Newer
Older
# Implementation Aspects of the goSDN Controller
## Why we do this in go
Because it rocks, but let's see afterwards what can be written here.
## Storing Information
Section XXX (Conceptual Design of a SDN Controller as Network Supervisor)
discusses the need to store information about for element inventories and
topology inventories.
### Element Inventories
Storing information about network elements and their properties is a relative
static process, at least when one considers potential changes over time.
Typically such network elements are added to a network and they will remain in
the network for a longer time, i.e., multiple minutes or even longer.
Martin Stiemerling
committed
### Topology Inventory
Every network has one given physical topology (G<sub>physical</sub> ) and on
top of this at least one logical topology (G<sub>logical1</sub>). There may be
multiple logical topologies (G<sub>n+1</sub>) on top logical topologies
Martin Stiemerling
committed
(G<sub>n</sub>), i.e., a recursion. Such logical topologies (G<sub>n+1</sub>)
can again have other logical topologies as recursion or other logical topologies
in parallel.
A topology consists out of interfaces, which are attached to their respective
network elements, and links between these interfaces.
Mathematically, such a topology can be described as a directed graph, whereas
the interfaces of the network elements are the nodes and the links are
the edges.
G<sub>physical</sub> ist a superset of G<sub>logical1</sub>.
Martin Stiemerling
committed
The topology inventory has to store the particular graph for any topology and
also the connections between the different levels of topologies. For instance,
the G<sub>logical1</sub> is linked to G<sub>physical</sub>. (needs to be clear
if changes in n-1 graph has impact on n graph).
For further study at this point: Which type of database and implementation of
databases should be used to store the different topology graphs and their
pontential dependencies? How should the interface between gosdn and this
database look like?
Here is an attempt to describe the above text in a graphical reprensetation (kinda of...not perfect yet):
```mermaid
graph TB
SubGraph1 --> SubGraph1Flow
Node1_l1[Node1_l1] <--> Node2_l1[Node2_l1] <--> Node3_l1[Node3_l1] <--> Node4_l1[Node4_l1] <--> Node5_l1[Node5_l1] <--> Node1_l1[Node1_l1]
subgraph "G_physical"
Node1[Node 1] <--> Node2[Node 2] <--> Node3[Node 3]
Node4[Node 4] <--> Node2[Node 2] <--> Node5[Node 5]
Net_physical[Net_physical] --> SubGraph1[Reference to G_logical1]
### Potential other Inventories
There may be the potential need to store information beyond pure topologies,
actually about network flows, i.e., information about a group of packets
belonging together.
## Database
A database will be used for the management and persistence of network
topologies and their associated elements within goSDN.
Since network topologies are often depicted as graphs, it was obvious to stick
to this concept and, also due to their increasing popularity, to use a graph
database. After a more intensive examination of graph databases it was found
that they (with their labels, nodes, relations and properties) are well suited
for a representation of network topologies.
The first basic idea was to create different single graphs representing the
different network topologies and label each node and edge to ensure a clear
assignment to a topology.
This would mean that nodes and edges of a graph have 1...n labels.
Therefore if you want to display a simple network topology in a graph, you can
display the different network elements as individual nodes and the edges between
network elements as their respective connections, such as Ethernet.
This works with both physical and logical topologies, which are described in
more detail [here](#topology-inventory).
So a simple topology in a graph database could look like shown below.
```mermaid
graph TD
A[Node 1 - Label: 'Host,physical'] -->|Ethernet - Label: 'physical'| B[Node 2 - Label: 'Hub,physical']
C[Node 3 - Label: 'Host,physical'] -->|Ethernet - Label: 'physical'| B
B -->|Ethernet - Label: 'physical'| D[Node 4 - Label: 'Host,physical']
B -->|Ethernet - Label: 'physical'| E[Node 5 - Label: 'Host,physical']
```
For this purpose some experiments with the [Redis](https://redis.io/)-Database
module [`RedisGraph`](https://oss.redislabs.com/redisgraph/) were carried out
first. The basic implementation was possible, but the function of assigning
several labels to one node/edge is missing (originally we considered this to be
indispensable especially to map different topologies).
For this reason we looked around for an alternative and with
[neo4j](https://neo4j.com/) we found a graph database, which gives us the
possibility to label nodes and edges with a multitude of labels and offers a
wide range of additional plugins such as [apoc](https://neo4j.com/labs/apoc/).
### neo4j
TODO: add a little description for neo4j in general
#### Implementation With neo4j
The current implementation offers the possibility to persist different network
elements (e.g. devices, interfaces...) and their physical topology and mainly
serves to represent the prototypical dataflow of goSDN to the database.
The following figure shows our first idea of a persistence of network
topologies with neo4j (to save space, only the labels were included).
PND[PND 1]
A --> |belongs to| PND
B --> |belongs to| PND
C --> |belongs to| PND
D --> |belongs to| PND
E --> |belongs to| PND
A[Label: 'Host,physical,logical1'] --> |Label: 'physical'| B[Label: 'Hub,physical,logical1']
D[Label: 'Host,physical,logical1'] --> |Label: 'physical'| B
B --> |Label: 'physical'| C[Label: 'Host,physical,logical1']
B --> |Label: 'physical'| E[Label: 'Host,physical,logical1']
A --> |Label: 'logical1'| B
B --> |Label: 'logical1'| C
C --> |Label: 'logical1'| D
D --> |Label: 'logical1'| E
E --> |Label: 'logical1'| A
The basic idea is to assign the different network elements to a specific
Principal Network Domain (PND). The different topologies are represented by a
neo4j relationship between the network elements that are stored as neo4j nodes.
