A Hierarchical Characterization of a Live Streaming Media Workload

1
A Hierarchical Characterization of a Live
Streaming Media Workload

• IEEE/ACM Trans. Networking, Feb. 2006
• Eveline Veloso, Virgílio Almeida, Wagner Meira,
  Jr., Azer Bestavros, and Shudong Jin

2
Motivation

• The characteristics of live media and stored
  media are different.
• Stored media object user driven
• Be directly influenced by user preferences
• Live media object content driven
• Be directly influenced by aspects related to the
  nature of the object

A Traffic Characterization of Popular On-Line
Games http//vc.cs.nthu.edu.tw/home/paper/codfile
s/clchan/200507191203/A_Traffic_Characterization_o
f_Popular_On-Line_Games.ppt
3
Basic statistics of the trace used in this paper
Microsoft Media Server
7 Kbps 18 Kbps 32 Kbps 57 Kbps
stream 1
48 different cameras

stream 2
4
Characterization hierarchy

• Client layer
• Session layer
• The interval of time during which the client is
  actively engaged in requesting live streams that
  are part of the same service such that the
  duration of any period of no transfers between
  the server and the client does not exceed a
  preset threshold Toff.
• Transfer layer
• In session ON time
• During transfer ON time, a client is served one
  or more live streams.
• Transfer OFF times correspond loosely to think
  times.

5
Relationship between client activities and ON/OFF
times
6
Client layer characteristics

• Topological and geographical distribution of
  client population
• Zipf-like distribution
• Most requests are issued from a few regions
• Client concurrency profile
• Client interarrival times
• Client interest profile

7
Client diversity IP addresses over ASs
Autonomous System (AS) the unit of router
policy, either a single network or a group of
networks that is controlled by a common network
administrator
8
Client diversity transfers over ASs
9
Client diversity transfers over countries
10
Client layer characteristics

• Topological and geographical distribution of
  client population
• Client concurrency profile
• Periodic behavior
• Client interarrival times
• Client interest profile

11
Cumulative distribution of number of active
clients
(cumulative)
12
Temporal behavior of number of active clients
over entire trace
13
Temporal behavior of number of active clients
daily
Weekend have slightly higher clients than weekdays
14
Temporal behavior of number of active clients
hourly
15
Client layer characteristics

• Topological and geographical distribution of
  client population
• Client concurrency profile
• Client interarrival times
• Pareto distribution
• Piece-wise-stationary Poisson process
• Client interest profile

16
Client interarrival times frequency

• What is the unit of frequency?
• It might be
• instance/second (x)
• instance/request (?)
• percentage (?)

17
Client interarrival times CCDF
CCDF Complementary Cumulative Distribution
Function
18
Discuss

• The client arrival process is not stationary in
  that it is highly dependent on time.
• It is natural to assume that over a very short
  time interval, such a process would be
  stationary, and may indeed be Poisson.
• Piece-wise-stationary Poisson arrival
• 15 min.

19
Client interarrival times piece-wise-stationary
Poisson process
20
Client layer characteristics

• Topological and geographical distribution of
  client population
• Client concurrency profile
• Client interarrival times
• Client interest profile
• Characterizing live content popularity is not
  meaningful ? characterizing the interest of a
  client in the live content is more appropriate
• Zipf-like distribution
• Most requests are issued from a few clients

21
Client interest profile client rank v.s.
transfer frequency
Rank number of transfers for that client
22
Client interest profile client rank v.s. session
frequency
Rank number of sessions for that client
23
Session layer characteristics

• Number of sessions
• Threshold Toff
• Session ON time
• Session OFF time
• Transfers per session
• Interarrivals of session transfers

24
Relationship between number of sessions and Toff
3600
25
Session layer characteristics

• Number of sessions
• Session ON time
• Lognormal distribution
• Session OFF time
• Transfers per session
• Interarrivals of session transfers

26
Distribution of session ON times
27
Session layer characteristics

• Number of sessions
• Session ON time
• Session OFF time
• Exponential distribution
• Transfers per session
• Interarrivals of session transfers

28
Distribution of session OFF times
29
Session layer characteristics

• Number of sessions
• Session ON time
• Session OFF time
• Transfers per session
• Pareto distribution
• Interarrivals of session transfers

30
Number of transfers per session frequency
31
Number of transfers per session CCDF
32
Session layer characteristics

• Number of sessions
• Session ON time
• Session OFF time
• Transfers per session
• Interarrivals of session transfers
• Lognormal distribution

33
Session transfer interarrivals frequency
34
Transfer layer characteristics

• Number of concurrent transfers
• Exponential distribution
• Transfer length and client stickiness
• Transfer interarrivals
• Transfer bandwidth

35
Concurrent transfers over all sessions
(cumulative)
36
Transfer layer characteristics

• Number of concurrent transfers
• Transfer length and client stickiness
• Lognormal distribution
• The long tail of the transfer length distribution
  is due to the clients willingness to stick to
  the live stream.
• Transfer interarrivals
• Transfer bandwidth

37
Transfer lengths
38
Transfer layer characteristics

• Number of concurrent transfers
• Transfer length and client stickiness
• Transfer interarrivals
• Like client arrivals
• Pareto distribution
• Transfer bandwidth

39
Transfer interarrival times
40
Temporal behavior of transfer interarrival times
over entire trace
41
Temporal behavior of transfer interarrival times
daily
Weekends have lower average interarrivals than
weekdays (but more clients)
? Due to channel browsing?
42
Temporal behavior of transfer interarrival times
hourly
43
Transfer layer characteristics

• Number of concurrent transfers
• Transfer length and client stickiness
• Transfer interarrivals
• Transfer bandwidth
• Client-bound bandwidth
• Congestion-bound bandwidth

44
Aggregate bandwidth
45
Frequency distributions of transfer bandwidth
client 58.6 Kbps 32.5 Kbps 17.6 Kbps 6.87 Kbps
congestion
46
Across multiple live media workloads

• Another live streaming server for a news and
  sports radio station
• The differences of two live streaming services
• Client interarrival times
• Session transfer interarrival times
• Transfer interarrival times
• These differences are due to the different
  interactions between clients and live streams in
  the workloads.

47
Summary of the characteristics of the Reality
Show and News and Sports