Articles

RSS

New and emerging Digital Signal Processor (DSP)-based digital imaging systems, especially those with sophisticated surveillance and multi-spectral capabilities, must handle data at enormous rates. System architectures have evolved over time to adapt to these demanding data acquisition, data movement, data processing, and data storage requirements. The trend has been from dedicated, through federated, to distributed systems. A key technology that will support the continued movement to distributed systems is the new ANSI Fibre Channel (FC) standard. This paper presents a typical DSP system architecture, provides a technical overview of Fibre Channel, then specifically addresses its application in advanced radar, sonar, medical and other DSP-based imaging systems.

1. THE GENERIC IMAGING SYSTEM

There are four primary elements in a typical image processing system. The Sensor Subsystem scans its environment and outputs unrelenting streams of high speed digital data. The DSP Subsystem processes these data streams and outputs usable image data. The General Purpose Computers display or further analyze the images. Finally, the Mass Storage Subsystems provide archive data storage at intermediate points along the data path.

The ANSI standard Fibre Channel allows these primary elements of the image processing system to be truly distributed with distances up to 10 kilometers.

2. FIBRE CHANNEL BACKGROUND

In 1988, the ANSI-sanctioned X3T11 committee was formed to develop an ultra high-speed transport for next generation mass storage, peripheral I/O, and network communications. The objective was to develop a serial gigabit communication link that supports the bandwidth and data reliability needed by I/O channels and the flexibility, connectivity, and distance of networking technologies. The result is "Fibre Channel", a universal carrier or transporter of data, capable of simultaneously handling networking and I/O channel protocols.

3. FIBRE CHANNEL OVERVIEW

Fibre Channel defines several implementation layers. Layers FC-0, FC-1, and FC-2 are specified in the FC-PH standard. Layer FC-3 and FC-4 are standards still in development or specified in FCA Profiles such as FC-PLDA or FC-IP. The next revision of FC-PH, FC-PH-2, is nearing approval. It further refines, clarifies, and extends the existing FC-PH standard.

The FC-0 layer defines the physical layer consisting of supported baud rates, media, and connectors. The serial baud rates include 133 Mbits/sec, 266 Mbits/sec, 531 Mbits/sec, and 1.0625 Gbits/sec. Additionally, work has begun to define 2 and 4 gigabit serial baud rates.

In spite of the name "Fibre Channel", FC-0 defines copper, and both short wave laser (SWL) and long wave laser (LWL) optic media. Typical distances at full speed over copper, SWL, and LWL are 30 meters, 300 meters, and 10,000 meters respectively.

The FC-1 layer defines the encoding scheme, ordered sets, frame delimiters, and other signaling primitives implemented in the Fibre Channel encoder/decoder. Data and special transmission characters and certain combinations of these are referred to as "Ordered Sets". Ordered Sets are used to identify frame boundaries, transmit primitive function requests, and maintain proper link transmission characteristics.

The FC-2 layer is the signaling and framing protocol layer for Fibre Channel. FC-2 specifies the addressing, flow control mechanisms, sequence management, topology, and classes of service. Within the FC-2 layer the data payload, used by the FC-3 and above layers, is transparent. This is an important feature, for this allows many new and existing Upper Layer Protocols (ULP) to be mapped and simultaneously operate over a Fibre Channel network. Fundamentally, FC-2 specifies a robust set of topologies and interconnection options referred to as the Fibre Channel "fabric".

A Fibre Channel fabric may consist of any combination of three topologies: point-to-point, switched, and arbitrated loop. This set of topologies allows the system integrator a great deal of architectural flexibility. The system integrator can mix and match the best architecture for the application.

FC-2 defines three classes of service. Classes of service are not topology-dependent. The class of service used is a function of specific hardware and software driver implementation.

Class 1 is a circuit-switched, connection-oriented service that guarantees full bandwidth and in-order delivery of data because the connection is maintained and not broken or changed between frames.

Class 2 is a connectionless, frame-switched service where frames are multiplexed on frame boundaries within the fabric. In-order delivery of data is not guaranteed, unless the fabric happens to be an arbitrated loop where only one path for data flow exists.

Class 3 is a connectionless, datagram service with buffer-to-buffer flow control, but no ACK to guarantee data delivery. The fabric is expected to make a best effort to deliver the data.

