INTRODUCTION
Network managers on the hunt for high throughput don't have time to play games. Still, they may want to spare a moment to consider this brief riddle: "What do supercomputers, workstations, and other high-end devices have in common?"
"Bandwidth--and lots of it," is the obvious answer. But it's not the only one. How about HIPPI? That's right, HIPPI (High-Performance Parallel Interface). True, the gigabit-per-second ANSI standard was originally developed to allow mainframes and their supercomputing kin to communicate with one another, and with directly attached storage devices, at supersonic speeds. But HIPPI isn't just for supercomputers anymore.
HIPPI’s first aim is to allow high-speed data transfer between different kinds of data-processing devices. It was especially optimized for large block transfers - thereby allowing high peak rates for supercomputers writing on a (parallel) filesystem, but also for interconnection of nodes in a missive parallel system with high volume data transfer between memory.
The HIPPI is a simplex channel, capable of transferring data in one direction only. Two HIPPI channels may be used to implement a full-duplex channel. The HIPPI is a point-to-point channel that does not support multi-drop. The point-to-point limitation considerably simplified the electrical and protocol aspects of the HIPPI. Crossbar switches and other networking methods are being considered to achieve the equivalent of multi-drop. An addressing mechanism is included to support these networking concepts.
Remember that HIPPI is simplex - and point-to-point. The simplex problem was solved simple - another cable will do it. But HIPPI has also facilities that arise the need to have a closer look on HIPPI managing a network of high-speed data processing equipment. Parts of the protocol enable HIPPI to drive special devices - so called HIPPI-Switches. By this means, HIPPI gains the ability to initiate a connection between two dedicated of several devices.
At present HIPPI is the only well-developed technology for gigabit-per-second networks, and as an ANSI standard it will continue to be useful for at least the next five or ten years. When the use of fiber-optic networks becomes widespread, HIPPI will probably serve as a local-network backbone that connects to long-distance links for special applications. Los Alamos has had a prominent and crucial role in the development of HIPPI and continues to refine its capabilities.
HISTORY
In the late 1980s, Los Alamos National Laboratory researchers looked at the links in their local-area network--specifically, the connections between supercomputers, file storage, and other devices inside the machine room--and saw potential bottlenecks. The network had a maximum data-transmission rate of 50 Mbit/s, one of the highest in the country, but numerical simulations on the fastest supercomputers were generating increasingly large data files. The time would come when simply moving the files would take almost as much time as generating them. Visualizing the data presented another bottleneck. For example, a researcher might want to simulate the formation and merging of eddies in a turbulent fluid and then view the simulation at a rate that the eye interprets as smooth, continuous motion--about 30 frames per second. To show 30 frames each second on a high-quality, 24-bit color monitor with a 1024-by-1024 pixel display requires a data-transmission rate of about 750 Mbit/s--15 times higher than what the network could support in 1987.The solution to that problem required more than just a system for high-speed data transmission. In 1987 there was already a sufficiently high-speed communication link--the proprietary Cray HSX interface, which operated at approximately 850 Mbit/s. However, Los Alamos (and other supercomputing facilities) wanted to be able to transmit data freely among computers, monitors, storage devices, and so forth. A data-transmission system for that purpose would have to meet two requirements that the HSX interface did not meet. First, the equipment for the system had to be nationally standardized so that any manufacturer could give its devices high-speed ports compatible with the new data link. Second, since connecting every device to every other device with a separate cable was impractical, the new system had to be a network in which switches would connect any device to any other as needed. Therefore Los Alamos researchers decided to develop and standardize a new generation of networks that would transmit data at about a gigabit (one billion bits) per second.
National standards are developed by committees of volunteers under the auspices of standards organizations such as the American National Standards Institute (ANSI). In 1987 Los Alamos personnel presented their plans for a gigabit-channel standard to an ANSI group working on high-speed data transmission. That committee had been considering a more modest standard, specifying transmission speeds of 100 to 200 Mbit/s, to be used in permanent connections between computers and data storage. The committee's reception of the Los Alamos proposal was decidedly cool--the members referred to advocates of gigabit speeds as the "lunatic fringe."
