next up previous
Next: Fiber Distributed Digital Up: Available Networking Technologies Previous: Fibre Channel

Asynchronous Transfer Mode (ATM)

ATM technology is a networking technology designed to permit telephony, video, and data communications on the same network using statistical multiplexing to provide bandwidth on demand. ATM technology differs drastically from HiPPI and Fibre Channel in the area of statistical multiplexing multiple data streams onto a single physical signal to obtain maximum utilization of the bandwidth. ATM uses small 53 byte cells with a 48 byte payload to provide the potential for rapid switching that will maintain the constant-bit-rate characteristics and timing of isochronous signal sources. The small, constant cell size permits processing of the cell in parallel within a switch --- the entire cell can be moved bit-parallel through a switch. This bit-parallel feature in ATM cell processing means that data rates can increase to 10+ gigabits-per-second without requiring expensive GaAs technology as required in other giga-switch technologies.

ATM is being developed with very strong backing from the telecommunications industry to support telephony and video for wide-area connections, although companies like Fore Systems and SynOptics have effectively pushed ATM technology into the LAN environment. While there may be overlap in the LAN environment for ATM and HiPPI/Fibre Channel, these are complementary technologies. All supercomputer vendors have ATM interfaces for the user to interface to their hardware, in the same manner that they have used Ethernet and FDDI networks for some time. As ATM-based enterprise networks arise, all supercomputer vendors are prepared to interface their hardware to users via ATM. Moreover, NCube has included specialized ATM output interfaces to support real-time video from their newest computer, the NCube 3. This computer will support video distribution over AAL1 or AAL5, where ATM Adaptation Layer (AAL) type 1 is for constant-bit-rate data and AAL type 5 is for limited overhead available-bit-rate data. Hewlett Packard is offering a version of their Computational Cluster with an ATM-based LAN as the interconnection network. HP has incorporated a Fore Systems ASX-200 switch in the rack with eight workstations.

ATM technology is an area of personal expertise, nevertheless, there was much to learn from discussions with exhibitors at the SuperComputing '94 trade show. I saw a demonstration of ATM congestion control at the Digital Equipment Corporation (DEC) booth. They demonstrated 135 megabit-per-second throughput on a 155 megabit-per-second OC-3 link between two Alpha-based workstations connected with a single DEC ATM switch. Without congestion control, throughput was less than 40 megabits-per-second. Due to the queuing/buffering architecture within the DEC switch, one would expect no degradation of isochronous signals, because these time sensitive signals are carefully separated in the buffering/priority architecture of the switch. A single buffer-memory space is divided into partitions for each signal type and priority. At present the layout of buffers is static, although future versions of buffer software will permit dynamic allocation of buffer memory. Discussions with IBM revealed that they also recognize that reserved bandwidth signals must be processed separately from workstation generated available-bit-rate data streams to avoid data loss.

The SCinet '94 included two Fore Systems ASX 200 ATM switches and a GTE SPANet switch. The Fore Systems switches were used to connect the MAGIC Testbed and the JPL HiPPI Testbed. The GTE SPANet switch was used as a gateway between the ATM and HiPPI networks. There were three Fore Systems LAX 20 legacy network interconnection devices to route FDDI traffic over ATM. A scan of the network diagram is provided in figure 1.

I believe that ATM will be the workstation networking technology of the future, due in part to the backing of the technology from the telecommunications industry. ATM networks collapse the concept of LANs/MANs/WANs because the same protocol is used throughout the entire network, eliminating the requirements for routers or gateways. The transition to ATM-based LANs for all workstations is being hindered because of the cost to replace legacy networking technologies. Existing workstations have some network technology (by definition) and costs are between one and two thousand dollars to add a workstation to an ATM-based LAN. The cost-point for transitioning to ATM technology must be compared to the cost of maintaining the legacy networks and especially the costs (in both dollars and degraded performance) for additional routers if the number of networked workstations are increased. Standards-based interoperability between ATM networks and legacy Ethernet and FDDI networks are not here today, but can be expected to arrive in the next several years. Proprietary solutions exist from Fore Systems and SynOptics for interconnecting legacy networks to and through ATM-based LANs, but these vendor's propriety protocols are not interoperable.

While HiPPI and Fibre Channel technologies have limited utility as LAN technologies because of the limited maximum connection distances, ATM has no such distance limitation. Rather, the limitations of ATM in the LAN environment are the overhead encountered with the small cell size. As available bandwidth permits users to readily access computing resources across larger physical distances, ATM-based networking will overcome the resistance induced by the cost to upgrade networking capabilities in existing workstations. The requirements for wider-area access to computing resources will spur the wide-spread deployment of ATM-based networking and promote the implementation of ATM on desktop workstations. Further discussions on the use of ATM at Sandia National Labs for large-scale enterprise networks is provided in section 3.



next up previous
Next: Fiber Distributed Digital Up: Available Networking Technologies Previous: Fibre Channel



David P. Koester
Sun Oct 22 13:05:27 EDT 1995