Initial Spreading
Résumé
With most internet connections being short–lived (i.e. ~10 segments), it is very tempting to enlarge the TCP Initial Window (IW). This would save two of the three RTTs needed to transfer most of the web pages through a legacy slow start. However it has been demonstrated that the bursts created by a larger IW greatly impair global performance. An intuitive solution is the TCP Pacing. By spreading the transmission over the whole RTT, Pacing smoothes the bursts and delays the congestion. While postponing congestion provides good performance for short-lived connections, it could significantly deteriorate global network performance insofar as the reaction to congestion is also delayed. This paper analyzes the weaknesses of large IW and TCP Pacing and proposes a fast Start-Up mechanism to speed up short-lived connections while preserving long-term connections. Extensive simulations and analysis demonstrate that our solution is as efficient as a larger IW would be in an uncongested network and better than current mechanisms in congested environments.