Two months later, proponents of gigabit networks from both Los Alamos and industry joined the committee. Within a year that committee produced the first draft of the standards for the new gigabit network, now called HIPPI, the High-Performance Parallel Interface. As soon as word got out that the new standard was on the way, vendors began to build components for the gigabit network. The first commercial products appeared in 1988: a HIPPI port for the IBM 3090 and a fiber-optic extender for HIPPI networks. In the same year, Los Alamos built a prototype of a HIPPI switch. By 1989, the first commercial HIPPI switches were marketed, and HIPPI ports on computers made by IBM, Cray Research, Thinking Machines Corporation, and others were being developed.
The ANSI committee developed the HIPPI standards rapidly by basing them as much as possible on off-the-shelf technology. In November 1991, three years after the initial draft was delivered to ANSI, HIPPI became the first national standard for gigabit-per-second data transmission. (More precisely, the HIPPI standard gives specifications for transmitting data at two speeds: either 800 Mbit/s--the minimum speed for a high-resolution, 30-frame-per-second movie--or twice that speed, 1.6 Gbit/s.) The HIPPI standard was adopted in a remarkably short time--a more typical time for acceptance of a new network standard is five to ten years. Today, HIPPI is the interface of choice for gigabit-per-second transmissions on most supercomputers and some high-performance workstations because it was the first standard at those speeds and because the technology is mature and stable.[1] For instance, Los Alamos now has one HIPPI network that is in regular use and another that serves as a testbed for HIPPI hardware and software development. Newer gigabit network standards based on new technologies are emerging, for example, Asynchronous Transfer Mode (ATM) and Fibre Channel, but such technologies are not yet readily available at gigabit speeds.
In 1991, as the official HIPPI standards were being completed, the National Science Foundation and the Advanced Research Projects Agency were beginning to build and evaluate wide-area gigabit networks. The Center for National Research Initiatives in Reston, Virginia, coordinated the establishment of five testbeds for gigabit-per-second communication and asked Los Alamos to join one of the projects, the Casa Gigabit Testbed. The other participants in the Casa testbed are the San Diego Supercomputer Center (SDSC) and the supercomputer centers at CalTech and NASA's Jet Propulsion Laboratory (JPL). The goal of the project was to construct a HIPPI-based network connecting the four supercomputer centers and to use the network to investigate the applicability of distributed computing--several computers working together on the same calculation--to large computational problems. The problems designated were global climate modeling, seismic modeling, and chemical-reaction dynamics.
When the project started, each of the supercomputer centers had a local HIPPI network. To send HIPPI data over distances of hundreds or thousands of miles, the Casa network used another technology, the Synchronous Optical Network, or SONET, which the telephone companies had developed and were installing across the country as the next generation of telecommunication transmission. SONET is a network of fiber-optic transmission lines designed to transmit data for long distances at various rates from 51 Mbit/s to 9.8 Gbit/s. Los Alamos engineers developed a HIPPI/SONET gateway that moves data over the SONET network at the HIPPI rate of 800 Mbit/s. The links between CalTech, JPL, and the SDSC were in place by August 1993; the Los Alamos link was brought up in June 1994 and connects a HIPPI switch at Los Alamos with a similar switch at the San Diego Computer Center. The project has now turned to distributed-computing research.
DEFINITIONS
Conventional:Used with respect to networks, this refers to Ethernet, FDDI and 802 LAN types, as distinct from HIPPI-SC LANs.
Destination:
The HIPPI implementation that receives data from a HIPPI Source.
Node:
An entity consisting of one HIPPI Source/Destination pair that is connected by parallel or serial HIPPI to a HIPPI-SC switch and that transmits and receives ARP and IP datagram. A node may be an Internet host, bridge, router or gateway. This memo uses the term node in place of the usual "host" to indicate that a host might be connected to the HIPPI LAN not directly, but through an external adaptor that does some of the protocol processing for the host.
Serial HIPPI:
An implementation of HIPPI in serial fashion on coaxial cable or optical fiber, informally standardized by implementor's agreement in the Spring of 1991.
Switch Address:
A value used as the address of a node on a HIPPI-SC network. It is transmitted in the I-field. HIPPI-SC switches may map Switch Addresses to physical port numbers.
Source:
The HIPPI implementation that generates data to send to a HIPPI Destination.
Universal LAN Address (ULA):
A 48 bit globally unique address, administered by the IEEE, assigned to each node on an Ethernet, FDDI, 802 network or HIPPI- SC LAN.
