https://dx.doi.org/10.1109/MILCOM.2013.172 describes an implementation of deep packet inspection into an SDN.

Very thanks for your recommendation. Actually, i have read that  article in not long ago. However, i want to know whether it is possible to introduce DPI into the control plane of SDN and can DPI  aware users' activities from the traffic flows? How much impact DPI will have on the performance of SDN controller?

If there does not exist an appropriate treatise in the literature that answers your questions, you have found a perfect research goal for yourself.

Having a network application utilizing DPI connected to an SDN controller over its northbound interface is standard practice. In fact, I would say that is the whole idea of reactive SDN - sending a first packet to be inspected, and then having the controller configure the switches to properly properly forward the following packets.

Y(J)S

I've come up with a very relevant white paper from Qosmos. http://www.qosmos.com/wp-content/uploads/2013/09/Intel-Qosmos_White-Paper_SDN-NFV_Architecture.pdf that discusses how SDN and DPI can work togerther.

Note: To be honest I do not know who the Qosmos guys really are and I've no intention to promote them. I just think the content of the white paper is directly related to Li's question.

We demonstrated an OpenFlow based (using OXM to communicate application layer match requirements to OF controller, which get passed on to a Snort box attached to the datapath) integrated DPI at GEC23.  A demo video is here: https://drive.google.com/open?id=0B97vODlLNDkoNkEyeEhyTkRZUzQ , and the technique has been published in our CoolSDN (ICNP) 2014 paper.  Hope this helps.

https://drive.google.com/open?id=0B97vODlLNDkoNkEyeEhyTkRZUzQ

Atay - I have worked with Qosmos, Sandvine, and Allot for this application, and they all do great jobs.

Y(J)S

Architecturally, why would you place the DPI functioning in the control plane and not the data plane? Research suggests that the DPI function would be performed by middleboxes in the data plane. What problem are you trying to solve? Are you looking for more visibility into the DPI function, the reason for moving the DPI function into the control plane?

Harold-- the white paper recommended by Atay may answer your first question. Moving the DPI function into the control plane may let us "see" deeper into the network. If so, L4~ L7 flow switching may be provided.  Benefits for security, more fine-grained traffic control and differentiated services may be achieved.  Of course, this may facilitate the extension of OpenFlow. 

All of  your answers gives me a lot of  inspiration.  Yesterday,  i attend a meet held in Shanghai, China. Two reporters came from Intel and Big Switch separately refered to this issue.  Yaakov--- can you introduce more about Qosmos or provide some better papers or other documents about it?