Nathaniel Smith published a set of patches for the Linux iwl3945 driver. He reports about two orders of magnitude improvement of his RTT under load, with no bandwidth performance difference. Latencies went from several seconds latency in his simple tests, to of order 30ms under load.
Good stuff, Nathaniel! Even better is that this may be usable on my own laptop, though I have to hunt a bit to see. :-).
jg's Ramblings 
February 13, 2011 at 9:38 pm |
The patches affect all of the devices supported by the iwlwifi driver (i.e., all the devices listed here: http://intellinuxwireless.org/). They’re just not *tested* on anything other than my iwl3945-using laptop, so if you have some other IWL-family hardware, then testing would be especially useful.
February 13, 2011 at 10:21 pm |
Thanks, Nathaniel. It looks like your patches may work on my laptop. I have a 6200agn, IIRC.
I’ll take them for a spin if I get a chance, though I have more writing deadlines coming up, I don’t know if I’ll get to it soon or not.
Also, have you looked at what Tianji Li, Douglas Leith, and David Malone have done, found at: http://www.hamilton.ie/tianji_li/ ?
February 13, 2011 at 11:09 pm |
Neat! It looks quite complementary — their goal is to intelligently estimate the *total* buffer required, which depends on both transmission rate and the need to buffer bursts. My patches are aimed at automatically adjusting the portion of that buffer which lives in the driver; its proper sizing depends only on transmission rate and the kernel scheduling latency.
Perhaps someone will implement their algorithm as a Qdisc. It occurs to me that to do this properly, drivers should have some way to export their tx queue size, so that any higher-level queue management can take it into account…