> "Slow start-> pkt loss -> Fast retransmit&Recovery-> congestion avoidance "
> 
> not
> "Slow start-> pkt loss -> rtx timer expire-> slow start"
> 
> I think the latter case will happen more frequenlty than the former case.

Well... I wouldn't really want to conjecture about the probabilities.
But, using SACK or NewReno loss recovery will help avert the RTO in this
situation.  In the case where half the segments are lost and there are
many outstanding segments, NewReno may not be quite what you want since
it will take a while and may or may not hit the RTO (depending on how
the implementation manages the timer).  But, SACK should deal with
multiple losses in a window of data in roughly one RTT without the RTO
(i.e., it's close to the first scenario you give above).  

(RFC3517 tries hard to be robust.  Of course, it can break down -- but,
I see no reason in your example why it would certainly breakdown.)

allman