Equipment:
A HIPPI network can be composed of nodes with HIPPI interfaces, HIPPI cables or serial links, HIPPI-SC switches, gateways to other networks and, possibly, proprietary equipment that multicasts or responds to ARP requests on behalf of the real nodes.
Each HIPPI interconnection between a node and a switch shall consist of a pair of HIPPI links, one in each direction.
If a link between a node and the switch is capable of the 1600 Megabit/second data rate option (i.e. Cable B installed for 64 bit wide operation) in either direction, the node's HIPPI-PH implementation shall also be capable of 32 bit operation (Cable B data suppressed) and shall be able to select or deselect the 1600Mb/s data rate option at the establishment of each new connection.
GETTING HIP TO HIPPI
CHARACTERISTICS:• Speeds: 800 Mbit/s and 1.6 Gbit/s
• Cabling: 50-pair STP, single mode and multimode fiber
• Distance: 50 meters point to point over copper. Cascaded switches can be extended 200 meters over copper; 300 meters over multimode fiber; 10 kilometers over single-mode fiber
• Connection time: Less than 1 microsecond for dedicated connection
• Latency: 160 nanoseconds on average
• Products: Switches, routers, mainframe and supercomputer channels, interfaces for mass storage devices, frame buffers, adapters for PCI workstations and PCs
• Cost per switched port: $2,000
• Cost per adapter card: $2,000 to $18,000, depending on the system and its bus architecture.
• Very high-speed data transfers -
HIPPI can be configured for either of two speeds--800 Mbit/s or 1.6 Gbit/s either simplex or full duplex.
• Very simple signaling sequences -
Basically, HIPPI connections can be set up with just three messages: Request (which the source uses to ask for a connection), Connect (which the destination uses to indicate that the connection has been established), and Ready (which the destination uses to indicate that it's ready to accept a stream of packets).
• Protocol independence -
HIPPI channels can handle so-called raw HIPPI (data formatted with the technology's framing protocol, without any upper-layer protocols), TCP-IP datagram’s, and IPI-3 (Intelligent Peripheral Interface) framed data. (IPI-3 is the protocol used to connect peripherals like RAID [redundant array of inexpensive disks] devices to computers). Thus, HIPPI is equally adept at internetworking (with Ethernet and FDDI) and high-speed data storage and retrieval.
• Physical-layer flow control -
HIPPI also offers a credit-based system for reliable and efficient communications between devices operating at different speeds. In effect, the source keeps track of the Ready signals and only sends data when the destination can handle them.
• Connection-oriented circuit switching -
Non-blocking circuit switches allow multiple conversations to take place currently. Thus, the aggregate bandwidth of these switches can be very high--equal to 800 Mbit/s or 1.6 Gbit/s times the number of ports.
• Compatibility with copper and fiber -
HIPPI uses 50-pair STPs (shielded twisted pairs) for short distances. It works with single-mode and multimode across the campus or metropolitan area. SONET is used for long-distance communication. HIPPI makes it possible to link a variety of high-speed devices. Copper connections can extend 25 meters; HIPPI-Serial interfaces allow runs of 300 meters over multimode fiber. With fiber extenders, devices can be separated by 10 kilometers.
HIPPI COMPONENTS
At its very simplest, a HIPPI network consists of two computers with HIPPI channels linked by two 50-pair copper cables. This configuration furnishes a full-duplex (or dual-simplex) 800-Mbit/s channel that can extend up to 25 meters. But there's no need to stop there. HIPPI switches can be cascaded locally over 200-meter cabling runs. The number of switches that can be cascaded depends upon the size of the switches themselves. HIPPI addresses are 24 bits long. Some switches work with 3-bit addresses; others, with 4 bits. If eight-by-eight switches are deployed (which use 3-bit addresses on their eight inbound and eight outbound ports), a maximum of eight boxes can be cascaded (for a total of 24 bits).The physical elements of the HIPPI system are ports for the devices that are connected to the network, cables for the network links, and a switch to connect the devices as desired. Figure 1 shows a representative HIPPI local-area network. (The HIPPI Tester, the Frame Buffer, and the disk array are described below.)
