5.5 System Performance Evaluation
5.5.3 Contention Control Analysis
5.5.3.3 Hybrid Recirculating FDL/Retransmission Analysis
In the hybrid optical buffering/retransmission scheme, the recirculating optical buffering (FDLs) is utilised as the primary contention resolution technique, and the traffic that overflows from FDL buffers is dropped and then recovered by the packet retransmission mechanism. However, frequent retransmissions in the network, as previously detailed, would lead to system overloading issue. To overcome this challenge, it is necessary to provide sufficient optical buffering capacity so that the amount of colliding traffic that needs packet retransmission is reduced to a sufficiently low level. With this arrangement, the recirculating FDL buffering and the packet retransmission mechanism
compensate for each other’s disadvantageous effects. In this section, the main goal is to measure the Quality-of-Service (QoS) of different hybrid FDL/retransmission congestion control schemes under both uniform and non-uniform traffic patterns in the proposed flexible network architecture. The QoS measurement is quantified by performance metrics such as the packet loss rate, the commu- nication latency and the packet retransmission rate. Simulation results indicate that by combining the recirculating FDLs and the packet retransmission mechanisms, packet collisions in the pro- posed datacentre network can be efficiently resolved without impairing the latency performance. Additionally, the robustness of the flexible network architecture with hybrid optical buffering/re- transmission scheme to different communication patterns is also verified.
Before proceeding to the performance analysis, the dimensioning of the FDL buffers is first re- solved, based on the performance evaluation of the FDL buffering conducted in Section 5.5.3.1. Consider the trade-off between the packet loss rate and the communication latency, the parameter S is set to 8. Further, to make sure that each switch input of the OPS, which comprises K (= 4) transmission channels, is assigned a virtual buffering channel, L is dimensioned as 16. Alterna- tively, on the basis of the packet retransmission analysis in Section 5.5.3.2, the retransmission buffer size is set to 25KB, which allows the storage of 40 packets. Given these measurements, the sys- tem performance of the proposed flexible optical network employing the coordinated recirculating FDLs/retransmission contention scheme is investigated through extensive simulations.
Figures 5.17 - 5.19 plot the network performance, measured in terms of packet loss, latency and retransmission rate, as a function of the offered network load ρ for two different cases of R = 1 and 2 under all-to-all traffic pattern. Figures 5.20 - 5.22 illustrate the performance characteristics of the proposed optical network under hot-spot communication pattern. For the purpose of comparison, the packet loss and the communication latency of the optical network using FDLs only (no retrans- mission B = 0) are also presented in figures (black dashed curves). Numerical results show that the combined FDLs/retransmission approach greatly enhances the contention and latency performance, when comparing to the standalone uses of the FDL buffers and the packet retransmission schemes, which were investigated in the previous sections. Also shown in figures is how the contention probability can be reduced by up to several orders of magnitude by increasing the number of FDL circulations, R, and the number of packet retransmissions, B. Note that in the scenario of FDL/RR, B = ∞, despite that an unlimited number of retransmissions is allowed, traffic loss still exists in the optical network. The same situation occurs in the scenario of FDL/PRR, B = 1. This performance degradation results from the overflow traffic from the limited retransmission buffer. Note that as
ρ
(a) Packet loss rate versus offered network load ρ
ρ
(b) Communication latency versus offered network load ρ
Figure 5.17: Network performance of the combined contention control scheme with R = 1 under uniform traffic pattern.
ρ
(a) Packet loss rate versus offered network load ρ
ρ
(b) Communication latency versus offered network load ρ
Figure 5.18: Network performance of the combined contention control scheme with R = 2 under uniform traffic pattern.
the offered network load ρ increases, this effect becomes more significant, due to the fact that packet contentions at higher network loads are more frequent, leading to more packets overflowing from the buffer. It is also noticed that increasing the number of allowable retransmissions B in the hybrid FDL/RR scheme yields an evident reduction in packet loss rate, without significantly increasing the communication latency. Nonetheless, increasing B to a certain value in the hybrid FDL/RR scheme, the performance gain appears to be less obvious. Indeed, there is a small difference in performance between the scenarios of B = 24 and B = ∞. The simulation results further show that the hybrid FDL/PRR strategy achieves considerably better network performance than the hybrid FDL/RR scheme under low and medium load conditions. Conversely, under very high load condi- tions, the FDL/RR strategy, especially in the case of B = ∞, outperforms the FDL/PRR scheme.
