Conference Paper

4
Modeling Results
Using the above described SDN model, we measured the network load, therefore for
given parameters a network administrator can simply establish required quality of
service for different network nodes by managing the delay, delay variation (jitter),
bandwidth, and packet loss parameters on a network. Experimental processes were
performed using a modeling time of 3600 s, and the memory size for the simulation was
set to 1024 MB.
To illustrate the impact of various network parameters on the quality of service exist
multiple options: arrival traffic rate changes, trigger sequences of packet-in messages,
controller performance impact on the overall packet processing mean time, etc. In
Fig. 
6
, the plot highlights the switch average packet processing time for different packet
arrival rates. The more packet arrival rate increases, the more average packet processing
time increases, thus the increasing arrival traffic rate will result in network throughput
decrease. The plot can be used to determine the maximum load that the network should
reach before its performance is compromised. For a fixed packet arrival rate on each
switch we measured changes in average packet processing time while increasing the
controller service rate. The simulation results are shown in Fig. 
7
. The average packet
processing time significantly decreases as controller service rate increases. Therefore,
the network throughput increases.
Fig. 6.
Average packet processing time of switches
Analysis and Performance Evaluation of SDN Queue Model
31

---

Fig. 7.
Average packet processing time of the controller
5
Analytical Modeling Framework
To assert the described above SDN model we proceeded to analytically evaluation of
OpenFlow switch. For that we considered a queueing model [
12
] for OpenFlow-based
SDN [
13
–
15
] as illustrated in Fig. 
8
. The switches and controller are modelled as
queueing systems to capture the time cost of the network.
Fig. 8.
OpenFlow-based SDN queueing model
32
S. Muhizi et al.

---

We assume that the packet arrival process in the network follows a Poisson Process and
the average arrival rate in the ith switch is λi, and that the arrivals in different switches are
independent. Packets may not match any flow entries in which case they are forwarded to
the controller via packet-in message. This happens with probability ρ. Packets are classi‐
fied into two classes, both of them arrive in a Poisson process with an average arrival rate
of λi*ρ and λi*(1 − ρ). The packet service time of switches is assumed to follow an expo‐
nential distribution, and the expected service time is denoted 1/μ1 and 1/μ2, respectively.
The mean service time of packet-in messages in the controller is denoted 1/μc. This service
time includes the transmission time from the switches to the controller. In other, to simplify
this model, both controller and switches are powerful enough for the traffic in the network,
and there is no limit on the queue capacity. We queue all the packets arriving at a switch
in a single queue instead of a separate queue on each ingress port and all the packets are
processed in order of arrival time. Moreover, we assume that when the first packet of a
connection arrives at a switch, the controller installs a flow entry. After that, the remaining
packets arrive to the switch and are forwarded directly. We also assume that all the
switches in our model have the same service rate, and the packet-in messages arrive the
switch following a Poisson process.