Soft Scheduling - Scalable Continuous Media Streaming Systems - page 61

4

Soft Scheduling

Most of the existing disk schedulers employed in continuous media servers use worst-case
analysis in capacity dimensioning. As discussed in Chapter 3, this worst-case approach is
relatively simple to implement and yet can provide deterministic performance guarantees,
which is particularly desirable in media streaming applications. Nevertheless, the worst-
case scenario rarely, if ever, occurs in practice and so the disk ends up under-utilized. In this
chapter we investigate a statistical approach to capacity dimensioning - soft scheduling,
where performance guarantee is probabilistic rather than deterministic. By tolerating a very
small probability of capacity overflow, we can significantly increase the disk utilization.
This chapter presents a disk scheduler supporting soft scheduling, a number of performance
enhancement techniques, and evaluates the potential performance gains.

4.1 Introduction
For media servers serving stored data (as opposed to data captured in real time), the disk
scheduler plays a vital role in providing glitch-free services to the end-users. Many excellent
disk schedulers have been proposed in the literature and were reviewed in Chapter 3, including
the SCAN [5], Circular-SCAN (CSCAN) [5], and the more general Group Sweeping Scheme
(GSS) [4] schedulers. Using worst-case dimensioning techniques, these schedulers are simple
to implement and are capable of providing guaranteed performance.
Nevertheless, using worst-case dimensioning techniques - hard scheduling - also have
shortcomings. As the usable disk capacity (in terms of maximum number of concurrent streams
that can be supported) is dimensioned according to worst-case scenarios, the disk will often be
under-utilized during actual operation. Moreover, most modern disk drives employ disk zoning
to improve disk capacity. Zoning divides the disk surface into multiple zones, where each zone
has a number of consecutive cylinders having the same track size (in numbers of sectors per
track). As the disk rotates with a constant angular velocity, outer zones can be allocated more
sectors per track than inner zones. Therefore, one side-effect is that outer zones will have
substantially higher transfer rate than inner zones. If we adopt hard scheduling, then we would
need to dimension the disk streaming capacity according to the (lowest) transfer rate of the
innermost zone, and hence sacrifice additional disk bandwidth available in the outer zones.