The FC-3 layer is intended to define a set of services that are common across multiple ports within a single node. No standards have been completed for the FC-3 layer at this time.

The FC-4 layer defines the mapping of ULPs such as HIPPI, SCSI, IP, and a native Lightweight Protocol to the lower layers of Fibre Channel. FC-4 is generally implemented in driver software on the host computer.

The FC-4 layer allows Fibre Channel technology to support IP network connection to legacy LANs, peripheral I/O connection to mass storage arrays, and high speed host-to-host communications in distributed computational systems.

The IP protocol stack of most operating systems has proven to be a very limiting factor to attaining the high levels of throughput needed by most imaging systems. Thus, two of the most important FC-4 layer protocols for imaging systems are the Private Loop SCSI Direct Attach (FC-PLDA) for mass storage and a Lightweight Protocol for high throughput data communications between nodes.

The FC-PLDA driver looks to the operating system like a standard SCSI disk driver, which allows a Fibre Channel mass storage subsystem to seamlessly replace SCSI-based storage solution.

There is no Fibre Channel standard or profile for a high throughput Lightweight Protocol driver. To date, each FC supplier has implemented a unique protocol to attain the highest possible throughput over their Fibre Channel hardware. Some of the current implementations use the same peripheral communications model of the FC-PLDA; and other implementations, such as Systran's FibreXpress FXLP, use a standard networking model for data communications. FXLP uses the concept of a bi-directional virtual channel established between two cooperating application programs.

4. APPLICATION TO IMAGE PROCESSING

For a number of reasons, the Fibre Channel Network is an excellent technology for use in distributed image processing applications. First, very high bandwidth products are available today, operating at a gigabit per second and delivering sustained throughputs of 50-80 Mbytes/sec. Second, it is a very robust technology and offers tremendous flexibility in topology and configuration so that performance and cost can be matched to a particular system requirement. Third, there is enormous commercial market acceptance of Fibre Channel providing the volumes required to attain true commercial-off-the-shelf (COTS) pricing and product longevity advantages. The first point of application for Fibre Channel in the image processing system is the connection between the sensor and the DSP. The connection to the sensor is usually very high bandwidth, located downstream from a high-speed A/D converter.

In many image processing systems, the sensor subsystem is completely implemented in pipeline hardware. A full Fibre Channel NL_Port connection requires a CPU to manage the protocol chip and to execute FC-4 ULP driver software. However, a CPU, and all the hardware support that goes with it, is an unnecessary expense in this type of sensor subsystem. In this case, a very simple simplex data link based on the Fibre Channel FC-1 layer encoder/decoder and some FIFOs, is all that is necessary to transfer data from the sensors to the DSP.

When a CPU is present in the sensor subsystem the lightweight protocol (FXLP) works well for the sensor connection. The advantage of using a full Fibre Channel FC-4 ULP at this connection is that multiple sensor ports can be connected through a FC fabric to a single port on one or more DSPs. The FC-4 ULP also provides two-way communication between the sensor subsystem and DSP which allows synchronization and control information to be easily exchanged.

The second point of application for Fibre Channel in the image processing system, is the connection between the DSP and one or more general purpose (GP) computers. These CISC or RISC based computers perform functions such as image display, or further analysis such as automatic target recognition algorithms. This connection will generally use the fast, efficient FXLP lightweight protocol to transport the image data.

The third point of application for Fibre Channel in the image processing system is the connection between the computers (DSP or GP) and mass storage arrays. The Fibre Channel replacement of SCSI in storage arrays provides the possibility of connecting very large, high-speed mass storage systems through a single light-weight serial data link. Properly designed into the image processing system, FC-based mass storage can archive the data stream at any point along the image processing path, from the raw sensor data to processed images, and any intermediate point in between.

At the heart of the system, implementation of the Fibre Channel interface within the DSP system is a very important task. To simplify this task, all of the Fibre Channel connections (simplex and FXLP to sensors, raw data and image data to mass storage, and connections to GP workstations) should share a common implementation methodology.

