Page 76 - ITUJournal Future and evolving technologies Volume 2 (2021), Issue 1

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 1

Fig. 12 – Total average latency for ONOS controller when network is partitioned into two clusters.

Our results shows that when two controllers are deployed
and the mastership management module is activated, the
optimum controller locations are Pretoria for cluster one
and East London for cluster two, with =48.9 ms. The
worst locations are Bloemfontein and Port Elizabeth for
cluster one and cluster two respectively, with =118.4
ms. These results coincide with the results from our
mathematical formulation in Section 6.2.

7.3.2 Control‑plane failover and signalling
overhead

The outcome of our failover tests was positive in that Fig. 13 – Average latency.
all nodes could reach each other regardless of the failed
control node. This means that the switch nodes under observed. This is likely because during the transmission
thesupervisionofthefailedcontrollerwereautomatically of the irst 150 000 packets, the switch has matching
reassigned to the active controller in the other cluster. entries in its low tables on how to route traf ic, which
The reassignment took approximately 0.5 seconds. The eliminates the need to forward incoming packets to the
reassignment time was measured by carrying out the control‑plane for routing decisions. After a certain time
ONOS performance benchmark test case provided in [60]. (from 150 000 packets upwards), the switch reaches
The failure recovery time increases signi icantly with the a hard timeout and clears its low tables leading to an
number of disconnected switches [61]. The recovery additional processing delay. The additional processing
time can potentially be improved by using more powerful delay is likely the cause of the increase in average
servers with more RAM and CPU resources [62]. delay. Similar results are observed with regards to
network jitter (as shown in Fig. 14). Last but certainly
Fig. 13 and and Fig. 14 depict the results we obtained not least, when the number of packets is increased
both before and after switch‑to‑controller placement to 150 000 and 200 000, we observe a percentage
balancing. As expected (see Fig. 13), the average delay
is in overall lower after switch‑to‑controller placement
balancing compared to the case of imbalance. We
believe this is primarily because, after balancing the
switch‑to‑controller placement, data‑plane monitoring
traf ic is fairly divided between the controller nodes thus
improving overall network performance.
The decline in average delay (both before and after
switch‑to‑controller balancing) is a result of an increased
matching probability of preserved low rules with newly
arriving packets, which reduces the number of packet‑In
messages to the controller, resulting in a reduction in
network delay. When the number of packets is increased
to 150 000 packets, an increase in average delay is Fig. 14 – Average jitter.

60 © International Telecommunication Union, 2021