The HIPPI-6400-PH ANSI standard, once known as Super-HIPPI, is commercially available under the Gigabyte System Network (GSN) name, and has stepped up even higher to address these same I/O bandwidth issues. With a raw throughput capability of 6,400,000,000 bits per second, the GSN technology is the first gigabyte-per-second networking to provide host-to-host communications that are significantly greater than all other LAN technologies in use today. GSN allows for more than six times the bandwidth capacity of Gigabit Ethernet, Fibre Channel, or HIPPI 800.
In short, GSN is the highest bandwidth and lowest latency interconnect standard, providing full duplex 6400 megabit (800 megabyte) per second of error-free, flow controlled data.
The first public GSN Demonstration was held at CERN on October 13, 1998. The first American demonstrations of these products were held in Orlando, Florida, November 9-12, 1998 at the High Performance Networking Computing show otherwise known as SC '98.
HIPPI is an abbreviation for "High Performance Parallel Interface" and is an ANSI standard. Serial HIPPI is the fiber-optic version of the High Performance Parallel Interface, originally developed in the late 1980s to serve the high-bandwidth needs of supercomputers and high-end workstations. A full line of network switches, gateways and network interface cards are available today from a number of vendors all of which are members of the HNF. HIPPI employs pairs of point-to-point, 800-megabit-per-second, simplex links (one incoming and one outgoing) that ensure full bandwidth to each station. By way of comparison: a true gigabit network provides connections between computer-system components-processors, shared storage, etc., at a signaling rate of at least 1000 megabits per second (Mbps). HIPPI is approximately ten times the speed of both FDDI and Fast Ethernet.
The GSN technology is by far the fastest networking technology available today. Recognizing that a gigabyte equals 8 billion bits whereas a gigabit of information equals only 1 billion bits, consider that GSN's duplex connection is capable of 6.4 gigabits per second in each direction simultaneously. That is, 12.8 gigabits per second per connection. This speed represents a flow-controlled, error-corrected data rate limited only by adapter hardware or DMA engine bandwidth. Beginning with a raw physical layer signaling rate of 10 gigabits per second, the final data throughput at 6.4 Gbps is calculated after subtracting a periodic link retraining sequence to keep the clocks synchronized to data lines, 1.6 Gbps used for addressing, signalling and error control functions, and 2 Gbps utilized for encoding.
GSN, which utilizes 32-byte micropackets and virtual channels, provides a low-latency structure for short control message communications and efficient transmission of messages up to 4 gigabytes (GB) in length. The technology allows network traffic to be segregated by throughput requirements, ensuring that large transmissions, such as digital film, do not delay the communication of short control messages and IP traffic, which reside on separate virtual channels.
The current GSN standard allows for copper connection cables in a star configuration up to 40 meters, 20 bits wide at 500 MHz. A fiber standard is currently defined at 1 Km with 10 bits at 1 GHz. In the short term, repeaters (back to back SuMACs) will allow additional sections of 40m cable to be strung together. In the longer term, the HNF foresees that optical transceivers will allow for 200m (18 months), and then 1Km connections (30 months).
GSN offers a significant technology enhancement in the problem area of network latency. While large data packets generally are a benefit, particularly when transferring huge files, occasionally the transmission of very large data packets can cause a problem because of the "single-mindedness" of the network connection. For example, a port may be unavailable until the packet transfer is complete, thus blocking and backlogging other transfers to the same port. GSN's ability to multiplex multiple traffic types at the physical layer can provide extremely low latencies for high priority messages.
High bandwidth technologies will be crucial for users of processor-intensive data or low latency applications. More specifically, GSN will be used in organizations that require timely movement of large amounts of information including such scientific and technical computing as data mining, transaction processing, film post production, system area networks, satellite imagery, seismic modeling, HDTV, video and film archiving, and storage management.
As with HIPPI, GSN will first be adopted in supercomputer centers and other very data-intensive environments where the number of processors operating in parallel (at higher and higher speeds) continues to grow. The quantities of data being manipulated are staggering, and the existing network connections are straining under the load.
Similarly, a variety of business concerns that depend on timely movement of large amounts of information are eagerly awaiting this next-generation technology. As an example of this, note the entertainment industry which relies more and more on digital tools, particularly in special effects and postproduction for feature films. GSN allows these users the ability to transmit uncompressed digital film images between network nodes in real time.
The list of organizations that have implemented HIPPI networks is a veritable "who's who" from the public and private sectors worldwide, including Los Alamos National Laboratories, CERN, NASA, Disney, Ford Motor Co., Boeing, and TRW.
GSN technologies are currently in use at CERN and the Los Alamos National Laboratories. HNF member customers for GSN, presently working with slower less-advanced network technologies, have expressed great excitement at the prospect of these forward-looking technologies.