5. FIBRE CHANNEL-BASED SIMPLEX LINK AS AN EXTENSION OF FPDP

The Front Panel Data Port(FPDP) was specifically invented to address the high speed connection between the Analog-to-Digital Converter(ADC) of a Sensor Subsystem and the Digital Signal Processors(DSP) of advanced DSP based image processing systems. An FPDP can be either an 80 or 160MB/s parallel connection made via ribbon cable across the VME front panel. The FPDP provides the simplicity, bandwidth, and reliability necessary to support these types of DSP systems.

However FPDP has one major limitation, the ADC must be located within the one meter maximum cable length of FPDP. For many reasons, it is often desirable to locate the ADC as close a possible to the antenna, which may be located more than one meter from the DSP system(s). The FibreXpress Simplex Link is a 1.062 GHz, 105MB/s serial link specifically designed to extend the reach of FPDP up to 10 kilometers while retaining its simplicity, bandwidth, and reliability.

The link is implemented on standard 6U VME cards with a Simplex Link Source Card (SLSC) at the sensor and a Simplex Link Destination Card (SLDC) at the DSP. When inserted into the system, the Simplex Link is completely transparent to the original FPDP connected devices except for an additional transport delay of less than 500 nSec.

The theory of operation is simple. The 32-bit FPDP input data passes through the transmit FIFO to the 8B/10B ENDEC (encoder-decoder). Once encoded the data it is inserted into the continuous stream of idle characters being sent to the serializer. The serialized data is then transmitted across the media to the destination card by the transceiver. The transceiver can be twinax up to 10 meters, shortwave laser up to 300 meters, or longwave laser over 10,000 meters. At the destination card the data is deserialized, decoded and filtered out from the idle characters, placed into the receive FIFO, and output to the FPDP cable. FPDP SYNC information, with or without data, is preserved through the use of special 8B/10B non-data characters embedded in the data stream and decoded by the ENDEC. Two PIO signals defined by FPDP are also transported across the link in a similar manner. SYNC information is placed in the FIFO with the data to preserve the timing relationship. Flow control is implemented using special 8B/10B characters on the return path already built into the ENDEC and transceivers used on the Simplex Link.

The VME SLSC and SLDC are actually implemented as CMC cards mounted on a VME carrier card which provides power and the electrical buffers to the FPDP cable. This allows the possibility that the CMC Simplex Link Card could be designed as a daughter card into either the sensor, DSP, or both eliminating the need for extra VME slots. This architecture also allows the possibility that adapter circuitry for other data interfaces, such as ECL parallel or C80 COM ports, could easily be implemented on the VME carrier card for Simplex Link transport.

Simplex Link Source and Destination Cards may also be connected directly to the I/O bus of workstations and single board computers by replacing the FPDP interface with a PMC interface. The PMC versions are completely compatible with the FPDP versions and can be mixed as needed by the image processing system. Dual DMA controllers with DMA chain list capability are used to move data to/from the PCI bus to the Simplex Link. Setup and control of the PCI interface is a relatively simple driver for most CISC and RISC based SBCs or can easily be provided as library functions for direct connection to DSP based SBCs with PMC sites.

The Simplex Link can also take advantage of the FibreXpress Network Transparent Switch (NTS). The NTS is a full crossbar switch which can copy the signal at any input port to any or all output ports. The NTS can be configured with up to eight port cards of four ports each. Thus the NTS can be configured in size from a 4x4 up to a full 32x32 crossbar switch. The NTS can also be cascaded up to five times. This allows a single Simplex Link input to be replicated and simultaneously routed to up to 32 million destinations over 50 kilometers apart for very distributed parallel processing.

The NTS in the system not only allows real-time parallel processing of the data stream, it allows the data stream to be copied to a high speed recording device. Once recorded, the NTS can reconfigure the system to allow the recording device to playback the data stream into the DSP system for post run processing and refinement of the processing algorithms.

5. SUMMARY

Fibre Channel allows all the major components of the image processing system (the sensor, the DSP, the GP workstations, and the storage) to be separated by distances up to 10 kilometers. It is available in a variety of speeds, media, and topologies so that performance and cost requirements can be easily matched. It is scaleable, providing a growth path for systems that start small, but must expand to meet the growing requirements of today's distributed image processing systems.

BIBLIOGRAPHY

[1] Fibre Channel Physical and Signaling Interface (FC-PH) Revision 4.3, August 24, 1994

[2] Fibre Channel Arbitrated Loop (FC-AL) Rev 4.5, June 1, 1995