Download presentation
Presentation is loading. Please wait.
1
Multi-level Application-based Traffic Characterization in a Large-scale Wireless Network Maria Papadopouli 1,2 Joint Research with Thomas Karagianis 3 and Manolis Ploumidis 1,2 1 Department of Computer Science, University of Crete 2 Institute of Computer Science, Foundation for Research and Technology-Hellas 3 Microsoft Research * This work was partially supported by General Secretariat for Research and Technology and by European Commission with a Marie Curie IRG grant COST-TMA: meeting @ Samos, September 22 nd, 23 rd 2008
2
2 Research interests Traffic modeling Impact of parameters (number of flows, flow inter-arrivals, flow sizes) on accuracy Topology & mobility modeling Traffic forecasting (moving averages, Singular Spectrum Analysis, etc) Client profiling Mobile p2p computing Data diffusion using realistic mobility models Efficient selection of appropriate network interface/channel based on network conditions/application requirements Efficient distributed monitoring Understanding the impact of network conditions on user experience
3
3 Roadmap Objectives Testbed, data acquisition & preprocessing Data analysis Aggregate traffic AP traffic Client traffic Conclusions Research in progress …
4
4 Objectives Classify flows into application types Identify dominant & popular application types Compare UNC network with other wired & wireless networks Characterize AP & client traffic
5
5 Infrastructure
6
6 Testbed, data acquisition & preprocessing Testbed 488 APs, 382 monitored 6,593 distinct MAC addresses – 9,125 distinct IPs Data acquisition Packet header traces from egress router Client SNMP data Data preprocessing Correlation of packet headers with client SNMP Classification of flows using BLINC
7
7 Classification with BLINC: heuristics Host behavior (e.g., client-server, collaborative) o Host popularity: number of distinct destination IPs o Clusters of hosts using a collaborative application o Number of source ports Transport layer protocol: TCP vs. UDP Cardinality of sets (ports vs. IPs) Per flow average packet size o Constant in several applications (e.g., malware) “Farms” of services: neighboring IPs Non-payload flows (e.g., attacks)
8
8 Graphlet library
9
9 Dominant application types Application typeFlows(%)Bytes(%)Packets(%) Network Management 9.950.421.54 Chat2.050.481.47 Web35.0657.5946.88 P2P30.0424.8534.46 Online Games1.110.010.07 FTP0.911.571.72 Mail0.070.330.21 AddScan6.40.120.58 PortScan0.390.320.28 Streaming0.10.170.19 Unknown13.214.0912.64
10
10 Popular application types Clients with at least one flow per application type Application typeClients(%) Network Management17 Chat73 Web99 P2P43 Online Games4 Ftp7 Mail1.5 AddScan73 PortScan1.4 Streaming0.5 Unknown84
11
11 Compare with other testbeds Traffic share for most dominant application types Wired & wireless testbeds UNC wired network Dartmouth wireless infrastructure Residential campus % Res. CampusUNC WiredUNC WirelessDartmouth Web37.548.6857.5928.6 P2P31.934.8524.8519.3 may have missed all Web traffic that was not accessed through one of the well-known ports for Web
12
12 Home application type of APs Traffic of this application type > than x% of total AP traffic Web most prevalent home application type xWeb(%)P2P(%)Ftp(%)Mail(%)Unkn 5085.96.170.2804.2 7555.80.28000.84 9025.20.28000
13
13 Client traffic characterization Client home application: Application type of which this clients transfer >X% of their traffic Clients have strong application preferences ~ 50% of clients have home application type (for X=90) Web: most prevalent home application type Clients with no home application are dominated by Web Only a minority of clients have P2P as dominant application
14
14 Wireless traffic load Wide range of workloads & log normality is prevalent Light traffic load but with long tails Dichotomy among APs: APs dominated by uploaders APs dominated by downloaders Majority of APs send & receive packets of small size Significant number of APs with asymmetric packet sizes: APs with large sent & small receive packets APs with small sent & large receive packets
15
15 Application-based characterization Most popular applications Web browsing & p2p accounting ~81% of total traffic These applications dominate most users and APs Web dominates both AP & client traffic share Network management & scanning activity ~17% of total flows Application-mix varies within APs of same building Wireless clients with strong application-type interests File transfer flows (e.g., ftp, p2p) are heavier in wired network than in wireless one Flow sizes per application type Different between wired & wireless network
16
16 In progress … Focus on applications with real-time constraints Impact of “extreme” network conditions on performance & user satisfaction Statistical analysis for client profiles Comparable analysis with other wireless networks
17
17 UNC/FORTH Web Archive Online repository of Wireless measurement traces Packet header, SNMP, SYSLOG, signal quality Models Tools http://netserver.ics.forth.gr/datatraces Login/ password access after free registration Maria Papadopouli mgp@ics.forth.grmgp@ics.forth.gr
18
18 Total network traffic across APs
19
19 Application traffic share across APs
20
20 Traffic asymmetry (2/2)
21
21 BLINC BLINd Classification Flows in application types Focus on end hosts rather than on flow 3-level host behavior analysis Social Functional Application Application signature based classification Accurate flows classification
22
22 Heuristics (2/2) 1. Community heuristic Farms of services in neighboring IPs 2. Recursive detection Interaction between servers Mail with Razor servers
23
23 Application level Transport layer interaction between hosts Based on TCP 4-tuple Empirically derived signatures – graphlets Nodes: Src,Dst IP & Src,Dst Port Edges: Flows through this TCP-tuple Protocol type Host behavior against graphlet library
24
24 Bldg level application usage patterns % of APs with home application type / bldg type Weak correlation between building category & # of APs with home application Distinct APs different configurations Uneven traffic distribution across APs of same bldg APs dominated by Web, P2P, or unknown traffic
25
25 Conclusions Three-level characterization of large scale infrastructure Support admission control & AP selection mechanisms Indicate user trends Assist application specific traffic modeling Web dominates both AP & client traffic share P2P systems bear a significant impact Clients have strong application preferences
26
26 Heuristics used in classification 1. Transport layer protocol: TCP vs. UDP 2. Cardinality of sets Ports vs. IPs Constant in several applications (e.g., malware) 3. Community heuristic Farms of services in neighboring IPs 4. Non-payload flows (e.g., attacks )
27
27 Attack graphlets Address-Scan attack Address-Scan attack for specific IP set Port-scan attack
28
28 P2P Graphlets
29
29 Traffic asymmetry (1/2) Asymmetry index = total downloaded / total uploaded traffic Certain APs dominated by uploaders Asymmetry index / application type Asymmetry index for P2P traffic < 1 for 40% of APs
30
30 Flow sizes per application type
31
31 Wireless user application preferences Similar between wireless & wired users Flow sizes / application type Different between wired & wireless network Possible reasons Application dependent User-driven
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.