How We Tested, Part 3
No Gigabit Ethernet network would be complete without gigabit server connectivity. We tested gigabit NICs from Alteon, Essential, Intel, Packet Engines, Silicon Graphics and 3Com. NICs designed for Intel-based systems were tested in a 64-bit Intel quad-processor Xeon system running at 400-MHz per processor with a 100-MHz front-side bus. The 64-bit PCI bus operated at 33 MHz. At the time we tested, this was the biggest, fastest server available on the market, and we thank Intel for providing this host server for our benchmarks. The SGI product was tested in a dual-processor 300-MHz Octane system, and Alteon's Sun Microsystems adapter was tested in a dual-processor Sun Microsystems Ultra-60 system with a 64-bit, 66-MHz PCI bus. We extend our thanks to Sun, as well, for providing this system for our test bed.

To benchmark the NICs in this review, we chose Ganymede Software's Chariot, version 2.11, Build Level 273, because of its multiplatform nature. Sixty-three Windows NT Workstation clients running on 200-MHz Cubix managed servers were connected via Fast Ethernet through a Lucent P550 Cajun switch to provide load against the server NICs. We used a modified version of Chariot's FILESNDL script to measure maximum throughput of the NICs in this benchmark. The Intel server ran Windows NT 4.0 service pack 3 with the TCP Window Size set to 64 KB. The clients were similarly configured.

Numerous benchmarks revealed that 14 clients provided sufficient load on the server without overburdening it with excessive data requests. We tested with 14 clients requesting TCP traffic from the NIC under test. We also tested with an additional 14 clients sending data to the server simultaneously. In all cases, single-direction traffic resulted in higher throughput than did full-duplex traffic. This can be attributed to the fact that receiving traffic is more CPU-intensive than sending traffic under Windows NT 4.0.

We performed other tests using UDP (User Datagram Protocol), but the unreliable nature of this protocol resulted in lower performance numbers than with TCP. Likewise, we tested SPX, but many of the NICs faltered when excessive SPX traffic burdened the server. The results of the benchmarks appear in Figure #11.