van@HORSE.EE.LBL.GOV (Van Jacobson) (06/14/91)
A new tech report (LBL-30840): "Measurements and Models of Wide Area TCP Conversations" by Vern Paxson, LBL Computer Systems Engineering Group is available for anonymous ftp from ftp.ee.lbl.gov (128.3.254.68), file tcpmeasurements.ps.Z (this is a 1MB compressed postscript file). This study is similar to that described in the recently announced paper by Danzig, et.al., of USC. However, a much more efficient means of data capture was employed (described in the paper) which allowed the collection & analysis of every conversation in to and out of the LBL campus over a period of two months, rather the the two hour period of the USC study. This greater temporal scope allowed the investigation of long-term TCP/IP traffic variation (i.e., hourly, daily, weekly and monthly). It also allowed a detailed investigation of the geographic distribution of traffic. (This and other studies have shown that there is substantial short term correllation in conversation patterns and it is not possible to reliably assess geographic distribution from samples spanning less than a few days.) The abstract of the paper is attached. - Van --------------------------- Abstract This paper describes measurements of all of the wide area network TCP conversations between the Lawrence Berkeley Laboratory (LBL) and the rest of the world for the months of November, 1990, and March, 1991. Some 500,000 conversations were recorded, encompassing 11 different major protocols. We look at aggregate characteristics of these conversations, both overall and by TCP protocol (e.g., smtp, ftp), computing the distributions of amount of data transferred, network bandwidth used, conversation lifetimes and conversation interarrival times. Temporal traffic variation is also investigated, showing the variation of number of active conversations and network bandwidth utilization over periods of 24 hours, 7 days and 30 days. Long term variation is also investigated by separately analyzing November and March data (which reveals a 10-20% increase in almost all aggregate traffic characteristics in just four months). We classify each conversation geographically and discover that the connectivity of the conversations was remarkably rich, including traffic to 48 of the 50 states in the U.S. and 23 foreign countries. Finally, we develop a number of models for describing conversations of the various protocols. From these models we can more readily assess how each protocol is used and how the use changes as network utilization grows.