Upcoming Events

Cloud Connect
Santa Clara
Feb 13-16, 2012

Cloud Connect brings together the entire cloud eco-system to better understand the transformation we're experiencing and promises to be the defining event of the cloud computing industry. Learn about the latest cloud technologies and platforms from thought leaders in Cloud Connect’s comprehensive conference.

Register Now!

ATM Backbo ne Switches

Load-Sharing a Bonus On the positive side, Collage's redundant LES/BUS configuration was the only implementation that permitted load-sharing among multiple active servers. Again, this is a proprietary feature, but it's still worth investigating. Like the Xylan switch, the Madge switch supports active broadcast control, which enables you to scale a single ELAN to much greater sizes. Another powerful feature, and one unique to the Madge switch, is its AMON (ATM Monitoring) management package. The switch also can be configured to provide RMON (Remote Monitoring)-style matrix utilization information--a feature no other switch could match.

The Collage 740 has four 622-Mbps adapter slots and one 155-Mbps adapter slot. At its maximum density, the 740 can support up to 17 OC-3 connections; OC-12 support is in the works, according to Madge technical support. In addition, 25-Mbps ATM and T1/E1 over ATM adapters are available today.

Overall, the Madge switch performed well during our testing, but it lacked some of the more up-to-date features of the competition. Madge is making every effort to bring the Collage 740 in line with its competition--features like PNNI, MPOA and ATM622 are slated for 1998, and a new 4-Gbps switching module will bring the product up to speed later in 1998, according to Madge representatives. When it comes to fault-tolerant Token-Ring LANE, you'll have a hard time finding a better contender.

Joel Conover can be reached at jconover@nwc.com.

How We Tested

Testing ATM backbone switches is more than just blasting bits through a switch. One fact that has been proven repeatedly in independent tests is that ATM switches have no problem forwarding cells at wire speed. The connection-oriented nature of ATM leads to other performance bottlenecks. Characteristics like call setup speed, maximum VCC (virtual channel connection) capacity, buffer capacity and management, and QoS (quality of service) enforcement are more critical to the operation of an ATM switch. We tested switches for their maximum call setup rate, as well as their EPD (Early Packet Discard) and PPD (Partial Packet Discard) performance.

Maximum call setup rate is a measure of how many simultaneous call setups the ATM switching engine can handle. Generally, your ATM switch will never need to exceed 10 to 20 calls per second. However, in a disaster situation, such as a power outage, the number of calls per second that a switch can process will determine how quickly your network recovers.

We tested maximum call setup rate using a Netcom Systems' SmartBits analyzer loaded with ATM smartcards and its Smart Signaling software. The Netcom solution let us benchmark maximum call setup rate in a 1-to-1 or full mesh. We tested both configurations (see "Peak Call Rate," at right).

When you use an ATM switch as a LAN backbone switch, EPD and PPD can enhance the performance of your network. During congestion situations, it is likely that some cells will be dropped by the ATM switch. To the application or server, this is a disaster, because it means the packet will have to be retransmitted, creating even more congestion. Rather than randomly discarding cells from multiple packets, EPD and PPD provide an intelligent packet discard mechanism for the switch. When the switch is congested and one cell is discarded from a packet, the switch discards all of the cells in that packet, creating a buffer for other cells that could potentially pass through the switch unhindered. The industry has coined the term "goodput" to refer to the percentage of valid PDUs (Payload Data Units) the switch delivers intact to the client. Higher goodput means fewer retransmits and a generally healthier network.

PPD algorithm, we used an Adtech AX/4000 to measure the overall good PDU throughput (goodput) from a congested port. We generated a single stream of traffic at 80 Mbps. This single stream was then multicast into three 80-Mbps streams by a FORE Systems ASX-1000 switch. A total of four streams were directed at a single 155-Mbps OC-3 port, which created congestion on that port. The output of the port was connected back into the Adtech analyzer, where we counted the total number of valid and partial PDUs received. (See "Early Packet Discard/Partial Packet Discard Test Setup" on page 80.)

Devices were tested with EPD and PPD turned off and turned on. Each test was run for one minute. The total number of valid PDUs received is graphed versus total PDUs sent. In this case, a PDU consisted of 2,730 cells--the size of two 64-KB TCP/IP windows. These PDUs burst at a rate of 80 Mbps.

Note each vendor's good throu ghput in the graphs at right and the number of partial PDUs transmitted. Generally, a high number of partial PDUs is bad, though a midrange number may not be unusual. Some algorithms permit the last cell of the PDU to be forwarded to the host, so the host knows an error occurred. Other algorithms (like FORE's) don't forward any cells once the EPD/PPD algorithm takes over. In either case, it is up to higher-level protocols like TCP/IP and IPX to retransmit the discarded cells.