Information Technology Reference
In-Depth Information
to deal with packets one by one, like NDN[3] etc. Openflow provides two ways
to deal with packets one by one. One way is sending packets to Openflow con-
troller, which poses a great burden on Openflow controller and is also contrary to
the architecture of separation between data-plane and control-plane. The other
way is sending packets to intelligent data processing platform, which directly
connects with Openflow switches. Due to the need of vendors for device clo-
sure emphasized by Openflow, it just provides the data-plane abstract but hides
the implementation details in data-plane. It is not conducive for researchers to
extend Openflow dataplane function to support per-packet processing.
In order to support the packet-by-packet processing of network architectures
such as NDN and etc., researchers use programmable hardware to extend the
network forwarding-plane like NetFGPA[4], NetMagic and etc., build a soft-
ware virtual router[5][6] based on general multicore processors[7][8] and deploy
Middleboxes[9][10] in networks so as to enable more functions of in-network
processing. Processing in networks provides an ecient implementation of in-
novative network functions, but the lack of a unified abstraction and control
interface is not conducive to the development and deployment of new network
protocols. We further enhance the scalability of the network equipment to sup-
port new network architectures by lending the vendor's need toward the device
closure of the resources, which lets device vendors provide intelligence processing
resources and development interface based on the network equipment resources.
At the same time, the isolation of experimental trac and production trac in
Openflow is a kind of scheduling based on packet options, which is unable to
meet the demand of dealing with special message options in new network archi-
tectures. We directly distinguish network architectures by labels, which simplify
packets classification and forwarding. Inspired by opening resources and unified
labels, we propose a software defined data-plane named LabelCast. LabelCast
identify network architectures with labels at the protocol level, while only mak-
ing abstraction for the ability of computing, storage and forwarding of network
devices, so that could support new network architectures.
In Labelcast, by dividing the data plane of the network into fast forwarding
plane and intelligence service plane based on semantics of processing in net-
works, and Label tables and Cast tables are proposed to abstract forwarding
resources and the computation and storage resources separately. LabeCast has
the following advantage: (1) Simplifying hardware implementation of forwarding
lookup based on fixed-length labels, (2) Enabling multiple user-defined applica-
tions running concurrently controlled by multiple Cast tables, (3) Achieving the
application isolation leveraging the extensible resource container, (4) Loading
multiple network architectures via LabelCast abstraction layer.
The structure of the paper is showed as follows, section 2 introduces the
evolution of network architectures, section 3 proposes a software defined data
plane, LabelCast, and the key mechanisms, such as the mapping and allocation
of labels, service atoms, section 4 provides the application of LabelCast, section
5 is the testing and analysis of the performance of the prototype, and the last
section is the conclusion of the work.
Search WWH ::




Custom Search