Studying and Modeling Real Audio Traffic

1
Studying and ModelingReal Audio Traffic

• John Heidemann
• 4 October 2000
• joint work with Art Mena and Kun-Chan Lan
• USC/ISI

2
Why Study Real Audio?

• streaming mediaa new class of traffic
• today real audio, Microsoft streaming media
  protocol, MBone audio/video
• tomorrow these formats, IP telephony, etc.
• streaming media is different
• to the network
• to simulate

3
Network Characteristics

• delay and jitter sensitive
• interactive traffic inherently sensitive
• (but can trade buffering for sensitivity)
• different transport protocols
• RTP, RTSP, etc.
• different congestion control mechanisms
• potentially very different traffic profile

4
Why Simulation?
from (Jain et al, 1988)
from sims by Xuan Chen
Answer what if?
For protocols, scales, scenarios outside
experimentation. (But depends on good models in
interesting part of space.)
5
Simulation Characteristics

• no source code---cant cheat to get accurate
  model
• opportunity to develop and apply multiple-scale
  modeling techniques

6
Digression Time-Scales
Self-similarity traffic that has similar
characteristics over wide range of
time-scales. Measured traffic is bursty at all
timescales. (Vs. Poisson which becomes smoother
when aggregated.) Multimedia traffic?
Measured
Poisson
7
Digression Why Does Burstiness Matter?

• router queueing, provisioning and packet loss
• aggregation of Poisson sources smoothes away
  burstiness gt enough buffering is possible
• self-similar traffic is always bursty gt cannot
  overprovision buffering gt must learn to tolerate
  or avoid loss other ways

8
Agenda

• Challenges
• Methodology
• Preliminary trace analysis
• Modeling Real Audio traffic
• Related work and future directions

9
Methodology

• Collect and analyze traces of traffic
• almost all real audio traffic
• collected at the server
• up to 18.2 hours/5.9M packets long
• dont have detailed description of content
• Much thanks to Henry Heflich, Gary Nelson, and Ed
  Luczycki of broadcast.com for making this tracing
  possible.

10
Trace Summary
Data and basic analysis reference (Mena
Heidemann, 2000).
11
Basic Statistics

• aggregate data arrival rate, duration, RTT,
  interdeparture times
• per-user data user arrival rate, duration,
  interdeparture times, TCP vs. UDP
• per-flow data packet interdeparture times, flow
  patterns

12
Basic Statistics

• aggregate data arrival rate, duration, RTT,
  interdeparture times
• per-user data user arrival rate, duration,
  interdeparture times, TCP vs. UDP
• per-flow data packet interdeparture times, flow
  patterns

Two time-scales users and protocol
13
Digression Structural Modeling
Structural modeling (term from Willinger et
al.) Hypothesis reproducing the structure of the
application is necessary to accurately reproduce
the traffic

• multiple levels of feedback
• TCP / HTTP / content / user
• across multiple timescales
• lt 1s / 1-10s / 10-100s / gt 100s

14
Digression2 Congestion Reactive Protocols
from (Jacobson, 1988)
Hypothesis TCP Congestion Control algorithms
allow net to scale over huge range of bandwidth,
load, loss. (Floyd Fall, 1999)
15
Basic Statistics Revisited

• aggregate data arrival rate, duration, RTT,
  interdeparture times
• per-user data user arrival rate, duration,
  interdeparture times, TCP vs. UDP
• per-flow data packet interdeparture times, flow
  patterns

Two time-scales users and protocol
16
Agenda

• Challenges
• Methodology
• Preliminary trace analysis
• Modeling Real Audio traffic
• Related work and future directions

17
Per-user Durations
Not heavy tailed (in the mathematical
sense). Hypothesis external factors limit
listening time (program length).
(5.5h) (10.5h) (18.2h)
18
Implications of User Duration

• Very long flows common (compared to TCP)
• User interest (content-specific?) an important
  factor
• gt Could be beneficial to aggregate different
  traffic content on a single server

