IT executives looking to pump up the performance on their back-end systems by marrying Gigabit Ethernet and Windows 2000 may want to take another look under the operating system's hood.
A recent study by Tolly Research, the independent testing arm of the Tolly Group, found that Windows NT delivers Gigabit Ethernet throughput equal to or better than Win 2000. Tolly's finding contrasts Microsoft's testing that found Win 2000 optimised to deliver gains in Gigabit Ethernet throughput.
Some end users have taken notice of Tolly's findings, but weigh them against other issues. "I have found generally that NT throughput is a little quicker, but with Windows 2000 I can run applications like Exchange 2000," says Josh Mitts, an IT administrator at Treasure on the Net, an online gaming company.
Microsoft officials say Tolly's conclusions are not a fair comparison, citing variables such as client operating systems, network adapters, LAN design, traffic-generating tools and methodologies.
But Tolly says enterprise customers looking at Gigabit Ethernet don't need to rush into Win 2000 for performance gains.
"It is a good [operating system], but have they embellished in certain areas? Definitely," says Chris Eichman, research manager for the Win 2000 track at Tolly Research. "If you are migrating [for] increased Gigabit Ethernet performance, now is not the time."
The Tolly test was a response to Microsoft tests last February that showed 1Gbps throughput on Win 2000 with a standard frame size of 1518MB and 2Gbps throughput using 9018MB Jumbo Frames with a dual-processor/dual network interface card (NIC) setup. In contrast, Tolly was only able to record a high throughput at 837Mbps using Jumbo Frames and a dual-processor/dual-NIC configuration on Win 2000. The highest throughput was on NT - 842.3Mbps - using the same configuration.
Microsoft is urging enterprise customers to combine Gigabit Ethernet and Win 2000 to improve performance of back-end cluster and transaction servers. Gigabit Ethernet use is expected to skyrocket in the next three years. Last year, there were 4.2 million switch ports and/or NICs shipped for LAN use, according to Cahners In-Stat Group. That number is expected to roughly double every year through to 2004.
Tolly says Microsoft's throughput numbers may be inflated by the NTttcp packet-blasting tool it used and by testing on a highly segmented LAN. Microsoft officials admitted their LAN had two clients per segment.
Tolly used the NTttcp tool, but also used NetIQ's application traffic simulation tool, Chariot 4.0, which it says is a more real-world test.
With Chariot in a single-processor/single-NIC configuration, Win 2000 throughput was 416Mbps compared to 442Mbps for NT with standard frames, and 539.5Mbps using Jumbo Frames compared to 534.9Mbps for NT. The latter is only one of two categories in which Win 2000 outperformed NT. The other was the single-processor/single-NIC test using NTttcp and standard frames where the comparison was 582Mbps versus 581.
Microsoft says the comparison is not apples-to-apples.
"Their methodology and configuration is unable to yield similar results . . . because of constraints introduced by their testing methods," says Steven Adler, product manager for Win 2000.
One key, Adler says, is Tolly used some NT 4.0 clients, which lowered throughput, because they don't have a TCP/IP stack optimised for Jumbo Frames and off-loading work to a NIC.
"We are not in disagreement with their results, but we do strongly disagree with the conclusions they have drawn," he says.
The conclusion, Tolly's Eichman says, is that "your own environment, topology and switch architecture will dramatically affect your results".