Modeling Internet Application Traffic for Network Planning and Provisioning Takafumi Chujo Fujistu Laboratories of America, Inc.
Traffic mix on converged IP networks IP TRAFFIC MIX - P2P SCENARIO IP TRAFFIC BY TYPE - JP MORGAN-MCKINSEY 100% 100% SHARE OF TOTAL TRAFFIC 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 2001 2002 2003 2004 2005 2006 2007 2008 WEB PAGES RICH MEDIA P2P S2S 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1999 2000 2001 2002 2003 2004 2005 WEB PAGES RICH MEDIA P2P S2S ROBERT B. COHEN, GRID COMPUTING AND THE GROWTH OF THE INTERNET, GGF 41
Next-generation application traffic demands PoP Core IP Flow Size = mean 47kB Metro Current metro collects traffic from local users and send it to core and distributes the traffic from core to the users. Mobile Mesh Network Metro Core PoP IP Flow Size = 600MB,5GB Web services Gaming Grid Appliance (PS3) Future metro Supports randomly fluctuating, bursty traffic with randomly distributed peers.
Future traffic modeling Develop understanding of future traffic properties on core and metro networks Traffic Growth Traffic Mix Traffic Pattern (Metro/Core) Traffic Characteristics Develop understanding of technical and economic impacts on core and metro network architecture. Identify new technical issues on network planning and provisionin.
Self-similarity of traffic W. Willinger, et. al., Self-Smilarity Through High-Variability Statistical Analysis of Ethernet LAN Traffic at the Source Level, Apr. 1997
Burstiness of traffic Characterize property of future Internet traffic in terms of number of users, access bandwidth, content size and application Access Bandwidth Future MAN Traffic Bursty?? Future WAN Traffic Bursty?? Self Similar, Bursty LAN Traffic Bellcore Poisson-like Smooth WAN Traffic Bell Labs Number of Users
Modeling Web traffic: Web user distribution Sacramento San Francisco Seattle Bakersfield Los Angeles San Diego Grand Rapids Cleveland Pittsburgh Boston Milwaukee Allentown Minneapolis New York Detroit Des Moines Chicago Philadelphia Salt Lake Washington D.C. City Denver Kansas St. City Louis Raleigh Dover Dallas Knoxville Greensboro Atlanta Phoenix Austin Orlando 40 Largest US Metropolitan Areas San Houston Antonio Albany Tampa Miami Manchester Hartford West Palm Beach
Modeling Web traffic: Web server popularity Grand Manchester Cleveland Rapids Pittsburgh Albany Hartford Seattle Boston Milwaukee Allentown Minneapolis Sacramento New York Detroit Des Moines Chicago Philadelphia San Salt Lake Washington D.C. Francisco City Denver Kansas St. Bakersfield City Louis Raleigh Dover Los Knoxville Greensboro Angeles Dallas Atlanta West San Diego Phoenix Palm Beach Austin Orlando Based on IRCache logs, Jun. 2002 San Houston Antonio Tampa Miami
Modeling P2P traffic: Control traffic Control traffic volume: 3PB/month Gnutella network Aug. 2002
Modeling P2P traffic: P2P user distribution Seattle San Francisco Los Angeles San Diego Phoenix Boston Milwaukee Minneapolis Buffalo New York Chicago Philadelphia Washington D.C. Denver Kansas St. City Louis Dallas Atlanta Houston Gnutella network Aug. 2002 Miami
21:00 Usage daily pattern Web P2P 18% 16% Web Usage Daily Pattern 1.4 1.2 1 0.8 0.6 0.4 0.2 0 Average 3,000,000 Daily pattern Variation 14% 12% 10% 8% 6% 4% 2% 0% 0:00 3:00 6:00 9:00 12:00 15:00 18:00 0.00 AM 2.00 4.00 6.00 8.00 10.00 12.00 2.00 PM 4.00 6.00 8.00 10.00 Time of Day Time (PST) Gnutella network Aug. 2002 Percentage
Content size distribution Web P2P 0.01 Content Size Distribution (Average ~ 47 KBytes) Audio Average 4.5MB 10KB 100KB 1MB 10MB 100MB 1GB 0.0001 1E-06 1E-08 1E-10 1E-12 Lognormal Pareto Software Average 34.5MB 10KB 100KB 1MB 10MB 100MB 1GB 1E-14 0 0 1/10 1 10 100 1000 10000 100000 File Size (KByte) Video Average 52.5MB 10KB 100KB 1MB 10MB 100MB 1GB Gnutella network Aug. 2002
Traffic simulation and visualization tool Traffic Matrix: 3D view Traffic Volume: 2D time series Mean/Peak Ratio: 2D time series
Test network configuration Total Population for Ring: 5,600,000 Total Population for each Node: 800,000 R1 700,000 R2 300,000 R8 1,200,000 R7 POP 300,000 R6 600,000 R5 500,000 R3 1,200,000 R4 Node Population Router 1 (R1) 800,000 Router 2 (R2) 700,000 Router 3 (R3) 500,000 Router 4 (R4) 1,200,000 Router 5 (R5) 600,000 Router 6 (R6) 300,000 Router 7 (R7) 1,200,000 Router 8 (R8) 300,000 POP (POP) -
Web traffic: Current scenario 10msec 100msec 1sec 10sec 100sec 9-node metro ring, 2.8 million online users, 1.5Mbps access
Web traffic: Future scenario 10msec 100msec 1sec 10sec 100sec 9-node metro ring, 2.8 million online users, 100 Mbps access
P2P traffic: Current scenario Traffic Volume (kbps) Mean / Peak Window size: 10msec 100msec 1sec 10sec 1min 10min Access BW (max.): 3Mbps/384kbps, File Size Distribution: 10KB-1GB P2P Population : 5% of total population(5,600,000)
P2P traffic: Future scenario Traffic Volume (kbps) Mean /Peak Window size: 10msec 100msec 1sec 10sec 1min 10min Access BW (max.): 100Mbps/100Mbps, File Size Distribution: 10KB-5GB P2P Population : 15% of total population(5,600,000)
Resource provisioning window Resource Provision Window In a provision window, link capacity is provisioned at the peak of the traffic System efficiency = system utilization = mean to peak ratio Provision Window (90% efficiency) Population Access Bandwidth 1.5Mbps 1 hour 3Mbps 10Mbps 1 minute Network Management System GMPLS 100Mbps 50Mbps 1 second Burst time
Conclusions Internet traffic projection P2P accounts for 50% of total Internet traffic P2P traffic in particular very large video objects are dominating the Internet traffic growth Application traffic simulations allow accurate estimation and prediction of inter-metro traffic Traffic being self-similar traffic being bursty Actual factors that affect traffic burstiness: Number of users, access bandwidth, content size and application Potential for network planning and proactive bandwidth provisioning Dynamic resource provisioning to improve system efficiency for bursty traffic