Round Trip Time

Figure 6.6 shows the average round trip time (RTT) of TCP segments for three values of TDC across a range of burst assembly parameters in both designs. The round trip time is measured from when a TCP segment is sent by the TCP application in a source server to when its acknowledgement is received at the source server. The RTT in all three cases of TDC across different burst aggregation schemes in the proposed

6.5. CONCLUSION

scheme is around 500µs. It can be noticed that the aggregation time that the proposed

technique considers is only 50 µs and 100 µs while the RTT is around 500 µs. This

is because TCP segments also spend some time in the queue of NIC in servers and ToR switches. When there is a high traffic load, these segments have to wait in the queue to get transmitted. This is not the case in the traditional methods of OBS using

one-way reservation. For example, with TDC = 4 and TDC = 8 as shown in Figures

6.6(b) and 6.6(c), the RTT in most of the cases is around 200µs because there is not much traffic to send due to a lower throughput observed in the traditional methods of OBS. So, TCP segments find it easy to go through the queue of NIC in low traffic that results in lower RTT. The RTT in the electronic network is higher than in both OBS schemes as shown in Figure 6.3(b). This is because of the additional delay which is incurred at the core node while in traditional and proposed methods of OBS delay is incurred only at the edge node.

The bandwidth delay product in a data centre network is low compared to that of a backbone long haul optical network due to its low RTT. The effect of BDP can be eliminated if we increase the number of TCP flows per server or the window size. By considering a default window size of 64K B and 40 TCP flows per server, good TCP performance comparable to that of the electrical data centre network has been achieved.

6.5

Conclusion

In this chapter, the performance of TCP over an optical burst-switched FOSA for data centre network is investigated by using network-level simulation. Burst loss is the ma- jor limitation of traditional OBS that degrades the performance of TCP. Due to the zero burst loss in OBS with two-way reservation, efficient TCP performance is observed.

Various burst assembly parameters were examined with various traffic workloads to evaluate the performance of TCP. The results show a significant improvement in the throughput, completion time and packet loss as compared to the traditional methods of OBS across all types of workloads. The proposed scheme also demonstrates better

6.5. CONCLUSION

TCP performance than the conventional electronic packet switching network for all types of workloads.

CHAPTER 7

CONCLUSIONS AND FUTURE WORK

7.1

Conclusions

This thesis presents three novel optical interconnection schemes for data centre net- works which are based on OBS. OBS was initially proposed for long-haul backbone optical core networks but it has not replaced OCS due to its limitation of high burst loss in this application. The proposed schemes consider OBS with a two-way reserva- tion protocol that ensures zero burst loss. In two-way reservation, the connection is established for each burst before transmission. The two-way reservation is not suitable for long-haul backbone optical networks due to the high RTT of the control packet but this RTT is not high in a DCN. The proposed designs employ a single-stage core topol- ogy with multiple optical switches that has the capacity to be scaled up and scaled out easily. Network-level simulation is considered to evaluate the performance of the proposed schemes.

First optical interconnection scheme proposed (HOSA) is based on the use of both fast and slow optical switches. HOSA leverages the strengths of both types of opti- cal switch. The hybrid architecture features MEMS OXCs for low cost and uses fast

7.1. CONCLUSIONS

optical switches to achieve low latency. The core idea is to use fast optical switches to hide the lengthy reconfiguration procedure of the slow MEMS switches from users. HOSA features separate data and control planes. The control plane comprises a cen- tralized controller while the data plane contains an array of fast and slow optical switches. A scalability analysis of HOSA, investigating various ratios of slow and fast optical switches, has been done. Scalability analysis shows that the proposed design is scalable to more than hundred thousand servers which is suitable for data centres of very large scale. A comparison of the cost and power consumption of the proposed design and those of conventional interconnects by using analytical modelling is also presented. The results show 50% power efficiency as compared to other conventional

electrical networks while 30− 35% improvement in power consumption is achieved

over hybrid optical/electrical network.

The trade-off between the performance and the capacity of both types of switch is also presented. The results indicate that the proposed hybrid technique, where only 40% of the interconnect capacity is provided by fast switches, shows performance comparable to that of an interconnect exclusively using fast switches till 83% load, while the cost of the hybrid architecture is reduced by 33% more than half compared to using fast switches only. However, the cost of the hybrid design is less than the Fat tree and the BCube networks.

The large aggregation time of bursts is the limitation of HOSA. In order to over- come this limitation, a second optical interconnection scheme (HOSA with TDS) is proposed. In HOSA with TDS, there is no need to aggregate large amount of traf- fic. The controller maintains a traffic demand matrix which updates traffic demand periodically and assigns slow paths to elephant flows. The results show low latency and high throughput for various workload communication patterns. The throughput achieved in HOSA with TDS is almost the same as that of the baseline electrical net- work while slightly higher latency than the baseline electrical network is achieved. However, the performance of HOSA has been improved by introducing HOSA with TDS scheme.