I found this paper hard to read, not because it is badly written, but because it discusses the measurements and results in extraordinary depth, discussing the measured effect, coming up with potential explanations, possible effects on higher layers and even suggestions for implementations. The lecture on Thursday luckily turned my frustration when reading the paper in admiration for the pioneering role Paxson played in introducing internet measurement methodology and the enormous body of results he produced.
Methodology
- large-scale experiment: 37 sites, N to N bulk transfer, done at Dec 1994 and 1995
- TCP bulk transfer (versus ICMP packet probing): 1. this is how most Internet traffic really looks like, 2. TCP adapts to avoid unduly loading the network
- TCP measurements intertwine network and transport, tcpanaly separates these behaviors, TCP implementation specific (Vegas versus Reno, Windows NT versus Linux), tcpanaly can not recover time series analysis
- Pathologies: Reordering (on certain paths as high as 36%, due to route flaps), Replication (rare) and Corruption (1 in 5000 packets)
- Bottleneck bandwidth (difficult to estimate, basic idea: send two consecutive packets, at bottleneck bandwidth, the second packet has to wait for the transmission of the first packet, limited by time resolution, refined through bunches)
- packet loss (distinguishes between data and ack loss, loss at loaded and unloaded links) and effect on TCP (causes of a redundant retransmission: unavoidable, coarse feedback (no SACK), bad RTO)
- packet delay (due to queing/bandwidth, but even compression observed)
- Justification of measurement (TCP) methodology over other measurement techniques
- wide variety of path properties
- underlying assumptions about the network often broken
- robust TCP implementations can work non the less
No comments:
Post a Comment