In my experience, TCP deals better with packet loss than a jittery RTT
(caused by huge buffering capability on linecards)
Unfortunately most people writing up SLAs have RTT measured as a very
long average (so a little bouncing around does not matter) but have quite low
packet loss targets. So they are biased towards large buffering to achive
near-zero loss. Obviously the best option is to offer different classes of
service, one providing constant latency with loss if neccessary and one
providing minimal loss with excessive buffering.
Pete