TCP-friendly marking for scalable best-effort services on the internet
被引:0
作者:
Lo Monaco, Giampiero
论文数: 0引用数: 0
h-index: 0
机构:
ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180
Scient Corp., San Francisco, CA 94111ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180
Lo Monaco, Giampiero
[1
,2
]
Feroz, Azeem
论文数: 0引用数: 0
h-index: 0
机构:
ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180
Packeteer Inc., Cupertino, CA 95014ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180
Feroz, Azeem
[1
,3
]
Kalyanaraman, Shivkumar
论文数: 0引用数: 0
h-index: 0
机构:
ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180
Kalyanaraman, Shivkumar
[1
]
Xia, Yong
论文数: 0引用数: 0
h-index: 0
机构:
ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180
Xia, Yong
[1
]
机构:
[1] ECSE Department, Rensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180
This paper proposes the use of TCP-aware network-based packet marking, which in conjunction with differential packet dropping, is a powerful way to improve the performance of buffer management for best-effort services on the Internet. We extend the notion of TCP-Friendly packet marking, proposed recently by us [6], and apply it to improve the performance of traditional best-effort services. The TCP-aware use of deterministic packet marking at the network edge allows us to protect selected TCP packets from suffering loss. This significantly reduces the total number of timeouts and avoids the resulting service degradation. In particular, we protect TCP sessions with very small windows, disperse packet loss across a given window, and protect retransmitted packets from encountering loss. We show baseline results which illustrate that the performance gains are considerable (orders of magnitude) when compared to stateful packet dropping algorithms like FRED [15], and even when TCP SACK [18] implementations are employed. The scheme has been implemented in Linux 2.2.10, and all the results are based upon experimental data.