USB devices and memory performance

[Alegria]

Today I'd like to show you a simple example of how a rather less known fact affects memory performance in benchmarking: USB devices conected to your computer (here I'll show you the impact of having the WiFi adapter enabled and then disabled on the P5K-E WiFi AP.
The first picture shows an Everest result with the USB device enabled, the second one without this USB WiFi card. This can easily reproduced with any memory benchmarking program such as cachemem, Sandra or anything else. They all will show an increase when running on a system with less USB devices but especially with those devices that generate interrupts more often such as networking devices. Go figure for yourself.
Interesting, isnt it?

Feel free to discuss this topic...

[k0NG0]

One may question the statistical validity of a change that is less than 3% for any given benchmark IMHO. Also, it's very seldom a task that will tax memory to an amount even remotely close to what a benchmark will won't be more held back by I/O than any USB interrupts that lurk around in the background noise.

Firewire is to be preferred over USB for any device that does block transfers (HDDs etc). Firewire is peer-to-peer and acts as a DMA peer, whereas USB is interrupt-driven and hierarchical. Thankfully, steadily increasing CPU performance and, lately, multi-cores will mask most of the performance degradation that USB incurs.

[Alegria]

Originally Posted by k0NG0

One may question the statistical validity of a change that is less than 3% for any given benchmark IMHO.

First of all it is reproducable, not only in my configuration but in several different configurations of the P5K-E WiFi running different Windows OSs, which is the first step to any validity. 3% may not be much of a performance loss, nearly unrecognizable but when it comes down to several devices the figures increase. Some people ask why their memory performance is not as expected and the first thing to blame is either the board or the memory itself. Less considered is the peripheral configuration with its impact on benchmarks. So for all these benchmark junkies this is a hint on investigating possible slower-than-expected results. Of course one could take a stopwatch and see if there is a difference in real-world performance or even better do a specific scripted real application task with a file creation and last access time stamp. Simple as that.