the impact the rapid growth of the number of wavelengths has on the network throughput of the three schemes (see Figure 5.1). In the experiments, the number of network nodes is 48 and the numbers of wavelengths (W) range between 2 and 24. As the total number of wavelengths increases, the total load is proportionally scaled up. When upgrading the overall network capacity to support the high bandwidth demand applications, both DQBR and WDMA unfortunately manifest a deteriorating throughput performance as the number of wavelengths is scaled up.
Results confirm the limitations of the multiple distributed queues approach and highlight the performance penalty of DQBR already discussed. Note that WDMA deteriorates the most. This is due to its collision and retransmission scheme, which produces noticeable bandwidth waste. We point out that the normalized throughput of DMGWR exceeds that of DQBR and WDMA several orders of magnitude for larger network capacity particularly under extremely heavy load (0.99). Furthermore, the DMGWR scheme achieves the same degree of bandwidth efficiency irrespective of the wavelength number and load (L) of the network. This is a very attractive scalability feature for the next-generation commercial WDM network systems.
As depicted in Figure 5.2, it compares aggregate throughput performance among three schemes when subjected to various loads and number of wavelengths. In the experiments, the number of network nodes is 48 and the numbers of wavelengths (W) range between 4 and 16 under various traffic load. Under the heavy load (0.75), the aggregate throughput of all of the schemes is proportionally scaled up as the total number of wavelengths increases. It observed that the aggregate throughput of DMGWR exceeds that of DQBR and WDMA several orders of magnitude for larger network capacity particularly under extremely heavy load (0.99). When upgrading W, both DQBR and WDMA unfortunately have a direct manifestation of the deteriorating aggregate throughput performance. This is understood by observing that a larger W leads to heavier traffic, i.e., mean packet arrival rate pr = (2×W×L)/N. As the ratio of
W/N increases, the pr on each node increases twice and more access collisions occur.
In case of the lack of statistical multiplexing gain, when the maximum transmission capacity of one protocol is reached, the congestion state is immediately deteriorated
2 4 6 8 10 12 14 16 18 20 22 24 0.7
0.8 0.9 1.0
Figure 5.1. Throughput comparison.
Normalized throughput
Number of wavelengths (W) 48 nodes
D = 10
WDMA DQBR DMGWR
L=0.8 L=0.99
by the almost constant arrival of packets and the node buffers are persistently filled with packets. As a direct result the throughput of either DQBR or WDMA abruptly decline when the congestion state is reached.
Furthermore, it compares the delay performance of the three schemes when subjected to various loads and burstiness. As expected, delay unfairness is manifested in both DQBR scheme and WDMA scheme. DMGWR outperforms DQBR and WDMA significantly, especially under highly bursty traffic. From Figure 5.3(a), one can see that as the burstiness increases, delays of WDMA and DQBR increase rapidly.
However, DMGWR achieves a bounded delay even in highly bursty traffic. Note that in Figure 5.3(b), when upgrading the network capacity as done in Figure 5.1, both DQBR and WDMA unfortunately have a direct manifestation of the larger mean delay as the wavelength number is scaled up. This is understood by observing that as the ratio of W/N increases, the mean packet arrival rate on each node increases twice and more access conflicts occur. When the maximum transmission capacity of one protocol is reached, the congestion state is immediately deteriorated by the almost constant arrival of packets and the node buffers are persistently filled with packets. As a direct result the mean access delay of either DQBR or WDMA increases abruptly when the congestion state is reached. As opposed to DQBR and WDMA, the DMGWR protocol is show to improve mean delay performance when upgrading the number of wavelength. The performance improvement is achieved by superior design of the DMGWR protocol that accommodates highly wavelength sharing among the competing nodes.
Figure 5.4 depicts the average access delay at each node under different inter-nodal distances (D). As expected, delay unfairness is manifested in both DQBR and WDMA. A larger ring length in WDMA leads to longer wait of upstream nodes for retransmission in the case of a packet collision. Consequently, the total delay of
upstream nodes drastically increases with the ring length in WDMA. We also observe that the delay of DQBR is a serrated curve on downstream side due to the impact of cyclic home channel assignment of DQBR. On the other hand, DMGWR enables fairness and efficiently shares the bandwidth between competing nodes. In addition, the mean access delay of DMGWR is independent of the ring length.
In Figure 5.5, we compare the average access delay of the three schemes when subjected to various burstiness. As depicted in Figure 5.5, DMGWR outperforms DQBR and WDMA significantly, especially under highly bursty traffic. From Figure 5.5, one can see that as the burstiness increases, delays of WDMA and DQBR increase rapidly. However, DMGWR achieves a bounded delay even in highly bursty traffic. We also observe in the Figure 5.5 that the delay of DQBR is plotted by a serrated curve (around nodes 36 and 44) on the downstream side due to the impact of a cyclic home channel assignment of DQBR. DMGWR guarantees delay under different burstiness settings. Indeed, with the ability of dynamically bandwidth allocatcation, the DMGWR can provide more efficiently resource allocation in response to bursty data traffic and time-varying traffic conditions.
