Measuring File System Performance

User measurement and tuning of computer hardware andcomponents is largely a lost art – we buy high performancehardware within our budget, and it is what it is. When we requiremore speed, or more likely just buy to keep up with technologicaladvances, we buy something new.

This came to mind as I wasevaluating why the media performance between a couple of home PCswas relatively poor: For whatever reason, streaming media fileswere stuttering, and photo replication was taking far too longgiven the quantities of data involved. While I just lived with thisfor a while (it is just a couple of home PCs), I finally decided toget some real metrics to know what I was dealing with. I decided tolook at the numbers from the file system level (rather thanmeasuring just the network itself to catch any high-level issues).Off I went with the fantastic application iozone (which also doubles as anexcellent test utility to gauge the performance of varioussegregated storage systems, which will be another entry coming verysoon), analyzing both the wired network and the wirelessnetwork.

To make a boring entry evenmore boring, it turns out that the source computer – running on annvidia nforce motherboard – was using the nforce chipset networkingadapter, with the secondary Marvell adapter disabled. All withstock settings. After running the first set of tests, seeing thatthe network performance over the 100Mbps network was yielding about800KB/second of actual transfer (despite the links all reporting a100Mbps signalling speed), I disabled that network adapter andenabled the Marvell, switching the ethernet cable over.

With that simple action,network throughput suddenly jumped to 12MB/second (the theoreticallimit of 100Mbps ethernet). A 15x performance improvement justbecause I finally decided to measure it and do something about it.Now I think I’ll play with the buffer settings to see what furtherbenefits can be gained. Then I’ll probably upgrade to 1000Mbps andstart again (amazing how inexpensive 1Gbps networking equipment isnow).

I always enjoy theseexercises because inevitably you go down a path and learn moreabout a fringe – yet still important – element of our computingworld. For instance some interesting details about the file cache of Windows 2000 (whichwas largely unchanged for Windows 2003).