Page 75 - ITUJournal Future and evolving technologies Volume 2 (2021), Issue 1

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

This procedure is carried out for two scenarios: 1) results from the controller placement experiment (for the
when the switch‑to‑controller placement is imbalanced case when two control instances are deployed). In other
(switch‑to‑controller assignment is two and ive switches words, two control instances were deployed at optimal
for controller one and two respectively) and 2) for locations to minimize propagation latency. Additionally,
the scenario where switch‑to‑controller placement is the ONOS mastership management module was activated
balanced (switch‑to‑controller assignment is three and to balance the switch‑to‑controller placement.
four for controller one and two respectively).
Failover is evaluated by shutting down one controller in
In addition to switch‑to‑controller balancing, the the cluster and calling the “pingall” function. If no packet
control‑plane has several tuneable parameters in the loss is observed, then it means all hosts can reach each
control‑plane, such as polling frequency and soft idle other and switch reassignment to the active controller
timeout [57]. Polling frequency is a parameter that was successful. We also take note of the time it takes for
speci ies how frequently statistics requests are sent to the controller to take mastership of the “controller‑less”
the data‑plane. Soft idle timeout speci ies the total time switches.
an inactive low entry is stored in the low tables before
deletion. Tuning these parameters impacts control‑plane 7.3 Results and discussion
overhead. In other words, increasing polling frequency is
likely to decrease the control‑plane overhead (of course This section presents and discusses the results obtained
at the expense of data‑plane protection and restoration) from following the procedures described above.
while increasing the soft idle timeout results in more 7.3.1 Controller placement
low rules in the low tables and reduces control‑plane
overhead (with the switch resource (e.g. memory and Fig. 11 and Fig. 12 present the results obtained from
storage) exhaustion as a trade‑off). In an operational our analysis of the SANReN network. As per Fig. 11,
environment, OpenFlow switches with TCAM (Ternary our results show that the optimum controller location
Content Addressable Memory) support are typically when one controller is deployed is Cape Town since this
preferred for fast processing [58]. However, TCAM is node has the lowest average latency ( =88.78 ms).
very expensive with very limited memory space [59]. Similarly, the worst location to place the controller when
Therefore, the soft idle timeout can only be increased up one controller is deployed is Bloemfontein since this
to a certain threshold to maintain the switch memory location yields the highest average latency ( =164.4
utilization around acceptable levels. ms).

To determine how the soft idle timeout affects Fig. 12 presents the results obtained when two
control‑plane overhead, we gradually increase the controllers are deployed. These results are intepreted
soft idle timeout and polling frequency (from 5 s to 40 s as follows: the blue bars indicate a scenario where one
in increments of 5 s) and measure the number of packets controller is placed in Pretoria (a region belonging to
(i.e. Packet‑In, Packet‑Out, Flow‑Mod, Stats‑Request and cluster one as described in Section 7.2.1), while the other
Stats‑Reply). In order to evoke control traf ic we generate controller’s location is iterated between Johannesburg,
200 000 packets between two hosts (one connected to East London, Port Elizabeth and Cape Town (regions
the node in Johannesburg and the other connected to belonging to cluster two). Similarly, the red and
a node in Cape Town). The duration, packet size and green bars indicate controller placement in Durban and
bandwidth are the same as for the switch‑to‑controller Bloemfontein (regions belonging to cluster one) while the
placement experiment. This experiment leveraged the other controller is placed in all regions within cluster two.

Fig. 11 – Total average latency for the ONOS controller without clustering.

© International Telecommunication Union, 2021 59