We furthermore draw a delay comparison between DMGWR and DQBR over a network with malicious nodes. In the simulation, nodes 10 and/or 15 are set as malicious nodes, where each node generated an excessive load of 0.09 per wavelength, in a network with a total load of 0.85 per wavelength. As displayed in Figure 5.6(a) and 5.6(b), DMGWR causes the malicious nodes to suffer severe delays, while leaving other normal nodes completely unaffected. On the other hand, the DQBR scheme results in unexpected delay deterioration (and thus unfairness) for the downstream nodes. As the number of malicious nodes increases, the delay unfairness problem worsens, as Figure 5.6(b) demonstrates. In this case, DMGWR can still guarantee a high grade of fairness among all nodes. Thus, the DMGWR scheme is
robust and fair even when under attack by malevolent nodes.
Moreover, we observe performance comparison between DMGWR and DQBR over a network with malicious nodes under extremely heavy loads. As displayed in Figure 5.7(a) and Figure 5.7(b), under the extremely heavy loads condition, DQBR scheme results in unexpected delay deterioration for the downstream nodes. Note that in Figure 5.7(b), when upgrading the network capacity as done in Figure 5.1, DQBR scheme unfortunately have a direct manifestation of the throughput performance as the network capacity is scaled up.
Furthermore, we examine the impact of wavelength sharing on the throughput performance between DMGWR and DQBR under various loads for the same ratio of the number of nodes and wavelengths. As depicted in Figure 5.8, the static wavelength sharing scheme undergoes severe throughput deterioration toward downstream nodes under heavy loads resulting from the performance penalty of DQBR. Under the same ratio of the number of nodes and wavelengths, the larger the number of nodes, the worse the normalized throughput. In contrast to this, as a result of the full wavelength sharing and multiple-granularity-window design, DMGWR achieves 100% throughput under all loads and network configurations. Finally, we used 40/80 nodes and 8 wavelengths in the network simulations to compare the impact of wavelength sharing on the throughput performance. Also compare Figure 5.9 that increasing the number of nodes by a factor of 2, DQBR results in a lower normalized throughput and worse performance.
Aggregate throughput
0.70 0.75 0.80 0.85 0.90 0.95 1.00 0
(a) Aggregate throughput comparison (W=4 and 8)
Load (L)
Figure 5.2. Aggregate throughput comparison under various loads and number of wavelengths.
0.70 0.75 0.80 0.85 0.90 0.95 1.00 0
(b) Aggregate throughput comparison (W=12 and 16) WDMA
DQBR DMGWR
Figure 5.3. Access delay comparison under various bursstiness and load.
(b) Delay comparison under various load
Mean delay (slots)
(a) Delay comparison under various bursstiness
WDMA DQBR DMGWR
Figure 5.4. Comparisons of delay Fairness under various inter-nodal distances.
0 6 12 18 24 30 36 42 48
0 10 20 30
40 D=5 D=1
Node ID WDMA
DQBR DMGWR
W = 4
L = 0.85
Mean delay (slots)
0 6 12 18 24 30 36 42 48
0 20 40 60 80
Figure 5.5. Comparisons of delay fairness under various burstiness.
Node ID WDMA
DQBR DMGWR
B=3 B=6
W = 4 L = 0.85
Mean delay (slots)
0 4 8 12 16 20 24 28 32 36 40 0
30 60 90 120 800 1000
DQBR DMGWR
Figure 5.6. Delay comparisons for netwok with malicious nodes. (a) Delay comparisons for malicious node (Node 10). (b) Delay comparisons for
malicious nodes (Nodes 10 and 15).
0 4 8 12 16 20 24 28 32 36 40
0 30 60 90 120 800 1000
DQBR DMWGR
(b) Delay comparisons for malicious nodes (Nodes 10 and 15)
Mean delay (slots)
Node ID (a) Delay comparisons for malicious node (Node 10)
Mean delay (slots)
Node ID
Figure 5.7. Performance comparisons for network with malicious nodes. (a) Mean delay comparisons for malicious nodes (Nodes 10 and 15) under extremely heavy load. (b) Throughput comparison for malicious nodes (Nodes 10 and 15).
0 4 8 12 16 20 24 28 32 36 40
0.80 0.84 0.88 0.92 0.96 1.00
0.7
(a) Mean delay comparisons for malicious nodes (Nodes 10 and 15) under extremely heavy load
(b) Throughput comparison for malicious nodes (Nodes 10 and 15) Load (L)
Mean dela y (slots ) Normalized throu gh put
DQBR DMGWR
0 4 8 12 16 20 24 28 32 36 40
(a) Normalized throughput fairness comparisons when N is 40 and W is 4.Normalized throughput (a) Normalized throughput fairness comparisons when N is 20 and W is 2.
Normalized throughput
Figure 5.8. The impact of wavelength sharing on throughput performance under the same sharing ratio.
Figure 5.9. The impact of wavelength sharing on throughput performance with the same (a) Throughput fairness comparison for 80 nodes.
Normalized throughput (b) Throughput fairness comparison for 40 nodes.
Normalized throughput