The figure shows a HIPPI crossbar switch connecting supercomputers, a high-performance disk array for file storage, a frame buffer, and a HIPPI Tester (to monitor performance of the network). A fiber-optic extension connects this small network to another HIPPI switch, and a HIPPI/SONET gateway connects the local HIPPI network to a wide-area SONET network. HIPPI cables are sets of 50 twisted-pair copper lines (for 800 Mbit/s) or 100 lines (for 1.6 Gbit/s). The data transmission is parallel; that is, many bits are sent simultaneously, one bit per twisted-pair line. Hence, HIPPI cables are similar to the buses, or ribbons of wires, connecting parallel ports on personal computers. HIPPI ports contain the circuitry that organizes and interprets the parallel transmissions, sending 32 bits of data on the 32 data lines and other necessary signals on the other lines. An 800 Mbit/s HIPPI port uses a 25 MHz clock to synchronize transmissions, so a 32-bit word is sent every 40 nanoseconds. If a device cannot supply a continuous stream of data, HIPPI allows idle time between groups of words.
HIPPI was designed to be a "machine-room" network; the standards specify copper cables extending 25 meters or less. Vendors are now able to supply fiber-optic cables (governed by an implementers' agreement, not a HIPPI standard) that can extend HIPPI links to a maximum distance of 10 kilometers so that HIPPI networks can cover wider areas. Even longer distances can be spanned by using ATM or SONET links from the telecommunications carriers.
Crossbar switches are used in HIPPI networks. The switches have multiple inputs and outputs and a separate connection path between every input and every output so that many simultaneous connections can be made.
As the standards were being developed, Los Alamos engineers were also developing equipment to help test and demonstrate HIPPI. The HIPPI Tester, for example, was developed to help ensure compatibility as well as to measure the performance of HIPPI ports developed by various computer manufacturers. The HIPPI Tester, a briefcase-sized unit containing a lap-top computer and a special version of a HIPPI port, receives test signals from computers and checks to see if they were transmitted correctly; it also generates and transmits signals so that the received versions can be checked. Since the HIPPI Tester is portable, Los Alamos personnel were able to travel around the country to test the HIPPI ports being developed for various supercomputers. By 1990 the technology for the HIPPI Tester had been transferred to private industry. The Tester is easy to use, and now that it is manufactured in quantity, manufacturers and system designers can test their equipment themselves.
Another device developed at Los Alamos was the HIPPI Frame Buffer, a device for the high-speed display of computer graphics--which was the original motive for the development of HIPPI. A frame buffer consists of a large memory split into two buffers that each stores a complete frame of display data. One buffer receives new data from a computer while the other buffer sends data to the display. When the "screen" buffer is empty, the roles of the two buffers are reversed. As long as the frame buffer contains data, the movie is not interrupted. The HIPPI Frame Buffer was designed to be connected both to a supercomputer and to a high-resolution color monitor. The technology for this system was transferred to private industry in 1992.
As HIPPI networks were being tested, researchers realized that the bottlenecks in the networks were the supercomputers, not the HIPPI links. For example, a 30-second HIPPI frame-buffer movie requires about 24 billion bits of data. On the one hand, a typical Cray supercomputer could supply data fast enough, but it had a memory capacity of only a few billion bits--less than what even a short movie requires. On the other hand, the new massively parallel computers had memory capacity several times the Cray capacity, but were inefficient at protocol processing--packaging data for transmission and checking the data for errors. For example, when the massively parallel CM-2 was doing the protocol processing, it could supply data to a HIPPI port at only about 80 Mbit/s or less.
In the last several years two new HIPPI components, the high-performance disk system (HPDS) and the crossbar interface (CBI), have been developed to solve these problems. An HPDS is an array of high-speed disks connected directly to a HIPPI port and controlled by a workstation. The array can hold more data and supply data faster than the memory of a supercomputer--a typical disk array has a capacity of tens of billions of bits and can transfer data at about 500 Mbit/s. Data are transferred from a supercomputer or file storage to the disk array and then sent through the HIPPI network for analysis or display.
The CBI can boost the rate at which a massively parallel computer supplies data to a network. A CBI, which is a small, special-purpose computer, was originally developed to manage HIPPI networks. However, a CBI can also perform protocol processing. If data are sent out "raw" from the HIPPI port on the CM-2 and the TCP/IP protocol processing is done in a CBI, the effective transmission rate is about 400 Mbit/s--a five-fold increase.
