Performance Comparison of Emerging Networking Technologies: ATM and Switched/High-Speed Ethernet


Andrew Rindos, Steven Woolet, David Cosby, Jerry Sents, Kishore Jotwani
IBM Corp., Networking Hardware Division,
P.O. Box 12195, Research Triangle Park NC 27709
Mladen Vouk
Computer Science Dept., Box 8206,
North Carolina State University, Raleigh NC 27695

The predominant LAN (local area network) technology is 10 Mbps Ethernet (10Base-T). However, a number of developments are occurring simultaneously to produce increasingly unacceptable congestion on today's LANs. Probably the most significant source of this congestion is the World Wide Web. Its explosive growth has dramatically increased the amount of file transfer activity across existing networks, while awakening previously relatively inactive network users to the entertainments and services available on the Internet. This increasing demand on available network bandwidth is developing in parallel with the ability of most personal computers and workstations to deliver it, as workstation processor power and overall performance increases at a relentless pace. This increased workstation capability has both arisen in response to the development of new applications that require such power/performance, and has spurred the development of new bandwidth-intensive applications. Some of this application development (especially in the area of video) has been the result of the availability of ATM (Asynchronous Transfer Mode) technology.

In an effort to relieve the congestion arising in existing Ethernets, the networking industry has rushed to provide new Ethernet products with a variety of capabilities that take advantage of the benefits of switching technology, providing congestion relief and collision reduction, while still utilizing the CSMA/CD physical access methods. 10-to-10 or 10-to-100 Mbps Ethernet switches can provide relief from current congestion by microsegmenting an existing LANs and providing higher bandwidth uplinks, thereby extending its lifetime. A number of choices are therefore available when a network is upgraded. At the LAN/campus level, the choice may be between higher-speed (possibly switched) 100 Mbps Ethernet and ATM-to-the-Desktop. For the backbone, the choice may be between ATM (155 Mbps and higher) and 100 Mbps full-duplex switched Ethernet, with future plans to migrate to a (promised) Gigabit Ethernet solution.

In an effort to assist a network designer/administrator in making these choices, we present measurements of existing ATM and high-speed/switched Ethernet products under a wide variety of operating systems, protocols and work station types. These include benchmark throughput measurements of: (1) 25 and 155 Mbps ATM adapters, for LAN Emulation and Classical IP over ATM, as well as early implementations of native ATM; (2) 10 and 100 Mbps shared Ethernet; (3) 10 and 100 Mbps switched Ethernet (half- and full-duplex); and (4) Ethernet-to-ATM bridging. These measurements were performed using properly tuned systems with sufficient resource capacity. Operating systems examined include Windows NT and 95, Novell NetWare, OS/2, DOS and AIX; protocols include IP, IPX and Netbios.

We will then present an evaluation of ATM and Ethernet switching based on the quality of support they provide MPEG1 and MPEG2 video sessions during concurrent large file transfers. MPEG1 video performance for both ATM and switched (and shared) 10/100 Mbps Ethernet is empirically evaluated, as is MPEG2 video performance for ATM. Since MPEG2 hardware/software is not currently available for Ethernet, MPEG2 comparisons with ATM are based on discrete event simulation studies.

Finally, we will discuss some of the underlying causes of observed results (both from the throughput studies, as well as from the video comparison studies). Using the collected data, we will also demonstrate which parameters and resources in the workstation/adapter/network have the greatest impact on end-to-end performance, and how they may be tuned or upgraded to improve overall performance. We will also present some simple mathematical models that illustrate some of effects.

As the measurements will show, when properly tuned and in an environment that provides sufficient supporting resource capacity (e.g., workstation processor speed), the ATM and Ethernet products examined are capable of producing throughputs close to their rated media speeds. The measurements will also demonstrate that the introduction of switched Ethernet technology to an existing Ethernet will produce significant improvements in network throughput and overall performance. It accomplishes this through the creation of of smaller isolated collision domains. Full duplex implementations further improve aggregate throughput capacity.

The video test results will indicate that switching technology (ATM or LAN) is required to support video transmission across networks. Shared Ethernet (even 100 Mbps) appears incapable of providing satisfactory support of multiple video sessions, so raw bandwidth alone is insufficient. Furthermore, ATM switching (with its 53 byte cell size) appears to be more robust in supporting jitter sensitive video, especially in the face of file transfers, than does Ethernet switching (with its variable, substantially larger frame sizes).