GSN and HIPPI products are shipping today. Initial GSN products such as network interface cards (NICs), switches, bridges, and hubs are being delivered to early adopters whose program requirements dictate the bandwidth and performance requirements of GSN that cannot be met with other technology.
There are many GSN and HIPPI products available today and coming available shortly.
GSN's commercially available products include ODS Network, Inc.'s GSN 6400, a non-blocking, 32 port switch, and Silicon Graphic, Inc.'s SuMAC™ GSN MAC controller. The Silicon Graphics GSN host adapters will be available in 1999. GENROCO, Inc. expects to be shipping its line of GSN products by year's end.
For detailed information on vendors of GSN and HIPPI technology, visit the HNF Products web page.
All GSN products currently under development are using the Silicon Graphics SuMAC™ GSN controller to implement the ANSI physical layer protocol - HIPPI-6400-PH. Silicon Graphics has a licensing program for SuMAC™, and has licensed it to a number of companies including Compaq, IBM, ODS, DaVinci, Genroco and PMR.
GSN and HIPPI products are competitively priced with other networking technologies in terms of cost per throughput. As GSN becomes more widely adopted, product costs will decrease with major price reductions expected by mid-1999.
The unique nature of this technology is its ability to bridge with any existing 802.x IEEE ethernet standard, making GSN the ideal bridging technology for high performance networks. With its superior ability to manage a broad range of message sizes, GSN is used as a high-speed LAN backbone. No other high-bandwidth technology is capable of aggregating and connecting the various existing and emerging gigabit LANs. The proposed ANSI standard provides for interoperability with Ethernet, Fibre Channel, ATM, HIPPI-800 and other standards.
The GSN throughput outperforms all other standardized networking technologies with at least six times the bandwidth. Fibre Channel supports 1.06Gbps today but won't approach GSN's 1GB per second for many years. Gigabit Ethernet is similarly bandwidth-limited.
ATM can surpass its commercial limit of 622Mbps, experimentally reaching 2.488Gbps, but it lacks flow control and retransmission. SCI (Scaleable Coherent Interconnect) technology runs at 4Gbps but does not feature retransmission, and most implementations are not interoperable.
Along with its very high throughput capacity, GSN features 4 Virtual Circuits (for bandwidth prioritization) and support for seamless bridging to all existing 802.x and TCP/IP technologies. This makes GSN the fastest, most efficient technology available for the high performance computer market built upon the same conceptual framework of commodity-based ethernet.
Additionally, GSN adapters are designed for a new upper layer Operating System bypass called Scheduled Transfer that allows for packets to bypass the network kernel and perform memory to memory data transfers. Scheduled Transfer allows for the direct DMA communication without involving the host operating system. This OS bypass allows the receiving node to set up memory buffers before the transfer begins, with buffer alignment being the responsibility of the sending node (which is where you want it to be). Virtual connections set up validated ports which then can be used for multiple ST operations. ST is a mechanism for avoiding one of the hardest to solve problems that all network media face which is the issue of congestion of network resources. By insuring that resources are available in the destination for storing incoming data, ST reduces the potential of congestion ever occurring.
Members of the HNF, GENROCO and SGI, have developed a method using encapsulation of SCSI over the new upper layer operating system bypass called Scheduled Transfer (ST) to do GSN storage. The ST model allows for the production of GSN disk arrays and gateways to Fibre Channel storage. It also allows access over ST from Gigabit Ethernet, ATM OC12, and HIPPI. The HNF expects these routes are likely to become even more accepted than IPI-3 over HIPPI. SCSI over ST is currently being formulated as an ANSI standard. (*For more information please see Dr. Steph Bailey's (GENROCO) white paper on this concept.)
Fibre Channel has a current bandwidth of 100 MB/s. It doesn't take a very large network of high performance servers interconnected to require a faster storage fabric. At eight times the speed of Fibre Channel, and with 32 port non-blocking switches, GSN is very atrractive for high performance storage area networks.
In the short term, using the above mentioned storage architecture technique, TBF-864 users can put eight 100 megabyte per second PCI Fibre Channel cards in separate PCI buses and stripe 800 megabyte GSN data over the bridge. In the mid term (7-9 months) users can utilize PCI buses to provide an 800 megabyte per second source or destination with 200 megabytes per second of back channel. In the long term (30 months) next generation PCI, PCI-X, will be as fast as GSN and will allow for such high-bandwidth throughput.
The Physical Layer standard, implemented by the Silicon Graphics SuMAC GSN Controller, was approved by the ANSI T11 Task Group at the August 1998 plenary meeting. It became an approved ANSI standard in December of 1998. Additional work is being completed on the Scheduled Transfer (ST) Protocol and ST API Mappings in the T11.1 technical committee.
SuMAC powered on in December of 1997, and prototypes were available shortly thereafter. Since June of 1998, production quality parts have been available in volume.
Email us!
Last Updated: 21-Jan-1999 ddw
(c)1996 HIPPI Networking Forum