THE SONET SOLUTION:
HIPPI takes advantage of SONET to extend over long distances. Of course, that means that until a national SONET infrastructure is in place, most HIPPI applications will span metropolitan areas.
When HIPPI and SONET are used in conjunction, the HIPPI network is effectively terminated at a HlPPI-SONET gateway, which frames HIPPI data for transport over SONET. This can be done fairly easily, since HIPPI maps well both to a single SONET OC12c (622-Mbit/s) circuit or multiple SONET OC3 circuits. In the latter scenario, the gateway stripes data onto the 155-Mbit/s circuits. This scheme has been demonstrated successfully for a sustained throughput of 783 Mbit/s over 2,000 km. A HIPPI-SONET gateway has even run successfully over a satellite link.
HIPPI INTERFACE
In relation to the Open Systems Interconnection (OSI) Basic Reference Model, the High-Performance Parallel Interface (HIPPI) covers the physical layer and a small portion of the data-link layer. The HIPPI uses a parallel data path with copper cable. The 800 Mbit/s HIPPI uses a 32-bit data bus, and the 1600 Mbit/s version of the HIPPI uses a 64-bit data bus. The major emphasis has been on the development of the 800 Mbit/s version and as such the RIO HIPPI Source is also a 800 Mbit/s implementation of a HIPPI interface.
HIPPI INTERFACE SIGNALS:
The HIPPI-PH specifies mechanical, electrical and signaling parameters of the HIPPI over 25 meters cable of the twisted pares caring differential ECL level signals. The HIPPI is unidirectional, synchronous channel with fixed clock frequency of 25 MHz. The theoretical maximum throughput of the HIPPI is 100 Mbyte/sec for 32 bit and 200 Mbyte/sec for 64 bit version. Separate links can be interconnected through HIPPI switches in HIPPI networks having a very high throughput. The HIPPI switches are commercially available from NCS, Essential Communications and Gigabit Labs.
The HIPPI is a point to point link from a data transmitting device "HIPPI SOURCE" to a data receiving device "HIPPI DESTINATION”. The interface signals are illustrated in Figure 3. The numbers in parentheses indicate the number of signal lines when using the 1600 Mbit/s option. The other numbers indicate the number of signal lines when using the 800 Mbit/s option. All signals, except for the INTERCONNECT signals, use differential emitter- coupled logic (ECL) drivers and receivers. The INTERCONNECT signals use single-ended ECL drivers and receivers.
The protocol is based on the relationship between the "REQUEST" and "CONNECT" and a look-ahead data flow control by the "READY" pulses. The physical framing hierarchy is: HIPPI connection, HIPPI packet and HIPPI burst - Figure 6.2.
The HIPPI connection consists of one or several packets. Upon requesting a connection the source asserts the I-Field, a 32 bit word, used as routing information when the connection is established through one or more HIPPI switches.
A connection is made in a fashion similar to the connection made when dialling the telephone. Once a connection is established, a packet (or multiple packets) can be sent from the source to the destination. Each packet contains zero or more bursts, and each burst contains one to 256 words. Bursts that contain less than 256 words may only occur as the first or last burst of a packet. Words are composed of 32 or 64 bits. The amount of wait time between packets and bursts may vary. Maximum wait times depend on the data flow to or from the upper-layer protocols and on the data flow to or from the opposite end of the channel.
The HIPPI burst is a train of up to 256 words of 32 or 64 data bits each and 4 or 8 byte parity bits. Every burst is followed by a length-longitudinal redundancy checkword (LLRC) which provides a proper data integrity for eventual error detection and correction.
Most of the DAQs for the HEP experiments need the movement of data only in one direction - from previous to the next DAQ level. In such case the use of a simplex HIPPI segment is relevant. The first application of the HIPPI based computer network is the second level of the DAQ for NA48 experiment at CERN [3]. The second level of the NA48 DAQ is an array of powerful workstations receiving data from an event builder - Figure 6.3.
The HIPPI switch is used as a distributor, forwarding large data blocks (approximately 180 MBytes) to the currently available workstation while other workstations are processing data, already transferred to them.
