Things You Can Do in Budapest During Summer Time

1
Things You Can Do in Budapest During Summer Time

• Aleksandar Kuzmanovic
• Rice University Ericsson Traffic Lab
• October 2001

2
Budapest

• Buda-Pest
• History
• Architecture
• Decadency
• Ericsson Traffic Lab
• Unix
• Young people
• Security

3
People

• Ericsson Traffic Lab, Budapest
• Andras Veres
• EMULab, The University of Melbourne
• Attila Pasztor, Darryl Veitch
• Budapest University, Department of Physics
• Gabor Simon, Gabor Vattay

4
My Goals

• Modeling and reality?

5
Active Probing

• Internal network
• inaccessible
• Poor understanding of
• origins of complex
• network dynamics
• Statistics is like the bikini

Goal Estimate the sample path of cross-traffic
6
Background

• Precise measurement infrastructure

One way delay (sec) -
Packet departure time (sec) -
7
Probing Uncertainty Principle

• Should not allow queue to empty between probe
  packets
• Small T for accurate measurements
• but probe traffic would disturb cross-traffic
  (and overflow bottleneck buffer!)
• Larger T leads to measurement uncertainties
• queue could empty between probes

8
Theory

• Lindleys equation
• CT information imbedded in delay
• Ideal case minimally backlogging condition QK99

9
Design Space

• No time synchronization
• between end points
• 2.000.000.000
• Delay difference
• Non-intrusiveness
• Stability

One way delay (sec) -
Time (sec) -
10
Probing scheme
11
Aside

• Experiments on LAN
• Time-sharing in Linux

Inter-arrival time (sec) -
Time (sec) -
12
Traffic Lab Experiment Setup

• Tcpdump, iperf (HTTP, TCP), sender, receiver

13
Routes

• Tech. University Budapest
• 6 nec2.ttt.bme.hu (152.66.247.2)
• 14.980 ms 14.212 ms 14.252 ms
• Columbia University, NY
• 16 bongo.comet.columbia.edu (128.59.
• 64.210) 150.725 ms 127.967 ms
• University of Melbourne
• 23 potoroo.ee.mu.OZ.AU (128.250.76.186)
• 384.271 ms 383.408 ms 382.977 ms
• Houston, we have a problem -

One way delay (sec) -
Time (sec) -
14
Wide Area vs. Metropolitan Area Network
15
Cross Traffic EstimationSample Path

• Differentiate
• delay, know C
• Cross traffic
• from Budapest
• to NY
• iperf - 1 TCP
• Probing traffic
• from Budapest
• to Melbourne

16
Cross Traffic EstimationMoving Average
17
Signs

• Delay difference
• two queues
• two cases

18
Wide Area Network revisited

• Differentiation btw.
• primary (256K)
• secondary
• (transatlantic)
• bottlenecks
• Secondary traffic
• underestimated
• for Csec./256K

19
Primary Secondary Bottlenecks

• LAN traffic
• controllable
• Separation
• between
• primary (256K)
• secondary (LAN)
• cross traffic

20
TCP Cross Traffic

• Iperf (number of TCP connections 3)
• delay decreases yet we see clusters of packet
  losses (queue size delaypck. loss?)

21
Role of packet sizes

• change packet size

22
Bit/sec. vs. Packet/sec.

• Up 540 Bytes
• Down 60 Bytes
• 540/60 9
• 256Kbps9
• 2.3Mbps
• UDP 1500/3050
• 256K50
• 12.8Mbps
• - 1.5M50
• 75Mbps

23
Secondary Bottleneck

• Difference of CT
• estimate and tcpdump
• CT followed by clusters
• of packet losses

24
Role of TCP

• Secondary CT (necessary
• for losses) picks up
• the periodicity of probe
• traffic!
• Through losses
• Through delay
• When phases match
• Long burst of small
• packets causes losses
• on secondary bottleneck

25
Ongoing work

• Short and long lived TCP connections
• Heterogeneous TCP traffic

26
Conclusions

• Scalable, edge-based tool for on-line network
  analysis, modeling, and measurement
• Scheme for estimating sample path of adaptive
  cross-traffic
• Differentiation between primary and secondary
  bottlenecks two queue model
• Role of TCP (delay variation) and variable packet
  sizes
• Low bit rate denial of service attack?
• QoS in access networks (xDSL)

27

• The End

28
One more picture