> We are also considering taking out Fast Recovery as a recommended
> mechanism. There could be some problems when using large windows
> and SACK makes it moot.
Our thinking stems from a conversation a couple weeks ago with Raj
about the problem with fast retransmit/recovery (I think Sally Floyd
has a note about this quirk, as well). Our ideas are not yet fully
baked which is why we have not dropped a note to the list yet. In
fact, I have not gotten a chance to do more than skim Raj's
documentation of the problem. But, since Dan mentioned it, I
thought I should go ahead and send out what we are thinking about...
Basically, the idea is that if we repair multiple losses in a single
window of data with more than one fast retransmit/recovery we are
sacrificing cwnd. For each repair we are pulling down cwnd by half.
And, building cwnd back up with congestion avoidance is really,
really slow over some satellite links. This problem is worse as
cwnd gets bigger. It seems to us that the right solution is to just
use some SACK algorithm. So, we are sort of kicking around what
exactly we should recommend as far as non-SACK algorithms.
Comments appreciated...
allman
--- http://gigahertz.lerc.nasa.gov/~mallman/
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:44 EST