HIPPI SERIAL COMMUNICATION:
As indicated, HlPPI-Serial allows connections over fiber to 10 kilometers, so switches can be linked over a HIPPI backbone (see Figure 6.4). If the switches are not equipped with serial interfaces, then HIPPI fiber extenders are required for long-distance links.
HIPPI makes it possible to link a variety of high-speed devices. Copper connections can extend 25 meters; HIPPI-Serial interfaces allow runs of 300 meters over multimode fiber. With fiber extenders and singlemode fiber, devices can be separated by 10 kilometers.
A real sense of the power of HIPPI as backbone technology can be gotten from a demonstration at last fall's Supercomputing '94 conference held in Washington, D.C. An all-fiber HIPPI backbone, consisting of some 18 miles of multimode cable, connected 16 exhibitors and delivered over 90 Gbit/s.
The backbone was used for numerous tests and demonstrations. For example, three HIPPI switch makers--Avaika Networks Corp. (Mountain View, Calif.), Essential, and NetStar--conducted the first public interoperability demo of HIPPI-Serial. The switches were linked directly via their serial ports--without fiber extenders--and ran at the full HIPPI speed of 800 Mbit/s. Direct switch-to-switch fiber connections will go a long way toward simplifying HIPPI networking.
HIPPI can be used to build a local area network (LAN) with various configurations. A simple network, configured as "star" has a HIPPI switch and number of network stations connected to the switch - Figure 6.5.
LAN applications requiring a full duplex use two HIPPI channels in both directions. The total throughput of the network is a sum of the particular throughput of all network segments.
HIPPI AND WIDE-AREA NETWORKS
In 1991, as the official HIPPI standards were being completed, the National Science Foundation and the Advanced Research Projects Agency were beginning to build and evaluate wide-area gigabit networks. The Center for National Research Initiatives in Reston, Virginia, coordinated the establishment of five testbeds for gigabit-per-second communication and asked Los Alamos to join one of the projects, the Casa Gigabit Testbed. The other participants in the Casa testbed are the San Diego Supercomputer Center (SDSC) and the supercomputer centers at CalTech and NASA's Jet Propulsion Laboratory (JPL). The goal of the project was to construct a HIPPI-based network connecting the four supercomputer centers and to use the network to investigate the applicability of distributed computing--several computers working together on the same calculation--to large computational problems. The problems designated were global climate modeling, seismic modeling, and chemical-reaction dynamics.
When the project started, each of the supercomputer centers had a local HIPPI network. To send HIPPI data over distances of hundreds or thousands of miles, the Casa network used another technology, the Synchronous Optical Network, or SONET, which the telephone companies had developed and were installing across the country as the next generation of telecommunication transmission. SONET is a network of fiber-optic transmission lines designed to transmit data for long distances at various rates from 51 Mbit/s to 9.8 Gbit/s. Los Alamos engineers developed a HIPPI/SONET gateway that moves data over the SONET network at the HIPPI rate of 800 Mbit/s. The links between CalTech, JPL, and the SDSC were in place by August 1993; the Los Alamos link was brought up in June 1994 and connects a HIPPI switch at Los Alamos with a similar switch at the San Diego Computer Center. The project has now turned to distributed-computing research.
At present HIPPI is the only well-developed technology for gigabit-per-second networks, and as an ANSI standard it will continue to be useful for at least the next five or ten years. When the use of fiber-optic networks becomes widespread, HIPPI will probably serve as a local-network backbone that connects to long-distance links for special applications. Los Alamos has had a prominent and crucial role in the development of HIPPI and continues to refine its capabilities.
COMPARISIONS
HIPPI vs. ATM:
For all its strengths, HIPPI is hardly standing still. Work is under way to improve the technology and integrate it with other networking schemes. For instance, a HIPPI MIB proposal is being prepared for SNMP. It features self-discovery of switch addresses and address resolution between MAC (media access control) addresses and HIPPI addresses. Developers also are working on performance and capacity planning applications and exploring ways to integrate HIPPI network management with OpenView, NetView, and other platforms. HIPPI interoperability with ATM, Fibre Channel, and SONET also is garnering a great deal of interest.
Because of ATM's strong momentum, some in the industry assume that it will displace HIPPI. But such a scenario isn't likely--if for no other reason than simple arithmetic. Any of today's ATM switches can deliver an aggregate bandwidth of 3.2 Gbit/s. That's fast, but it pales in comparison to the 12.8 to 25.6 Gbit/s offered by HIPPI switches.
