Open-Loop Algorithms - Scalable Continuous Media Streaming Systems - page 319

30

: Skyscraper Broadcasting
: Greedy Disk-Conserving
Broadcasting
: Staircase Data Broadcasting
: Poly-harmonic Broadcasting
: Consonant Broadcasting
: Pagoda Broadcasting
:

SB
GDB3

CB, m= 1
CB, m= 2
CB, m= 4
CB, m= 16

SDB
PHB
CB
PB
LB

20

PHB, m= 16
( without client
access bandwidth
constraint )
PHB, m= 1
( without client
access bandwidth
constraint )
LB

: Lower Bound

SDB

GDB3

SB

10

PB

0

2 b 4 b 8 b 12 b 16 b 20 b

Network bandwidth (multiples of video bit-rate b ) t- )

Figure 18.3 Start-up latency versus network bandwidth at small latency range

and PB and thus the results are not directly comparable. Nevertheless, this result shows the
performance loss due to limited client access bandwidth.
Except for PHB and PB, CB achieves the lowest start-up latency. This is true even for m 1,

which generates the least number of media segments (and hence system complexity) given
the same system parameters. Increasing m can further reduce the start-up latency but at the
expense of higher system complexity. For a network bandwidth of 5 b , CB with m

=

= 4 achieves

start-up latency 81%, 74%, and 60% lower than SB, GCB, and SDB respectively.
Figure 18.3 compares the start-up latency of the broadcasting schemes for larger network
bandwidth ranging from 2 b to 20 b . At this range the start-up latency is reduced to seconds,
well within the response time required in an on-demand media streaming service. Again the
observation is consistent with the results in Figure 18.2, showing that CB achieving the lowest
start-up latency. For example, with a network bandwidth of 10 b , CB with m 4 can achieve

a start-up latency of only 2 seconds, which is 96%, 95%, and 72% lower than SB, GCB, and
SDB respectively.

=

18.5.2 Start-up Latency versus Client Access Bandwidth
Figure 18.4 plots the start-up latency versus the client access bandwidth ranging from 2 b to
6 b , where b is the media bit-rate. The network bandwidth is equal to 6 b . There are three
observations.
First, CB clearly outperforms the other schemes, especiallywhen the client access bandwidth
is low. This is a significant property as the client access network in practice will likely have
substantially lower bandwidth than backbone networks. Second, the performances, of PHB
and PB degrade significantly when the client access bandwidth is reduced. This is because both