Internet Traffic Trends
A View from 67 ISPs
Craig Labovitz ([email protected])
Danny McPherson ([email protected])
Scott Iekel-Johnson (s[email protected])
Mike Hollyman ([email protected])
Internet Statistics
• Golden Age
– NSFNet 1986-1995
– Monthly ARTs reports
• protocol, ports
– Filtered prefix reports
• Today
– Some ISP specific traffic research and
commercial datasets
• e.g. Akamai, Google, etc.
– Lots of BGP data
– Many, many analyst reports
– Mostly people make stuff up
Internet Traffic Project
Global view of Internet traffic and attack trends
• Leverage commercial probe deployments
– Pool of 2,500+ active Flow / DPI collectors
– Across 250 ISPs / Content Provider / Higher Ed
– Deployed adjacent interesting bits of infrastructure
• Internet scale data collection
– Traffic, DPI, Mitigation and Security datasets
– Geographically and topologically diverse
Internet Traffic Project
• Outgrowth Fingerprint Sharing
Initiative
– 45 publicly disclosed participants
– and Annual ISP Security Survey
• All data voluntary anonymous data sharing
agreement with ISPs
• Still more research project than commercial
– Arbor, University of Michigan, Princeton (Intern)
– And 78 ISPs (and growing)
Internet Traffic Deployment
• 67 long-term participants (2 years)
• 17 unique countries
• 27 in US, though many have global footprint
Current Traffic Project Deployment
• 67 long-term ISPs (now 78)
– 5 MSO, 4 Tier1, 15 Tier2, 4 Content, 1 R&E
– Remainder not self-catagorized
• 1,270 routers
• 141,629 interfaces
• > 1.8 Tbps of average inter-domain traffic
• 485 days and counting (began September 2006)
Typical ISP Deployment
• Majority deployments are Flow from all peering
edge routers
– NetFlow / JFlow/ Sflow / IPFIX / etc
• Growing DPI from gigabit inline / portspan in
front of customers or server clouds
• Exported to commercial probes
– Usually 1/100 - 1/000 sampling
– Regexp or BGP based classification of border
interfaces to avoid double counting
– Data validated against interface SNMP counters
Anonymous XML Data
• Five minute traffic samples
– Traffic In/out of network (subset of
backbone traffic)
– Cross-products based on top N
protocols, ASNs, ports, applications,
etc.
• Traffic anomaly data
– Combination protocol signatures,
behavior and statistical variance from
baselines
– Distinguish Attack versus Flash Crowd
• Annotations and mitigation status
• Self-Categorization
– Tier1/2/3, Content, High Ed, etc
– Predominant geographic coverage area
Internet Traffic Project
statistics.arbor.net
ISP1
ISP2
ISP3
ISP4
ISP5
• Each participating ISP deployment submits XML
• Anonymous XML over SSL every hour
• Arbor managed servers collect/process
90 Day Protocol Distribution Trends
• No real surprises: TCP dominates followed by UDP
• Possible North America / Europe bias to dataset given
diurnal patterns
• Wither IPv6?
60 Day TCP Port Trends
• Again, no surprises: http/80 by far most prominent TCP
port
• In second place, TCP/4662 (edonkey) most prominent
inter-domain peer-2-peer file sharing protocol
• Rises of NNTP (ranks 3rd) as file sharing alternative
(alt.binaries!)
Internet Anomaly / Attack Summary
• 485 days
• Anomalies
– 616,631 total anomalies reported
– 1,271 average anomalies/day
• Attacks
– 353,588 classified attacks (57.34%)
– 729 average attacks/day
Internet Attack Propagation
• Each color
represents different
anonymous ISP (30
represented)
• Each line represents
different attack
• 7 Outbound ISPs,
10 attack streams (7
tcpsyn, 3 icmp)
generating 6.312
Mpps, one Russian
AV Vendor
Detect multi-ISP
distributed
attack
Internet Attack Packet Size Distribution
• Small packets predominate (pps attacks)
• Spectral analysis-like fingerprints of other attack types and tools
• Some issues with data collection methodology…
Internet Attack Types
• Data excluding floods
• After nine years, TCP SYN (31%) still dominates DDoS
• ICMP (27%) also prominent
Most Frequently Attacked Ports
• HTTP ports account for bulk of TCP-based attacks
• Fragmentation attacks lead the pack on the UDP front
Internet Attack Scale
• Unique attacks exceeding indicated BPS threshold for single ISP
• Average of three 1-Gbps or larger attacks per day over 485 days of
collection
• Two ~25 Gbps attacks reported by a single ISP (on same day, about
one hour apart, duration of ~35 minutes)
21 Days Y/Y
• Significant Y/Y growth
• Identify additional trends: Holiday Season typically slow
time for attackers
Challenges
• Balance commercial privacy with
research and business interests
• Data normalization / extrapolation
– Differing notions tier1
– Many business units within an ISP
• Data availability to other researchers
Questions?
Craig Labovitz ([email protected]),
Danny McPherson (d[email protected])
Scott Iekel-Johnson (s[email protected])
Mike Hollyman ([email protected])