ρ
(a) Number of FDL circulations R is 1
ρ
(b) Number of FDL circulations R is 2
Figure 5.19: Packet retransmission rate in the combined recirculating FDL/retransmission scheme under uniform traffic pattern.
This performance difference is attributed to the low resource utilisation of the PRR policy under heavy traffic conditions, because of the pre-reservation feature.
Under both uniform and non-uniform traffic patterns, an evident improvement in contention perfor- mance is obtained by increasing the number of FDL circulations R from 1 to 2. This is expected as increasing R results in a more efficient use of the shared buffering resources, thus allowing a greater number of the contending packets to be carried by the optical buffering. This in turn largely reduces the packet retransmission rate in the network, which will be illustrated later. Alternatively, it is observed that for a target packet loss rate equal or lower than 10−6, the overall network utilisation is also greatly improved. For example, in Figures 5.17(a) and 5.18(a), the analysis shows that in the scenario of hybrid FDL/RR scheme with B = 5, by increasing R from 1 to 2, the network utilisation is enhanced by more than 10%, approximately up to 34%. In addition, a network utilisation close to 45% is supported by the hybrid FDL/RR scheme with B = 24. Furthermore, the hybrid FDL/PRR scheme achieves a network utilisation of 55%, which corresponds to a mean resource utilisation of 220% per input/output port of the OPS, since each switch port contains K (= 4) transmission channels, as described in Section 5.5.1.1. Nevertheless, this improvement comes at the cost of a significantly increased end-to-end communication latency.
As mentioned above, in the hybrid scheme, the inclusion of the optical buffering can efficiently ad- dress the packet collisions in the optical network, thus the packet retransmission rate is significantly decreased. This aspect is illustrated in detail in Figures 5.19 and 5.22, where the packet retrans- mission rate is plotted as a function of the offered network load ρ. Comparing to the results shown in Figures 5.12(b) and 5.14(b), the packet retransmission rate in the network is dramatically low-
ρ
(a) Packet loss rate versus offered network load ρ
ρ
(b) Latency versus offered network load ρ
Figure 5.20: Network performance of the combined contention control scheme under hot-spot traf- fic pattern. Note that R = 1.
ρ
(a) Packet loss rate versus offered network load ρ
ρ
(b) Latency versus offered network load ρ
Figure 5.21: Network performance of the combined contention control scheme under hot-spot traf- fic pattern. Note that R = 2.
ered, by several orders of magnitude. Moreover, a further reduction in packet retransmission rate is gained by increasing the number of allowable FDL circulations R from 1 to 2. It should be noted that at high network loads, the recirculating FDLs become increasingly less effective at resolving contentions, and therefore the amount of traffic that needs retransmissions increases greatly. As a consequence, the packet loss due to traffic overflowing from the retransmission queue dominates the overall system performance.
Simulation studies presented in this section give an indication as to the level of performance en- hancement that can be attained by combining the recirculating optical buffers and the packet re- transmission techniques in the proposed flexible datacentre network. It is indicated that a high net-
ρ
(a) Number of FDL circulations R is 1
ρ
(b) Number of FDL circulations R is 2
Figure 5.22: Packet retransmission rate in the combined contention control scheme under non- uniform traffic pattern.
work link utilisation, approximately up to 55%, can be achieved by using the hybrid recirculating FDL/PRR scheme, which corresponds to an average resource utilisation of 220% per input/output port of the OPS, thus highlighting the efficiency of the proposed optical network structure.