[OpenAFS-devel] .35 sec rx delay bug?

[Tom Keiser]

On 11/6/06, Jeffrey Altman <jaltman@secure-endpoints.com> wrote:
> Rainer Toebbicke wrote:
>
> > . With bigger windows, and a routed network, the windows of 350 ms for
> > ACKs is actually low, the price for retransmits is high. Here is makes
> > sense to increase the timeout.
>
> I wouldn't want to implement a longer fixed timeout.  From my work on
> Kermit we discovered a long time ago that fixed timeout values cannot
> provide optimal performance across all combinations of throughput,
> window size and latency.  Instead, per connection dynamic timeout
> computation is much more effective.
>
> The algorithm used in Kermit is:
>
>   receive timeout (msecs) = AVG(packet interval) + 3 * StDev(RTT)
>
> I suspect it would be fairly easy to implement something similar for RX.
>

Alternatively, there's always the RTO algorithms from rfcs 2988 and
2960 section 6.3.  It would interesting to benchmark several different
published algorithms on "typical" AFS workloads.  An NS model of Rx
would come in handy right about now...

-Tom