Hi, On Fri, 21 Dec 2007, Ilpo Järvinen wrote:
> > I need to re-read properly, but I think the same problem affects the > > microsecond values where TCP_CONG_RTT_STAMP is set (used by vegas, veno, > > yeah, illinois). I might follow up with another patch which changes the > > behaviour where TCP_CONG_RTT_STAMP when I'm more sure of that. > > Please do, you might have to remove fully_acked checks to do that right > though so it won't be as straight-forward change as this one and requires > some amount of thinking to result in a right thing. The TCP_CONG_RTT_STAMP code does need to be fixed similarly. A combined patch will follow this mail. As I understand it, the fully_acked checks kick in where the ACK is a portion of a TSO chunk and doesn't completely ACK that chunk. Leaving the checks in place means you get one rtt for each TSO chunk, based on the ACK for the last byte of the chunk. This rtt therefore is the maximum available and includes the time-lag between the first and last chunk being acked. Removing the tests gives you an RTT value for each ACK in a tso chunk, including the minimum and maximum. It seems the minimum rtt is the best indicator of congestion. On the other hand having all available RTTs gives the congestion avoidance greater knowledge of how the RTT is evolving (albeit somewhat coloured by TSO delays which don't seem too severe in my tests). The patch I'll suggest for now takes all RTT values available, rather than the old maximum or adding extra logic to pluck out the minimum. I've captured some rtt values from timestamps and from tcp_clean_rtx_queue() to demonstrate the effect of this patch: http://www.hamilton.ie/gavinmc/linux/tcp_clean_rtx_queue.html#usec As always, comments are most welcome, Gavin -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
