Continuous Media Storage and Retrieval - Scalable Continuous Media Streaming Systems - page 52

1

0.9

0.8

0.7

0.6

0.5

0 2 4 6 8 10

Media Object Index

Figure 3.7 Relative popularity of media objects following the Zipf law with c = 0 . 237

However, this data partition approach has a significant shortcoming. Specifically, in most
applications (e.g., entertainment videos) the popularity of media objects can vary significantly.
Previous studies [2-3] using data collected from video rentals, for example, suggested that the
popularity in those scenarios follows the Zipf law [4]:

1
i c

P i ∝ (3.9)

where is the popularity of media object i , sorted in decreasing order

of popularity; and c is a constant characterizing the skewness of the popularities. A value of
c

{ P i | i = 1, 2, . . . , M }

= =

P j ∀

i , j ), and increasing values of c represent
increasing skewness. Figure 3.7 shows the relative popularity for a 10-media system following
the Zipf law with c

0 represents uniform popularity (i.e., P i

0.237. The difference between the more popular media objects and the
less popular media objects can be quite substantial.
The implication of the Zipf-distributed popularity is that a small portion of the media objects
often account for the majority of the user requests. Consequently, the disks storing the popular
media objects will have a much higher load than the disks storing unpopular media objects.
Researchers have long recognized this load imbalance problem and have proposed various
solutions, e.g., careful placement ofmedia objects to even out the load of the disks, or replicating
the popular objects on multiple disks for load sharing. Interested readers are referred to the
literature for more detail of these approaches [5-7].
At the other extreme, if we replicate all media objects onto all disks in the server, then the
load balance problem can be eliminated altogether. However, this approach is only suitable
for systems with a small set of media objects, where the extra storage incurred in repli-
cation is available and is cost effective. For systems serving a library of, say, hundreds or
even thousands of feature-length movies, the multiplied storage costs will quickly become
prohibitive.

=