WAN Measurements Carey Williamson Department of Computer Science

WAN Measurements Carey Williamson Department of Computer Science University of Calgary 1

Introduction u There have been several studies of wide area network traffic (i. e. , Internet traffic) u We will look briefly at two of these: – Caceres, Danzig, Jamin, and Mitzel - 1991 – Paxson 1994 2

Measurement Method u The wide area network measurements collected by Caceres et al. focus solely on TCP/IP traffic u Uses tcpdump u Collected data from several different sites (e. g. , UC Berkeley site: 5. 9 million pkts/day) 3

Data Collection u Used Sun 3 workstation u Added special timer with 10 usec resolution u Collect 56 bytes of data from each packet, including data link, network, and transport layer information u Use well-known port ids to classify app’s u Packet loss: 0% (estimated) 4

Data Analysis u Characterizes different network applications based on a number of criteria u Number of bytes transferred u Duration of connections u Number of packets sent u Packet sizes u Packet interarrival times 5

Main Observations u Both interactive and bulk transfer traffic have a large number of short conversations u 75 -90% of bulk transfer connections send less than 10 kilobytes u 90% of interactive conversations send less than 1000 packets, and 50% last less than 90 seconds in duration 6

Main Observations (Cont’d) u Most conversations are bidirection (even the bulk transfer ones) u Bulk transfer accounts for over 50% of the packets and bytes transferred u Interactive traffic: small packets u Bulk transfer: big packets u Bimodal packet size distribution results 7

Summary u One of the first detailed studies of wide area network TCP/IP traffic u Identified characteristics of the traffic for different network applications u Proposed models for packet size distribution, packet interarrival times, etc u Modeling package tcplib is available (free) 8

More WAN Measurements u ‘‘Measurements and Models of Wide Area TCP Conversations”, Vern Paxson, 1993 u More wide area network measurements u All TCP conversations between Lawrence Berkeley Laboratory (LBL) and the rest of the Internet world 9

Measurement Methodology u Uses tcpdump u Captures only SYN (start) and FIN (end) packets of TCP connections (conversations) u Provides all the information needed for conversation level characterization u Dramatically reduces storage space needed u Makes longer trace durations possible 10

Measurement Tools u Sun 3/50 on a network between LBL and the Internet u Capture all packet headers with SYN or FIN u Saved to local disk for later analysis u Headers have port id info for identifying specific network applications 11

Measurements u Results are reported for one month of data (first data set collected November 1990) u 123, 703, 757 packets seen u Average of 3. 5 million per day (40 per sec) u 84 Mbyte of raw trace data u 210, 868 conversations u Represents 5. 6 Gbyte of user data 12

Measurements (Cont’d) u Elevent different application protocols identified in the traces u Results reported separately for each, as well as in aggregate u Repeated measurements at later times (e. g. , March 1991) to assess growth in traffic with time 13

Data Analysis u Presents aggregate workload characteristics, as well as a per application breakdown u Volume of data transferred u Network bandwidth used u Conversation lifetime u Conversation interarrival times u Geographical distribution u Develops models for each application as well 14

Summary u. A nice detailed study of wide area network TCP/IP traffic u Novel aspects: trace collection methodology, geographic analysis, models, growth u Identified significant growth in Internet traffic over fairly short time span u Growth continues to this day (and beyond!) 15