lee1997
lee1997
Abstract
This paper discusses a novel data placement scheme which optimizes the storage utilization of a
NVOD system. The scheme is most distinctive in the following two aspects:
1.It considers the file blocks placement ofprogramsfeatured different number NVOD channels.
2. The file blocks grouping scheme optimizes the storage utilization of a NVOD system.
Keyword
near video on demand, NVOD, VOD, disk layout
Introduction
Unlike Video On Demand system that requires one channel per user, Near Video On Demand
(NVOD) system broadcasts the program periodically that the time between two successive play-
ings is much shorter (say 15 to 30 minutes) than the program playing time (usually between
90 minutes to 2 hours) and each playing uses a different channel. Consequently, NVOD system
only needs a less number of channels than VOD system as it allows users to share the channel
with the penalty of short start latency and no VCR-like functions. Generally, because NVOD
is communication-intensive, it makes best sense to employ this service in conjunction with video
compression technology.
For 10 two-hour programs,
1. if program begins every 15 minutes, then 80 channels are required.
2. if program begins every 30 minutes, then 40 channels are required.
The mathematics here is quite simple. A program that lasts two hours will utilize 8 channels
if it is to be shown every 15 minutes. This results from the fact that there are 8 15-minute
intervals in a 2-hour period. Similarly, if the same program is shown every 30 minutes, it will
require 4 channels as there are 4 30-minute intervals in a 2-hour period. Thus, for some popular
Related Works
In a NVOD system, the video programs start at every fixed time interval according to their pop-
ularity. If a NVOD program broadcasts on m channels, the system has to read m blocks at every
access period. Traditionally, every neighboring pair of those file blocks are located at the (Total
file blocks of this program)/m distance apart. Rangan and Vin[4] propose a disk layout scheme for
different playback rates of programs. Two questions are answered in the constrained allocation of
blocks of a media strand: (1) What should the size of the blocks(i.e. the granularity) be ? and (2)
What should the separation between successive blocks(i.e. the scattering parameter) of a strand
be? The guiding factor in determining the block size and separation is the requirement of contin-
uous retrieval at different strand rates. The relative values of block size (M) and the scattering
parameter (G) for each strand must satisfy the real-time retrieval and playback requirements.
The pair (M,G) defines the storage pattern of a strand, and the strand consists of repetitions of
its storage pattern. Notice that the larger M is , so is G. It is flexible to merge different strands,
but more buffer are required due to the larger service cycle time.
At the first glance, this scheme is suitable for a NVOD storage system design. But the scheme
assumes that the lengths of strands are unbounded. In their scheme, the retrieval data size of
each strand in a service cycle is proportional to their playback rates. If lengths of strands are
the same, data of the higher playback rate strand will have more data interleaved with the lower
playback rate strand in each service cycle. Sooner or later, while the higher playback rate strand
is placed in the disk, the lower playback rate strand still has data left to be placed. Without any
choice, the blocks at the end of the disk will contain data from the lower playback rate strand
only. This placement pattern will induce many seeks overhead as the disk needs to retrieve from
different blocks in order to supply data from both strands.
Since contemporary SCSI disks usually have read-ahead caching capability, it is essential to
access disk blocks sequentially to maximize I/O throughput. In this paper, a model is proposed
that relates SCSI read-ahead cache capability and novel data placement scheme to maximize the
storage utilization of multiple NVOD programs.
391
File blocks
,/ Region 0
jR
I, File blocks
'I, Region 1
I jR+1
partitioned into
R regions
jR+(R-1)
Figure 1: The general scheme of the proposed disk layout. File blocks of a program with index j x R + lc are
placed in region k, j is the integer runs from 0.
R is the number of regions into which the disk is partitioned, j is the integer runs from 0, and
k is the index of the regions and runs from 0 to R — 1. In the proposed disk layout design, the
basic storage unit is a disk track. That is, a file block(or retrieval unit) comprises one or more
disk tracks. The reason behind adopting this practice is to eliminate rotation latency during disk
access, which is one of the major overheads of disk accesses. If the disk features on-arrival read-
ahead and file blocks always start and end at disk track boundaries, then disk rotation latency
can be completely eliminated.
