in the multimedia applications.
HU
et al.: SURVEY ON SDN AND OPENFLOW: FROM CONCEPT TO IMPLEMENTATION
2201
and be upgraded solely by replacing the corresponding virtual
machine software units in the control plane servers. The MFFEs
remain intact. The entire testbed is interconnected via COTS
LAN switches.
Note that the above system adds MobileFlow level above
regular OpenFlow level. Thus it can directly use OpenFlow
language syntax for network definition. For example, it can use
the following language abstraction to monitor a port 80 traffic:
Def switch_
join(
switch)) :
P =
inport : 2
, tp_
src : 80
Install(
switch, p, DEF AU LT, [ ])
Query_
stats(
switch, p)
Such a high-level language can run in NOX and interpreted
by all OpenFlow-compatible routers. Overall, the MobileFlow
system is a complete SDN/OpenFlow implementation in mo-
bile applications. It allows flexible re-configuration of mobile
channel allocation, QoS parameters, mobile access, and service
roaming between wireless networks.
IX. R
ESEARCH
T
RENDS
There are still many questions on how to make the SDN more
efficient, how to optimize it across all the network sets, and how
to achieve tradeoffs between different implementations. There
is a need to have quantitative metrics/approach for evaluating
the performance and efficiency of the SDN such as its scalabil-
ity, availability and latency. In the following we point out some
important future research topics.
A. Intelligent Flow Table/Rules Management
It is true that the motivation of designing SDN/OpenFlow
is to simplify the network switches’ operations: instead of
using vendor-dependent, embedded software in each switch,
SDN only assumes ‘dump’ data forwarding functionalities in
each switch. The switch just simply checks the flow table to
determine how to forward each received data packet.
The tricky part is: although the switches do not analyze
the traffic, they can always report forwarding results to the
higher layer—OpenFlow controller. The results could be simple
success or failure for a data forwarding operation, or some error
messages (such as switch failure), or other forwarding status
data. In the current OpenFlow specifications, they do not spec-
ify how to handle those feedback data. They just point out that
the flow table should be set up based on a set of rules defined
by the controller. But the issue is: since the switches could have
high traffic burden, it will slow down their data forwarding
operations if (1) network traffic is heavy, and/or (2) the flow
table is large and has complex rules. It is important to perform
self-learning in the controllers based on the pattern analysis of
the traffic flowing through each switch.
We believe that the future trend of SDN/OpenFlow will in-
clude high intelligence in flow table/rules control. We illustrate
our idea in Fig. 23. The network controller can learn what’s
going on based on the feedback from the switches. For example,
if the switches report a long delay for a source-destination IP
Fig. 23.
Intelligent flow table/rules management.
pair, it indicates a possible routing loop or switch congestion
somewhere. When the controller analyzes the statistical pat-
terns from the switches’ feedback data, it can use any of the
recent learning tools (such as Bayesian learning, reinforced
learning, etc.) to deduce the optimal ‘actions’ in the future in
order to obtain a higher accumulative reward. The reward could
be defined based on the QoS performance metrics. The ‘actions’
could be any packet handling operations or any policy changes.
Statistical analysis could be based on any traffic pattern data
mining schemes. Through dimension reduction, we could ex-
tract the intrinsic features from complex, multi-attribute traffic
data. Since some dimensions are not useful in pattern recog-
nition, they could be removed by using Principle Component
Analysis (PCA), Non-negative Matrix Factorization (NMF), or
other dimension reduction schemes.
Do'stlaringiz bilan baham: