N. Hu (CMU) L. Li (Bell labs) Z. M. Mao. (U. Michigan)

P. Steenkiste (CMU) J. Wang (AT&T)

Infocom 2005

Presented ByMohammad Malli

PhD student seminar

Planete Project

A Measurement Study of Internet Bottlenecks

November 28, 2005 2

Goals Recently, many active probing tools have been

developed for measuring and locating bandwidth bottlenecks, but

Q1. How persistent are the Internet bottlenecks?– Important for measurement frequency

Q2. Are bottlenecks shared by end users within the same prefix?– Useful for path bandwidth inference

Q3. What relationship exists between bottleneck and packet loss and queuing delay?

– Useful for congestion identification

Q4. What relationship exists between bottleneck and router and link properties?

– Important for traffic engineering

November 28, 2005 3

Related Work• Persistence of Internet path properties

– Zhang [IMW-01], Paxson [TR-2000], Labovitz [TON-1998, Infocom-1999]

• loss, delay, pkt ordering,..• The persistence of the bottleneck location does not

considered

• Congestion points sharing – Katabi [TR-2001], Rubenstein [Sigmetrics-2000]

• Flows-based study and not e2e paths-based

• Correlation among Internet path properties – Paxson [1996]

• e2e level and not at the location level

• Correlation between router and link properties– Agarwal [PAM 2004]

November 28, 2005 4

Data collection

• Probing– Source: a CMU host– Destinations: 960 IP addresses– 10 continuous probings for each destination

(1.5 minutes)• Repeat for 38 days (for persistence study)

S

DDDDDDD

CMU 960 InternetDestinations

Day-1

Day-2

Day-38

…

November 28, 2005 5

Pathneck

• An active probing tool that can detect Internet bottleneck location– For details, refer to

“Locating Internet Bottlenecks: Algorithms, Measurements, and Implications” [SIGCOMM’04]

– Source code: www.cs.cmu.edu/~hnn/pathneck

• Pathneck characteristics– Low overhead (i.e., in order of 10s-100s KB)– Single-end control (sender only)

• Pathneck output used in this work– Bottleneck link location– Route

November 28, 2005 6

Recursive Packet Train (RPT) in Pathneck

Load packets

60 pkts, 500 B

TTL

255255255255

measurement packets

measurement packets

30 pkts, 60 B 30 pkts, 60 B

2 130301 2

Load packets are used to measure available bandwidth Measurement packets are used to obtain location information

UDP packets

November 28, 2005 7

Gap value

RouterSender

Packet train

Time axis

November 28, 2005 8

RouterSender

Drop m. packetSend ICMP

Gap value

November 28, 2005 9

RouterSender

Drop m. packetSend ICMP

Recv ICMP

Gap value

November 28, 2005 10

RouterSender

Drop m. packetSend ICMP

Recv ICMP

Drop m. packetSend ICMP

Gap value

November 28, 2005 11

RouterSender

Drop m. packetSend ICMP

Recv ICMP

Drop m. packetSend ICMP

Recv ICMP

Gap value

Gap valueRPT probing is repeated 10 times for each pair of nodes

November 28, 2005 12

Terminology

Persistent probing set is the probing set where all n probings follow the same route

November 28, 2005 13

Route Persistence

• Route change is very common and must be considered for bottleneck persistence analysis– Consistent with the results from Zhang, et. al. [IMW-01] on route

persistence

AS level

Location level

over 9 days

November 28, 2005 14

Bottleneck Persistence

• Persistence of a bottleneck R

• Bottleneck Persistence of a path Max(Persist(R)) for all bottlenecks R

• Two views:1. End-to-end view ― per (src, dst) pair

– Includes the impact of route change2. Route-based view ― per route

– Removes the impact of route change

Persist(R) =# of persistent probing sets R is bottleneck

# of persistent probing sets R appears

November 28, 2005 15

Bottleneck Persistence

1. Bottleneck persistence in route-based view is higher than end-to-end view

2. AS-level bottleneck persistence is very similar to that from location level

3. 20% bottlenecks have perfect persistence in end-to-end view, and 30% for route-based view

3

2

2

1

November 28, 2005 16

Results summary• Only 20-30% Internet bottlenecks have perfect persistence

– Application should be ready for bottleneck location change

• Bottleneck locations have a strong (60%) correlation with packet loss locations (2 hops away)

– Bottleneck and loss detections should be used together for congestion detection

• Only less than 10 % of the destinations in a prefix cluster share a bottleneck more than half of the time

– End users can not assume common bottlenecks

• Bottleneck has no clear relationship with link capacity, router CPU load, and memory usage

• A clear correlation between bottlenecks and link loads– Network engineers should focus on traffic load to eliminate

bottlenecks

November 28, 2005 17

Limitations Interesting study but ..

• How much the obtained statistics are representative for the whole Internet, since

– the few sources used for probing are a CMU node, 8 Planetlab nodes, and 13 RON nodes

– the number of probed destinations are 960 • <<< # of Internet paths

• Pathneck limitations– Load pkts are larger than what the firewalls permit

• only forward the 60 byte UDP packets – Anyway, Pathneck is not able to measure the pkt train length on

the last link due to ICMP rate limiting • theoricaly, the destination must send a ‘destination port

unreachable’ for each pkt

November 28, 2005 18

Thank you for your listening

November 28, 2005 19

Backup

November 28, 2005 20

Bottleneck vs. loss | delay

• Possible congestion indication– Large queuing delay– Packet loss– Bottleneck

• They do not always occur together– Packet scheduling algorithm large queuing delay– Traffic burstiness or RED packet loss– Small link capacity bottleneck

• Bottleneck ? link loss | large link delay

November 28, 2005 21

Trace

• Collected on the same set of 960 paths, but independent measurements

1. Detect bottleneck location using Pathneck2. Detect loss location using Tulip

– Only use the forward path results3. Detect link queuing delay using Tulip

– medianRTT – minRTT• [ Tulip was developed in University of Washington, SOSP’03 ]

• The analysis is based on the 382 paths for which both bottleneck location and packet loss are detected

November 28, 2005 22

Bottleneck Packet Loss

November 28, 2005 23

Bottleneck Queueing Delay