However, with this current variant it is not possible, as required in
[Topology Inventory](#topology-inventory), to represent topologies that are hierarchically
interdependent, since neo4j does not allow relations to be stored as properties
(as described [here](https://neo4j.com/docs/cypher-manual/current/syntax/values/#structural-types)).
Furthermore, multiple links between the same nodes which belong to the same
topology are difficult to represent, since this model only provides a single
link between nodes of a certain topology.
For the reason mentioned above, a more complex idea for persistence is available
for the further development, which hopefully allows us to persist and map
network elements, PNDs and topologies with all their hirarchical dependencies.
The following figure tries to visualize this idea.
subgraph "dependencies of topologies"
logical1 -->|related_to| physical
logical5 -->|related_to| physical
logical3 -->|related_to| logical1
end
subgraph "every node belongs to a specific PND"
Node1 -->|belongs_to| PND
Node2 -->|belongs_to| PND
Node3 -->|belongs_to| PND
Node4 -->|belongs_to| PND
Node5 -->|belongs_to| PND
end
subgraph "relationship between nodes (nodes can be linked by 0...n links)"
lp2[link_physical]
lp3[link_physical]
lp4[link_physical]
lp5[link_logical1]
lp2 --> |connects| Node4
lp2 --> |connects| Node2
lp3 --> |connects| Node2
lp3 --> |connects| Node3
lp4 --> |connects| Node2
lp4 --> |connects| Node5
lp5 --> |connects| Node1
lp5 --> |connects| Node2
end
subgraph "links are part of a topology"
lp1[link_physical]
lp1 --> |connects| Node1
lp1 --> |connects| Node2
lp1 --> |part_of| physical
end
subgraph "links can contain 1...n layers"
lp2 --> |contains| ODUH
lp2 --> |contains| OTUCN
lp2 --> |contains| ODUCN
end
The basic structure explained in the upper part remains the same.
However, the relations, which previously served as links between the respective
nodes, now become **separate nodes**. These nodes now act as links between the
respective network elements and are part of a network topology (which itself
is represented as a separate node in the graph). By this change, network
topologies can now be interdependent. Furthermore, as can be seen in the figure
above, you can add additional nodes to the link nodes by using this scheme.
So a physical link between two nodes could e.g. **contain** several cables.
All other information can be stored in the properties of the respective nodes/edges.
The above idea is not yet approved and there are still open questions.
- Is there a better solution for the assumption that there are several different physical connections between the same nodes than separate link nodes between them?
- Can topologies run over different PNDs -> membership to different PNDs?
- Where can we benefit from using different layers? (e.g. possible saving of unnecessary relations between nodes)
- Do the sdn controllers provide us with the necessary information to map the topologies in this way?
## YANG to code
The base of the development of goSDN are YANG modules. The RESTful API used for RESTCONF is defined in an OpenAPI 2.0 file. This API documentation is generated from the YANG module. The YANG module description is also used to generate code stubs for the goSDN RESTCONF client.
YANG defines an abstract network interface. It is the foundation of the RESTCONF protocol. Several code generators exist to generate code stubs from a given definition.
### OpenAPI
OpenAPI - formerly known as Swagger - is a framework that defines RESTful APIs. We use OenAPI documentations to define the RESTCONF server implementation of the cocsn YANG modules.
### Toolchain
We use 3 different tools for the code generation workflow. For the RESTCONF server `yanger` is used to generate the OpenAPI documentation from the YANG file. `go-swagger` is used to generate a RESTCONF server with stubs for the REST calls.
The RESTCONF client stubs used by goSDN are generated from YANG files using YGOT.
### Dependencies
For now we can only use the OpenAPI 2.0 standard. This is because `go-swagger` does not support OpenAPI 3.0 specifications yet.
## Storing Information
This section keeps by now some loose thoughts about what information has to be stored how and where.
There seem to be two classes of information to be stored in the controller:
* short-living information, such as, current configured network flows or obtained network configuration out of use case #1 (CoCSN)
* long-time information, such as, information about principle network domains, elements in such a domain if directly learned from SBI, etc
Long-time information should be persistenly stored in the database and survive reboots of goSDN etc. Short-Living information doesn't have to survive reboots of goSDN
### Some more details for implementation for the database(s)
We define the principle network domain (PND) and each piece of information of any PND has to be stored in relation the particular PND.
Specification of a PND:
* Human readable name of PND
* Textual description for further information
* Set of supported Southbound-Interfaces, e.g., RESTCONF, TAPI, OpenFlow etc
* Physical Inventory Network Elements, hosts and links, pontentially only the SBI SDN controller
A PND entry must be explicitly generated, though some information can be automatically be generated, e.g., the physical inventory for use-case #1 (CoCSN) would mean that the information about the SBI domain specific SDN controller is entered.