4.5. SDN Design and Implementations

VLCP SDN framework allows the modules to operate Flows on a high level. With the support of lower-level modules, flow generations are easy to understand and easy to implement.

4.5.1. State Management of OpenFlow

VLCP controller always flush all the flows in the switch when it is connected to the controller. This makes sure the flows in the switches are consist with the view of the controller.

While the connection is alive, controller tries to add/remove/update minimal flows when necessary. This makes the network stable on state changes.

Modules use notifications from the transaction layer to update flows. The steps are:

  1. Query data from the central database
  2. Update flows base on the latest data
  3. Wait for update notifications from objectdb
  4. If it is necessary to restart the query, goto 1; else goto 2

With the ACID guarantees from objectdb module, it is easy to update flows in a safe way.

4.5.2. Plugable Tables

In OpenFlow, tables are identified by table ID, which is a number. Processing on a packet is done from the lower IDs to the higher IDs. This makes it difficult to extend an existed model: we need to insert or remove tables between existing tables.

VLCP uses an extensible way to allocate table IDs for each module. It uses an unique name to identify a table. Each table should declare none or more tables which must have smaller IDs than this table, they are called ancestors of this table. This makes sure a flow can use GOTO instruction from these tables to the defined table.

A table is always in one path. A path is a chain of tables which are processed one by one. VLCP inserts a default flow for each table in a path to GOTO the next table in the same path, so you can insert extra tables in a path without disturbing the original processing order. Each path also has an unique name, and the name of the default path is an empty string “”. Flows in tables can use GOTO instruction to jump to another path for extra processing. Modules may also replace the default flow in a table to drop unmatched packets or upload the packet to controller with OPF_PACKET_IN message.

On module loading, each module starts to acquire tables from openflowmanager with acquiretable API. openflowmanager module queries each module with a gettablerequest API. The API should return a tuple:

(table_requests, vhost_bind)

vhost_bind is a list of vhosts this module is binding to. It defaults to [“”], which binds only the default vHost.

table_requests is the following structure:

((name1, (ancestor1_1, ancestor1_2, ...), pathname1),
 (name2, (ancestor2_1, ancestor2_2, ...), pathname2),
 ...)

Each line acquires a table. The first element name is the unique name of this table; if multiple modules acquire the same name, it is considered to be the same table. ancestors are tuples of table names, they may not be defined in this table_requests structure. pathname is the path name of this table.

For example, the module ioprocessing defines two tables:

(("ingress", (), ''),
 ("egress", ("ingress",),''))

An ingress table and an egress table, all in the default path. The egress table must have larger ID than the ingress table. If ioprocessing is the only SDN module loaded, there will be only two tables used in the switch.

In module l2switch, more tables are defined:

(("l2input", ('ingress',), ''),
 ("l2output", ('l2input',), ''),
 ('egress', ('l2output', 'l2learning'), ''),
 ("l2learning", ('l2output',), 'l2learning'))

This creates l2input, l2output and l2learning tables. They must be in ingress -> l2input -> l2output -> l2learning -> egress order. The l2learning table is not in the default path, so a packet does not go through l2output to l2learning by default.

4.5.3. Strategies

Some modules can have different strategies. Usually there are three types of strategies:

Prepush
The controller pre-creates all flows which endpoints may need to use. This has the best stabilities and performance for reasonable sized logical networks. When load on the central database is high, there may be a delay of a few seconds before the flows are created.
Learning

The controller uses information from the incoming packets to create flows for outgoing packets. For example, input port of a packet with specified MAC address is memorised and saved to a flow. When an outgoing packet with the specified MAC address as the destination MAC is forwarded, the flow directs the packet to the original input port. If the needed flow is not created by the incoming packets, switch uses broadcast instead. This is the triditional way for switches to process packets. Extra broadcasting packets may be sent in this mode. There are two types of learning techniques:

nx_learn
This is an OpenFlow extension of OpenvSwitch. This action allows the learning procedure executed directly on OpenvSwitch, thus has better performance. This is recommended for very large scale of logical networks.
controller learning
This is a replacement for nx_learn. If you are not using OpenvSwitch (e.g. using physical switches), this uses OFP_PACKET_IN to upload the packet to controller for the learning procedure, which may increase the load of controller.
First-Upload
The switch sends a packet which does not match any exising flows to controller via OFP_PACKET_IN message. The controller looks up the information for this packet and generate a flow for it. Further packets with the same properties are processed by the created flow. This introduces a quite large delay for the first packet, but eliminates the broadcasting packets. Usually this is not recommended.

These strategies can be configured from the module configurations, see Configuration Manual for details.

4.5.4. Flow Table Design

Current SDN modules (with L3 support) and tables they used can be expressed with the figure OpenFlow Tables:

OpenFlow Tables

OpenFlow Tables

Description for each table:

ingress
This table do inital processes on the packets, initializing registers
l2input
This table drops packets which should not be forwarded (e.g. STP packets, packets with broadcast source MACs). If learning is enabled, this table uses nx_learn action or OFP_PACKET_IN to creating learning flows which matches the source MAC with the input port.
vxlaninput
If learning is enabled, this table uses nx_learn action to create learning flows which matches the source MAC with the tunnel source IP address.
arp
ARP responders. Endpoints send broadcasting ARP packets to look up the MAC address for a specified IP address. This table directly responds these broadcasting ARP packets with the correct MAC address to eliminate these ARP packets.
l3input
Embedded DHCP service uses this table to upload DHCP requests to the controller. Virtual routers uses this table to redirect packets sent to the router gateway to l3router table.
l3router
Routing tables for each virtual router. When there is a next-hop IP address, source MAC of the packet is changed to the router MAC, destination MAC of the packet is changed to the next-hop MAC address; when the next-hop is on a connected network, goto l3output
l3output
Lookup destination MAC address for L3 outgoing packets.
l2output
Lookup the output port for this packet
l2learning
If nx_learn is used, this table contains the learned flows, and is used by l2output
vxlanoutput
Lookup the tunnel destination IP address for packets in an overlay network (VXLAN)
vxlanlearning
If nx_learn is used, this table contains the learned flows, and is used by vxlanoutput
egress
Output of packets