Network Computingwww.networkcomputing.com

We used the Smart Windows software to throw a variety of different types of traffic at the devices under test. For starters, we tested 100-Mbps to 1,000-Mbps performance. We connected all of the available 100-Mbps ports on the tested device to the Smartbits analyzer, then incrementally sent traffic from the Fast Ethernet ports to the Gig abit Ethernet port on the device under test. This test was designed to see how well the switch could multiplex multiple data streams into a fat pipe.

To test switch throughput, we used Smart Windows to generate streams of wire speed 100-Mbps and/or 1,000-Mbps traffic, where applicable. All of the switches tested operated at wire speed.

To level the playing field, we tested Alteon Networks' AceNIC in a 200-MHz Sun Ultra-2 and Essential Communications' JackRabbit Gigabit Ethernet in a 195-MHz Silicon Graphics Octane. Both machines had a single processor.

We used Ganymede Software's Chariot benchmark utility as well as the publicly available NetPerf. Both products were tested as server adapters and connected to Foundry Networks' FastIron Backbone Switch. Eight Pentium Pro clients with full-duplex Fast Ethernet adapters were connected to t he switch to provide load. These clients were first benchmarked at approximately 96 Mbps each, using the respec tive benchmarking programs. All TCP window parameters were set to 64 KB for maximum throughput. The Chariot script used was FILESNDL, with a file send size of 1 MB. NetPerf was run with a variety of different message sizes; we found the best to be around 24 KB.

One fact to keep in mind when considering NIC performance is overall system throughput. Our benchmarks were designed to stress the NIC, not the entire system.

The Buffered Distributor
The first incarnations of Ethernet supported networks with diameters of up to 1,000 meters. This magic number came from the time it took an Ethernet signal to traverse the length of a copper wire. When Fast Ethernet arrived, the tenfold speed increase essentially limited the length of an individual cable length to 100 meters when operating in half duplex; full duplex doesn't have to deal with collisions, so this restriction was re moved. To carry Ethernet's CSMA/CD technology to 1,000 Mbps would require another tenfold decrease in the size of the network, but a 10-meter network wouldn't be of much use. To overcome this limitation, a collision-free technology was required, and one solution is full-duplex communications. The buffered distributor is the result of combining full-duplex technology with repeater technology. It is, in essence, a full-duplex repeater.

To achieve any reasonable distance with a Gigabit Ethernet repeater, full-duplex connections are essential. The buffered distributor works by allowing each device to be connected at full duplex. When a packet arrives from one port, it is put on the backplane of the distributor, and then sent to other ports on the device. Since all of the devices are connected in full-duplex, it is possible for the repeater to become oversubscribed. To avoid an oversubscribed situation, 802.3x Flow Control is implemented on a per-port basis. The result is a device capable of handling 1 Gbps of traffic in a repeated fashion, while maintaining a reasonable overall network diameter. When more than 1 Gbps of traff ic is presented, the device uses flow control to limit the amount of incoming data, yielding a collision-free environment with an aggregate throughput of 1 Gbps.