OpenFlow Data Plane Abstraction (OF-DPA) is an application software component that implements an adaptation layer between OpenFlow and the Broadcom Silicon SDK. OF-DPA enables scalable implementation of OpenFlow 1.3 on Broadcom switch devices. Figure 1 shows the OF-DPA software component and its relation to other OpenFlow system components.
In Figure 1, user applications obtain services from an OpenFlow Controller via a northbound API. The controller's northbound API enables applications to communicate with and control one or more OpenFlow switches. The controller maintains a channel with each OpenFlow switch, over which it exchanges OpenFlow protocol messages. At the switches, OpenFlow agents maintain their end of the channel, process received OpenFlow protocol messages and send OpenFlow messages in response to local events. OF-DPA fills a key role in this architecture by providing the OF-DPA API to the agent. This API represents hardware tables to the agent in terms of OpenFlow constructs such as flow tables, the group table, and ports. As OF-DPA maintains all of the state that matches the hardware tables to OpenFlow, the agent is expected to do a relatively straightforward and stateless translation of OpenFlow messages into OF-DPA API calls and vice versa. The application writer programs to the description of the OF-DPA Abstract Switch. Refer to the "OF-DPA Architecture Specification" document for details of the abstract switch model implemented by OF-DPA software.
OF-DPA Software Architecture
Figure 2 shows a block diagram representation of the various components of OF-DPA.
- API Layer: Provides the northbound OF-DPA APIs. These are available via RPC in the inter-process mode or as direct function calls to a user application that links with the OF-DPA libraries. OF-DPA does not have a user interface (e.g. a CLI). All control is done via APIs. All APIs calls are synchronous and control is returned to the caller after changes are committed to the hardware.
- OFDB Layer: This is the OF-DPA database layer. OFDB is the software database for the flow, group and port tables. It provides APIs to manage these tables. OFDB APIs are invoked by the following layers of OF-DPA:
- API: OFDB stores the management state of the system. OFDB APIs are invoked by the API layer for configuration updates like flow addition, deletion, etc.
- Mapping: OFDB also stores system status information like port link status, etc. The mapping layer receives these port state updates from the hardware and invokes OFDB APIs to update the port tables in the database.
- Datapath: The datapath layer invokes OFDB APIs to traverse through flow tables etc. and perform housekeeping operations like flow aging.
- Datapath Layer: Implements a task that wakes up periodically and traverses flow tables to perform flow aging operations. On detecting aged flows, it sends event notifications to client applications.
- Mapping/Driver Layer: Provides system mapping and driver functions.
- All system initialization is done here. The driver, OFDB, and datapath layers are initialized during system startup. Functions to retrieve system specific information like table sizes, valid VLANs, etc. are implemented in this layer.
- The driver functions provide an abstraction layer over the Broadcom SDK. These functions are invoked to program and query the hardware. The driver layer defines the interface OF-DPA requires when ported to other platforms.
- Platform Layer: Provides platform specific initialization structures and functions. These functions are invoked during system startup.
Operational Modes
OF-DPA software can be built and used in two modes:
- Interprocess Mode: OF-DPA daemon runs as an independent process on the switch. The user application invokes OF-DPA APIs using RPC. Only a single client user application can interface with OF-DPA at a time.
- Single Process Mode: The user application is statically linked with OF-DPA. A single Linux executable contains the user application, OF-DPA layers, and the Broadcom SDK.
Interprocess communication overhead results in a performance difference between these two modes. Applications implemented using the single process mode are able to invoke OF-DPA APIs at a higher rate than when using the interprocess mode. Interprocess mode is useful for experimenting with the OF-DPA software. An example of this is using the sample client applications included in the OF-DPA software package to configure the pipeline for test scenarios. Applications designed to configure the switch for network operation, like an OpenFlow agent, should use the single process mode.
- Note
- The interprocess interface can also be available when running in single process mode.
Initialization Overview
The system initializes in the following sequence:
- SDK: The SDK is initialized with a default configuration for the chip.
- Platform: The platform's ID, description and physical port mapping information is read from the platform specific structures.
- Driver: The hardware is initialized here.
- OFDB: Memory is allocated for the various tables in the OFDB including flow, group, and port tables.
- Datapath: A socket is created that is used to send event notifications generated by the datapath layer. A task is started that provides flow entry housekeeping operations like aging.
The RPC server task is started after system initialization. This task listens on a socket for RPC messages from client applications.
Steady State Overview
In steady state, a few background threads like the SDK counter, packet rx and datapath threads run periodically. All other processing is triggered only on API invocations.
SDK Configuration and Setup
OF-DPA software is built on top of the BCM APIs and data structures. A reference integration of OF-DPA software with the SDK is provided as part of the OF-DPA software release. The reference integration builds the SDK in Linux 'User' mode. SDK initialization that is needed for OF-DPA software is done as part of the OF-DPA initialization. Customers can change the OF-DPA SDK integration to fit their environment.
