els with a cabling length of at least 3 meters.
With SCSI, or even with Wide and Ultra SCSI's maximum cable lengths of 3 and 1.5 meters, respectively, this is hardly possible.
The same problem is prevalent in the local networking area. Ethernet provides a reach of a few hundred meters. Above that limit, connections become flaky. The Fibre Channel architecture aims at eliminating throughput and distance limitations for both peripherals and LAN infrastructures.
The basic concept of a transparent high-end connection technology that encapsulates already established and proven transfer protocols has been around for several years (see "Fibre Channel: Fast and Flexible," May 1996 BYTE). One of the most distinctive features of the Fibre Channel is compatibility with existing software. Unlike SCSI and asynchronous transfer mode (ATM), the Fibre Channel protocol does not have its own command set, but simply manages the data transfer from port to port and thus interoperates with existing upper- level protocols (ULPs).
To achieve the very
high transfer speeds of up to 1 Gbps and to avoid a time-consuming protocol management in software, Fibre Channel's data flow control needs to be laid out in hardware. The serial standard transfer rate of 1 Gbps is 10 to 250 times higher than that of other connection technologies. R&D groups are investigating even higher transfer speeds (2 and 4 Gbps). At the same time Fibre Channel technology has a lower latency and allows for cable lengths of up to 10 kilometers. The high transfer speeds are an essential advantage for RAID applications that need fast and parallel access to multiple mass storage devices.
Though the word suggests it, Fibre Channel technology is not limited to fiber- optic media. (Different spellings help distinguish between the connection technology and the transfer media: fiber for the former, fibre for the latter.)
Besides the transfer speed and protocol characteristics, topology plays an important role simply because it determines the number of simultaneous connections as well
as the available bandwidth. The Fibre Channel supports three connection topologies with different complexities: Point-to-Point, Arbitrated Loop, and Cross-Switched Fabric. The three topologies can be combined, providing for easier upgrades of existing systems.
To connect two devices or to incorporate a single system into an existing configuration, the simplest topology is a Point-to-Point connection. In contrast to SCSI, Fibre Channel Point-to-Point configurations have no termination problems.
The Arbitrated Loop (FC-AL) topology is capable of connecting up to 126 participants in a closed loop. The address of an FC-AL device can be set by hardware, software, or automatically by its position in the loop. The system then transfers data and control commands simultaneously on both parts of the loop.
The Cross-Switched Fabric (FC-XS) topology is based on a switched matrix architecture similar to a telephone network. It is a connection-oriented architecture that allows several participants to utilize si
multaneously the full bandwidth. In addition, it establishes a network that automatically bypasses failed nodes, thus providing higher reliability. The 24-bit addressing system of a fabric supports more than 16 million network nodes, so address space is not a problem.
The hardware and the Fibre Channel protocol handle the routing of the data packets as well as transfer errors, which eliminates extensive station management at the nodes. It is very easy to connect FC-XS peripheral devices to LANs.
A Class of Its Own
The Fibre Channel supports five different service classes either unidirectionally or bidirectionally. All devices are allowed to define the service classes, which isn't the case with other protocols. In addition, it is possible to use different service classes simultaneously.
Class 1 establishes a dedicated circuit-switched connection, granting full bandwidth for all participants. This service class guarantees high data transfer rates in combination with low latency,
but it uses the connections exclusively and is not able to free up bandwidth for other connections. This method is useful if large amounts of data have to be transmitted from one point to another very quickly.
Class 2 provides connectionless data transfer with verification and acknowledgment services. It routes data packets individually, as in an Ethernet or Token Ring LAN, and lets packets share the same bandwidth.
Class 3 services are similar to class 2, but they have no means of packet verification and acknowledgment. Because of its lower delays, class 3 is particularly suitable for audio and video applications. This service class also applies if ULPs such as SCSI perform data verification. It is also possible to "intermix," using temporarily available bandwidth of a class 1 connection for class 2 and 3 data packets transmission, improving the overall net throughput.