In addition, the ATM market is primarily focused on services at 155 Mbit/s and below, with particular interest in 25 Mbit/s. That may be what many users want, but it cannot satisfy the significant number of users who need gigabit throughput right now. Finally, HIPPI--unlike ATM--allows storage devices to be directly connected to a network at high speeds. That's a critical concern, since end-users on high-speed networks must be able to access compute and storage servers.
Actually, it's not a question of HIPPI or ATM. A gateway from a HIPPI workstation cluster to ATM would give end-users the best of both worlds: the unsurpassed throughput of HIPPI in the local area and the ATM's wide-area connectivity. With such a gateway, even multimedia data could fan out from HlPPI-attached servers to multiple ATM desktops.
Sounds great, but creating such a gateway is hardly a trivial undertaking. HIPPI is a connection-oriented, circuit-switched transport. ATM is a connection-oriented, cell-switched scheme. HIPPI can handle raw HIPPI, as well as IP and IPI-3 datagrams. ATM slices and dices everything into 53-byte cells. For all the complexity, though, the ANSI committee and the HIPPI Networking Forum are hammering out HIPPI-ATM interfaces. HIPPI-ATM connections work by encapsulating HIPPI data, shipping it over the ATM network, and then reconstituting the HIPPI data and format at the other end. In other words, HIPPI is tunneled through the ATM network. This is done at AAL5 (ATM adaptation layer 5). NetStar is beta-testing a version of its GigaRouter with a HIPPI-ATM interface based on this draft standard. GTE's HIPPI-ATM switch uses an interface based on an earlier version of some of this work.
HIPPI vs. FIBRE CHANNEL:
Ironically enough, HIPPI and Fibre Channel are both the work of the same ANSI committee, X3T11. Unlike HIPPI, however, Fibre Channel is marked by complexity. It specifies four data rates, three kinds of media, four transmitter types, three distance categories, three classes of service, and three possible fabrics. The idea was to make sure that Fibre Channel was very rich in features, an ambitious and admirable goal. Unfortunately, it also has meant that Fibre Channel products have been slow to make it to market. And those that have emerged thus far generally run at 200 Mbit/s (266 Mbit/s signaling speed), a quarter of the standard's highest rated speed.
Despite their differences, HIPPI and Fibre Channel can be complementary technologies. An ANSI standard has already been defined that specifies how to send upper-layer Fibre Channel protocols over the lower-layer HIPPI media. The complementary ANSI standard (defining how to ship HIPPI upper-layer protocol over Fibre Channel lower-layer media) is in process. IP-level routing offers another way to map HIPPI to Fibre Channel, and vice versa. At the High-Performance Computing and Networking Conference in Milan, Italy, both HIPPI and Fibre Channel were deployed on the backbone network used for the Technology Demonstration Display.
ADVANTAGES & SHORTCOMMINGS
ADVANTAGES:
HIPPI is the current interface of choice, largely because it was the first standard at close to the gigabit speed. It came to fruition quickly because of a "keep it simple" goal, and a well-focused direction in the standards committee that avoided adding lots of bells and whistles. Some of the advantages of HIPPI include:
• It is simple, elegant, and easy to understand.
• It has a good physical level flow control. The flow control even works with very long links by the addition of extra buffering at the receivers (approximately 1 kilobyte per kilometer of distance).
• A good tester was developed early on which allowed vendors to test implementations in-house so that interconnection with other vendor's equipment was usually a plug-and-play.
• A variety of products with HIPPI interfaces from a fair number of vendors currently exist. Many are second generation designs, incorporating improvements from earlier designs.
• HIPPI crossbar switches are available from multiple vendors.
• HIPPI specific integrated circuits are available. Even so, some vendors find that small scale integration parts are more suitable due to the simplicity of the physical interface and limitations of the HIPPI specific ICs.
• Serial-HIPPI is being built into native switch and workstation interfaces, increasing the distance and improving the physical properties.
• HIPPI to SONET adapters are available for very long distance links using telephone network facilities.
