2008年3月12日 星期三

End-to-End Internet Packet Dynamics

Vern Paxson
End-to-end internet packet dynamics
ACM/IEEE Transactions on Networking, 7(3):277-292, June 1999

[Summary]
This paper mainly focuses on the TCP traffic analysis between sites, while TCP is widely used in real world. They made an analysis based on the measurement of TCP transfer traffic, and network behavior. The measurement is obtained by sending and receiving 100 Kbyte TCP transfer between 2 of 35 sites, which are separately located in 9 countries in US and Europe, and they were taken in 1994 and 1995, and these two set of data are called N1 and N2. N2 has enlarged window size ensuring TCP connection throughput. In this paper, they listed and analyzed some prevalent network events like packet loss, network pathologies, detection of bandwidth bottleneck in the path, and packet delay.
Packet loss is the important thing that every networking people care about. In this paper, the author wants to find out a general pattern of packet lost. The author claims that acks are better for determining packet loss rate because of transmission rate of acks is stable. It’s also claimed that packet loss is somehow correlated. That is, if a packet is lost, the next packet has good chance to lose.
Packet delay is also a hot topic. Due to asymmetric path of packets, delay is calculated in one direction, called one-way transit time (OTT). And this would be modeled to a certain distribution. Furthermore, the author tries to find out available bandwidth.
Network pathology, meaning unusual or unexpected packet delivery, is crucial to advance analysis. This paper talked about out-of-order delivery, packet replication, and packet corruption. For out-of-order delivery, which has little impact on TCP performance, effects TCP "duplicate acks" that determines if the packet needs re-transmission. They concluded that out-of-order delivery is site independent, and asymmetric. Packet replication, which is an unnecessary retransmission event that sender mistake the packet was not properly received, is also site independent. Packet corruption occurs when the transmission channel has noise or some error while going through routers.
In this paper, bottleneck bandwidth is also discussed. Bottleneck bandwidth means the upper limit on how quickly the network can deliver the sender’s data to the receiver. The bottleneck estimation technique used is based on the idea of “packet pair”. The sender sends consecutive packets and measures the time interval of these packets at sending and receiving. But somehow this method suffers from several potential problems like out-of-order delivery. Motivated by the shortcomings of “packet pair”, packet bunches, instead of pairs, are used in bottleneck estimation.

[Comments]
Clarity: I think this paper is a good at clarity. The author made a topic he wants to discuss at first. And the reason of this topic is explained. The measurement method and result analysis is illustrated. Several graphs and tables are used to express the measurement result, and the author explained these figures and tables well.
Correctness: The author shows some graphs of measurement result that is close to what I think it would be and the phenomenon behind these measurements are discussed. However, several problems are stated but still not having a proper answer yet.
Originality: This paper measured TCP network traffic and brought up some new view to these measurement.
Significance: I will say this paper is significant. For all of networking people, TCP traffic and packet dynamics analysis is crucial.