19
Detailed Flow AnalysisTime-Sequence Plots
At large time-scales appears to be constant-bit
rate traffic (CBR). (Some evidence of congestion
control.)
20
Interdeparture Stats
Mean and quartiles are more complex mean grows
smoothly, quartiles are clustered.
mean
21
Time-Sequence Revisited
More complex behavior at small-time
scales bursts and gaps of 1.8s.
bursts
1.8s inter-burst interval
22
Adjacent Interdepartures
(flow A)
Comparing adjacent interdepartures (for packets
A, B, C plot AB delay vs. BC delay) clustering
at 0.04s and 1.8s.
1.8s
0.04s
previous packet interdeparture (s)
line- speed
packet interdeparture (s)
23
Interdeparture CDFs
(flow A)
(flow B)
cumulative pkts
cumulative pkts
Slower flows have similar pattern at multiples of
1.8s.
packet interdeparture (s)
packet interdeparture (s)
24
Small-Scale Behavior Implications

• Burstiness suggests more frequent loss of
  multiple packets
• affects coding schemes (FEC) and loss repair
• Burstiness will have much different interaction
  with routers, other traffic (than CBR)
• prior work with CBR based simulations
• Other protocol design choices?

25
Agenda

• Challenges
• Methodology
• Preliminary trace analysis
• Modeling Real Audio traffic
• Related work and future directions

26
Basic Model of Real Audio

• Each user picks number of flows
• For each flow, sequentially
• Pick an overall rate , packet size (fixed),
  duration
• While flow is active
• Pick on off duration (on fixed)
• Calculate of packets to send in on-time to
  satisfy rate

( See next slide)
27
Model Limitations
No source code nor content details, so

• Each user picks number of flows
• For each flow, sequentially
• Pick an overall rate , packet size (fixed),
  duration
• While flow is active
• Pick off duration (on fixed)
• Calculate of packets to send in on-time to
  satisfy rate

• Doesnt model content-specific effects

• Doesnt model Real Audio congestion control.

work in progress but, does pretty well (much
better than CBR).
28
Evaluating the Model

• Basic stats
• mean, CDF, etc.
• (not presented here since straightforward)
• Time-variance plots
• Wavelet scaling plots

29
DigressionTime-Variance Plots

• Details from (Willinger Paxson, 1998)
• For X a stationary sequenceX X(i), i ? 1
  (pkts per smallest timescale)

• Define X(m) as the number of packets at
  timescale mX(m)(k) m-1 ? X(i), k(k-1)m1 to
  km

30
DigressionScaling Plots

• Details in (Feldman et al, 1999)
• Rough description E(j) is normalized
  coefficients of the Haar wavelet for X at that
  timescale.
• Intuitive description energy at timescale j
  corresponds to burstiness at that timescale

31
Time-Variance (Model Version 0)
model (version 0)
to generally smooth decay
Basic shape of time-variance graph correct, but
details not correct.
need to look closer.
32
Time-Variance (Model Version 1)
model (version 1)
Problem flows are synchronized---different flows
fire at the same time. Correcting this gives
model version 1.
33
Scaling Plot (Model Version 1)
trace
model (version 1)
Scaling plot also shows close match. (But model
still preliminary)
34
Modeling Observations

• Basic statistics good to get started
• Multi-scale statistics (t-v, scaling) critical to
  gaining model confidence (found multiple
  problems)
• But still work in progress
• need to validate model against fresh traces
• need to model Real Audio feedback/congestion
  control

35
Agenda

• Challenges
• Methodology
• Preliminary trace analysis
• Modeling Real Audio traffic
• Related work and future directions

36
Related Work

• Several similar studies of web traffic
• (Mah, 1997)
• Crovellas SURGE (earlier work Crovella et al,
  1996)
• Ramon Cáceres mmdump (Cáceres et al, 1999)
• looks in the structure of multimedia flows

37
Network Simulationthe Bigger Picture

• Simulation validation
• how do we know were real enough?
• Just-in-time models
• but is this my traffic?

38
Simulation Validation

• Simulation validation an open problem
• How do we know were close enough?
• Question cannot be answered in the abstract
• Better question Are we close enough for question
  X?
• I.e., close enough for relative comparisons of
  queueing disciplines.
• More discussion (Heidemann et al, 2000)

39
Just-in-Time Modeling

• Problem how to simulate my traffic, not 1999
  traffic
• Approach combine parameterized models with
  network measurements to get just-in-time modeling
• Challenges
• right kinds of parameters to models
• integrating measurements from many sources

40
SAMAN Project

• Simulation and modeling in the SAMAN project
• model generation
• just-in-time model measurement and
  parameterization
• simulation scenario pre-analysis and filtering
• network failure analysis (simple and cascading
  failures)

41
Conclusions

• Real Audio traffic is not just CBR, there is
  relevant internal structure
• modeling must consider multiple timescales
• much more work to do
• both Real Audio
• and larger problems in network simulation

42
References (1)

• (Cáceres et al, 1999) R. Cáceres, C. J. Sreenan,
  and J. E. van der Merwe. mmdump--A Tool for
  Monitoring Multimedia Usage on the Internet,
  July, 1999. lthttp//www.research.att.com/ramon/pa
  pers/mmdump.ps.gzgt.
• (Crovella Bestavros, 1996) Mark E. Crovella and
  Azer Bestavros. Self-similarity in World Wide
  Web traffic evidence and possible causes. In
  Proceedings of the ACM SIGMETRICS, pp. 160-169.
  Philadelphia, Pennsylvania, May, 1996.
  lthttp//www.cs.bu.edu/best/res/papers/sigmetrics9
  6.psgt.
• (Floyd Fall, 1999) Sally Floyd and Kevin Fall.
  Promoting the Use of End-to-End Congestion
  Control in the Internet, ACM/IEEE Transactions
  on Networking, V. 7 (N. 4), pp. 458-473, August,
  1999.

43
References (2)

• (Heidemann et al, 2000) John Heidemann, Kevin
  Mills, and Sri Kumar. Expanding Confidence in
  Network Simulation, ISI Research Report 00-522,
  USC/Information Sciences Institute, April, 2000.
  Submitted for publication, IEEE Network.
  lthttp//www.isi.edu/johnh/PAPERS/Heidemann00c.htm
  lgt.
• (Jacobson, 1988) Van Jacobson. Congestion
  Avoidance and Control, in SIGCOMM '88, pp.
  314-329, Stanford, California, August, 1988.
  (Updated version at ltftp//ftp.ee.lbl.gov/papers/c
  ongavoid.ps.Zgt.)
• (Jain et al, 1988) R. Jain, K. Ramakrishnan, and
  D. Chiu. Congestion Avoidance in Computer
  Networks with a Connectionless Network Layer.
  Technical Report N. DEC-TR-506, December, 1988.
  ltftp//ftp.netlab.ohio-state.edu/pub/jain/papers/c
  r5.pdfgt.

44
References (3)

• (Mah, 1997) B. Mah. An Empirical Model of HTTP
  Network Traffic, In Proceedings of the IEEE
  Infocom, pp. 592-600, Kobe, Japan, April, 1997,
  lthttp//www.ca.sandia.gov/bmah/Papers/Http-Infoco
  m.psgt.
• (Mena Heidemann, 2000) Art Mena and John
  Heidemann. An Empirical Study of Real Audio
  Traffic, in IEEE Infocom, pp. 101-110,
  Tel-Aviv, Israel, March, 2000, lthttp//www.isi.edu
  /johnh/PAPERS/Mena00a.htmlgt.
• (Willinger Paxson, 1998) W. Willinger and V.
  Paxson. Where Mathematics meets the Internet,
  Notices of the American Mathematical Society, V.
  45 (N. 8 ), August, 1998. ltftp//ftp.ee.lbl.gov/p
  apers/internet-math-AMS98.ps.gzgt.

45
More information

• http//www.isi.edu/saman/