SHORTCOMMINGS:
HIPPI is not without limitations and shortcomings. Perceived shortcomings include:
• It is not a mass-market item, the number of applications that require the bandwidths are not that numerous. Hence the price is higher. It is questionable whether competing gigabit/s technologies, e.g., fiber Channel or ATM would be any cheaper.
• It does not support speeds slower than 800 Mbit/s. slower speeds would help make it more of a mass market item.
• It does not support multiplexing. If you transfer a megabyte over a HIPPI channel as a single entity then it will take at least 10 milliseconds. During this time the channel cannot be used for any other communications.
• HIPPI does not support time-critical or isochronous data.
• Error detection, but not error correction, is provided by the HIPPI.
• The HIPPI specification limits the distance to 25 meters (82 feet) with copper twisted-pair cable. Serial-HIPPI defines a fiber-optic extender that is useful for distances up to 10 kilometers, but it is an added expense.
• The cable is somewhat bulky and stiff.
• The cable connector is large and somewhat fragile.
APPLICATIONS
• HIPPI switches are now an accepted technology for LAN interconnects. What's more, RFC (Request for Comment) 1347 from the IETF (Internet Engineering Task Force) specifies how these boxes are to be used on IP networks. Most vendors of computers with HIPPI channels also have software drivers for TCP/IP.
• Some in the industry have questioned whether HIPPI, which is a circuit-switched technology, can handle datagram traffic efficiently--especially given the small size of most TCP/IP packets. But HIPPI's streamlined signaling sequences allow connections to be set up and torn down in less than 1 microsecond. Thus, a single port on a HIPPI switch can deliver hundreds of thousands of IP datagrams every second, outperforming IP routers or hosts, which would typically be the sources of this traffic. What's more, switch latency is on the order of 160 nanoseconds. HIPPI is so fast, in fact, that any bottleneck would be associated with protocol processing and data handling at the end-devices.
• HIPPI has long been employed as a high-speed channel to peripherals like disk and tape controllers. HIPPI switches also can be used to link diverse storage devices, thus allowing for shared access across a network. Further, IPI-3 is employed for host-peripheral communications. This protocol is specifically designed to deliver high I/O with minimal CPU overhead.
• HIPPI also is well suited to hierarchical internetworking. In a three-tier arrangement, for example, Ethernet segments could be connected via FDDI rings, which in turn are connected by one or more HIPPI switches (see Figure 10.1). And HIPPI works equally well with fast Ethernet, full-duplex Ethernet, and straight Ethernet--anything, in fact, that speaks TCP/IP.
• In another demonstration at Supercomputing '94, Triplex Systems Corp. (Columbia, Md.) transmitted full-motion video through a HIPPI switch to a high-resolution monitor--with no loss of quality. Scenes from the Clint Eastwood thriller In the Line of Fire were shipped from a tape unit through a HIPPI switch to a frame buffer from PsiTech Inc. (Fountain Valley, Calif.); they were displayed on the hi-res monitor at a rate of five frames per second. Given that each frame contained more than 6 Mbytes of data, the sustained throughput was over 30 Mbyte/s. What's the big picture? Only HIPPI can deliver this sort of aggregate throughput. By comparison, ATM running at OC3 (155 Mbit/s) can deliver a maximum of 55 percent of that throughput.
• So-called workstation clusters could well become the high-performance computing systems of the future, capturing a good portion of the supercomputing market thanks to their dramatically improved price/performance ratios. But realizing this promise means that networking technologies must deliver the exponential increase in bandwidth required in such environments. Fortunately, HIPPI makes it simple and relatively inexpensive to build gigabit-speed, switched networks running TCP/ IP.
HIPPI has four major roles to play in such networked systems. First, it can link workstations and other hosts. Second, it can connect workstations with storage systems at higher speeds than any other currently available technology. Third, it can attach display peripherals for real-time visualization. Fourth, it can connect the entire system to another network.
• There are several potential wide-scale applications for HIPPI in workstation clusters--especially animation, scientific visualization, and high-fidelity imaging. All of these require very high data transmission rates to display large-size, high-quality color video.
• HIPPI-Serial connections, in combination with desktop HIPPI, make it possible to transfer images to their point of use at the requisite speed.
• HIPPI also makes it easy to set up affordable parallel processing workstation clusters that can be used to divide and conquer complex problems.
0 comments:
Post a Comment
Thanks for your Valuable comment