Efficient Server Design for Hybrid Multicast Streaming - Scalable Continuous Media Streaming Systems - page 356

former simply cannot be applied to a SS-VoD system as the video placement policy and I/O
scheduler typically do not allow random data retrievals. The latter, on the other hand, can still
be applied to a SS-VoD system but the efficiency will be sub-optimal as the static channels'
periodic retrieval patterns are not exploited to increase retrieval efficiency.
To support both static batching channels and dynamic patching channels, we develop a new
video placement policy and retrieval algorithm that can support random data retrievals, while
still be able to exploit the periodic retrieval pattern to increase disk efficiency. To tackle the
disk-zoning problem, we develop a Weighted Group Segment Pairing Scheme (WSGP) based
on SGP [8] for video placement. By pairing an outer zone with an inner zone and allocating
video data according to the zone's storage capacity, we can achieve increased disk utilization
without sacrificing disk storage capacity.
Compared to conventional server designs using round-based schedulers, this efficient server
design can increase the system capacity by as much as 60% with the same buffer requirement.
This chapter presents details of this new server design, derives a performance model, and
analyzes it using numerical results.
20.2 Background
Video server design has been studied extensively in the literature. Gemmell et al . [2] provide an
excellent overview of the area, explaining the key challenges and reviewing existing solutions.
Most of the existing server designs for TVoD systems are centered around round-based algo-
rithms, such as the CSCAN scheduler [7] and the Grouped Sweeping Scheme (GSS) scheduler
[6]. Common among these round-based schedulers is the assumption that there is no correla-
tion between the active video streams, i.e., the video streams play back video independently
at arbitrary schedules. While this assumption is valid and necessary for TVoD systems, it is
sub-optimal for a SS-VoD system where some of the multicast video streams are prescheduled
and thus have a fixed temporal relation with one another.
At the other extreme of the spectrum is NVoD systems where all video streams are broad-
cast repeatedly and periodically in a fixed schedule. Armed with complete knowledge of the
broadcasting schedules, one can then design an optimized video placement and disk retrieval
scheme to increase disk efficiency. The principle is to take advantage of the fixed temporal
relation between broadcast video streams and place video data in an interleaved manner so
that the server can retrieve video data continuously with minimal disk seeking.
For example, Chen and Manu [8] proposed a video placement policy called Segment Group
Pairing (SGP) to allocate video data in zone-bit-recording (ZBR) disk for NVoD servers. With
SGP, data blocks that are to be retrieved in the same round are divided evenly into two groups.
The first group is stored continuously in the outer zone while the second group is stored
continuously in the inner zone. This continuous placement reduces seeking overhead in data
retrieval. In a service round, the disk head will first retrieve the group of blocks located in the
outer zone and then seek the inner zone to retrieve the remaining data blocks. This scheme
enables the data rates of both zones to be averaged and thus results in a higher deterministic
disk throughput.
Nevertheless, this algorithm did not account for capacity differences among different zones.
In particular, inner zones usually have lower capacity compared to outer zones. As a result, SGP
will likely fill up the inner zone before the outer zone is fully utilized and thus the remaining