During the access operation, the system retrieves and buffers one file block for each channel
when the disk head scans across one partition region. Later in time, when the disk head moves
to next region, the system broadcasts the buffered data to the clients and buffers the incoming
392
Tseek
(RS1SZ
(b) space constraint:n x S RS
393
seekf 'd—
ime — 3.24+O.4OO/i, if d<383
8.00 + O.008d, if d 383,
where d is the distance of disk head movement in number of cylinders and the unit of time is
millisecond. Table 2 shows the NVOD system specifications of our design example. There are
two programs, A and B, in our storage system.
We assume the hard disk is partitioned into 8 regions, and the round_access_time from region
0 to region 7 to provide the NVOD service is 4 second. Notice that the selection of 4 seconds
is arbitrary and this number determines the size of buffer needed and the number of channel
supported. The periodA of program A is 12 and the periodB of program B is 19 according to our
definitions. Figure 2 shows the file blocks grouping of the 3th region of the program A. Program
B is also placed in the disk similarly. Table 3 shows that the file blocks in a region are placed
without grouping scheme, the system can not provide the sufficient NVOD channels for program
A and B, but with grouping scheme, the system can. The reason is obvious, as the disk seeking
time overhead is reduced with grouping scheme.
394
- - the9ithblockofprogram A
the 187th block of program A
File blocks
grOuplngOf
program A
Figure 2: An example of the proposed file blocks grouping in a region-based disk system. The periodA is equal to
12
395
Program A Program B
Media type MPEG-i MPEG-i
Demand bandwidth per channel 1.5Mbps 1.5Mbps
Program duration 5 minutes 5 minutes
Periodically start interval 48 seconds 76 seconds
Random Grouping
placement placement
scheme scheme
Total retrieval and seek time
510.82 msec 475.02 msec
in a region
Meet t e real-time criteria of
No Yes
NVO[) service
Table 3: Comparisons of random placement scheme and grouping placement scheme. Notice that the time bound
for real-time criteria of disk access operations in one region is 500 msec
396
1 .It considers the file blocks placement of programs featured different number of NVOD channels.
2.The file blocks grouping scheme optimizes the storage utilization of a NVOD system.
Currently, we only consider an optimal disk layout for NVOD service on one disk system.
The main challenge for a multiple disks system supporting NVOD service is to decide which
programs should be placed together on the same disks to meet the programs' popularity bandwidth
requirement and space size requirement with minimum storage cost. In[2] [3] , we propose an
optimization procedure. The proposed procedure determines, with respect to overall system cost,
which programs should be placed together on the same disks for supporting VOD service. Based
on the decision, the designer then can figure out the appropriate configuration of the multiple
disks system. The optimization target is to minimize the overall cost of the storage system. We
will apply the optimization procedure in NVOD system in the near future.
References
[1] Yen-Jen Oyang. A tight upper bound of the lumped disk seek time for the scan disk scheduling
policy. Information Processing Letters, 54, 1995.
[2] Meng-Huang Lee, Chun-Hung Wen, Chih-Yuan Cheng, Fu-Ching Wang, and Yen-Jen Oyang.
Storage Hierarchy Design in Multimedia On-Demand Servers. Proceedings of the 1996
IS&T/SPIE Symp. on Electronic Imaging: Science and Technology, Conference on Stroage
and Retrieval for Still Image and Video Database IV, San Jose, CA. Jan. 1996.
[3] Meng-Huang Lee, Meng Chang Chen, Jan-Ming Ho, Ming-Tat Ko, and Yen-Jen Oyang. De-
signing a ReadWrite Multimedia On Demand File Server. Proceedings of the 1996 IS&T/SPIE
Symp. on Electronic Imaging: Science and Technology, Conference on Very High Resolution
and Quality Imaging, San Jose, CA. Jan. 1996.
[4] P. Venkat Rangan and Harrick M. Vin. Efficient Storage Techniques for Digital Continuous
Multimedia. IEEE Transactions on Knowledge and Data Engineering, Vol. 5, No. 4, 1993.
[5] Chris Ruemmier and John Wikles. An Introduction to Disk Drive Modeling. IEEE Computer,
Vol. 27, No. 3, 1994.
397