Network Management Strategies for the CIOOn the 1st run, having 100-Mbps cards in both the PCs, the time taken is 33 minutes. On the second run, having 1000-Mbps cards in both the PCs, the time taken is 29 minutes.
Should that be all -- a 4-minute difference? What realistically should be expected from Gigabit technology?
Requires Free Membership to View
With this very good test scenario, the problem lies in the PC architecture. What you are seeing is the limitation of PCI bus, networking cards and Microsoft Windows.
Up until the very latest versions of WinME and WinXP, Microsoft has programmed the network adapter like a serial port. This has some big problems since every packet forces a software interrupt and makes a very CPU intensive operation. As a result, most MS Windows systems are actually quite slow across the network.
The PCI bus in Intel machines is also a bottleneck; typically you can only get about 200-300 Mbps across the bus (I'm not exactly sure what the number is, but its theoretical maximum is about 600 Mbps). Now some of the new PCI architectures do better than this, like the 133 MHz buses. And certain machines will do better, like the major vendors (Dell, Compaq, HP, IBM) who have much better hardware design and chip integration and therefore no wait states on the PCI bus, which speeds things up by a staggering amount.
But wait?there's more!
A lot of the cheaper network adapters save money by using less buffer RAM and lower powered chips. These types of cards may well be Gigabit Ethernet, but they aren't much use in fileservers. They run pretty badly. The better quality NICs from the major vendors always outperform the cheaper NICs.
If you want really good performance, get a server NIC; these actually do CPU offload and have special PCI bus mechanisms and lots of other things that let them get up to as much as 400 Mbps.
Don't get me started on just how badly Microsoft performs -- it really is a reason to use Linux.
This was first published in July 2002