Class 4 provides a simultaneous exchange of data packets between two participants in a network by utilizing several routes so th
at bandwidth is guaranteed. This service class is particularly beneficial in a fabric where the load changes continuously but connections must have a minimum data transfer rate. In this case, the remaining bandwidth can be used by other applications.
Class 5 service allows for simultaneous (isochronous) data transfer to several participants and is especially applicable for audio and video servers in broadcast mode.
Except for class 1, all classes require a fabric network. Class 4 and 5 specifications are not yet fully fleshed out.
A Layered Architecture
Encapsulation of the existing transfer protocols, the so-called ULPs, and support of the different physical transfer media made it necessary to separate the Fibre Channel protocol into five layers, FC-0 to FC-4.
FC-4 links the Fibre Channel architecture to the ULPs and maps ULP data packets to a data buffer. The FC-2 layer then assigns these sequences to FC transfer packets (frames). This way, different ULPs can be transfer
red simultaneously.
FC-4 specifications have been finalized for the peripheral protocols HIPPI, IPI-3, SBCCS, and SCSI and the networking protocols IP, IEEE 802 (Ethernet and Token Ring), and ATM. The number of supported ULPs will grow in the near future and will soon include memory buses like VME or Futurebus+.
The frames assembled by FC-2 functions consist of an obligatory administration part and a variable-size payload area of 0 to 2112 bytes (
see the figure
). The frame headers contain routing and transfer information. The 4-byte cyclic redundancy check (CRC) code at the end of each frame helps detect transmission errors.
Another substantial part of the FC-2 layer functions is data-flow control for both fabric- (class 1) and frame-controlled (class 2) data transfers. For class 2 services, the FC-2 layer reassembles the original transfer sequence, even if the frames are routed through different connections and arrive out of order. The FC-2 protocol also is responsible for a
cknowledgment and automatic reconnection of transfers. The FC-3 layer has not yet been specified completely. It will define common service function for groups of networking participants (hunt groups).
FC-1 controls and synchronizes serial data transfer. It uses an 8-/10-bit encoding scheme that combines 8 data bits with 2 additional bits so that the resulting 10 bits have, on average, as many high bits as low bits (50 percent signal balance). This offset-free transmission assures a high signal quality. Even if fast transfer rates are combined with long distances, it guarantees a low bit-error rate. The trade-off is, of course, that 25 percent more data has to be transferred, which necessitates higher transfer rates. To avoid confusion about the real transfer rates, the Fibre Channel community uses MBps for the net transfer rate and Mbps for the overall transfer rate. Thus, the Fibre Channel's standard net transfer rate of 100 MBps is equivalent to a total of 1 Gbps.
The lowest protocol layer, FC-0, is
the physical link module. It defines the physical interface characteristics, including connectors, cables, speed, transmitters, and receivers. The Fibre Channel supports single-mode and multimode fiber optics and coaxial and twisted-pair cables.
FC-AL Architecture
Almost all the products coming up this year will use the FC-AL architecture for two reasons: The specification of the more complex fabric will not happen soon, and the gigabaud link modules (GLMs) that couple different FC-AL transmission media are still very expensive. These developmental limitations restrict Fibre Channel technology for now to high-performance connections between peripherals.
For the same reasons, mass storage systems in arbitrated loop topology that employ the SCSI protocol will play a predominant role in the near term. For larger RAID systems, the use of multiport devices in parallel loops will prevail.
Parallel FC-AL loops enhance redundancy as well as performance. If a RAID controller and a har
d disk array connect in two parallel loops, two independent paths are available for data transfer, almost doubling data transfer rates.
Fibre Channel has significant advantages such as high reliability, fault tolerance, real-time transmission, and flexibility of services. The goal of the Fibre Channel community is a network of corporate servers, workstations, and peripherals connected in one large chain.
illustration_link (22 Kbytes)
Fibre Channel frames feature variable payload lengths.
Otto Lehner is a software engineer and firmware specialist with ICP Vortex Computersysteme (Flein, Germany). You can reach
him by sending e-mail to
lehner@vortex.de
.
In the Line of Fibre
