Reliable and Fault-Tolerant Storage Systems - Scalable Continuous Media Streaming Systems - page 96

Now consider the retrieval process. Given u active streams, the number of playback buffers
required is given by

B p ( ρ ) = u (2 N D − 1) Q (5.27)

For the rebuild process, the buffer requirement when there are u active streams is given by

B r = ( K − u ) N D Q (5.28)

Hence, the combined buffer requirement with buffer sharing is then given by

B share = u (2 N D − 1) Q + ( K − u ) N D Q

= (( N D − 1) u + KN D ) Q

≤ (2 N D − 1) KQ ∀ u (5.29)

which surprisingly just equals the buffer requirement for the retrieval process. Therefore, with
the proposed buffer-sharing scheme one can use the same round-based disk scheduler for both
reading and writing without any additional buffer.

5.5 Track-Based Rebuild
Most, if not all, modern disk drives employ zoning to increase disk storage capacity as discussed
in Section 3.6. A side-effect of zoning is the variation in track size. In particular, inner tracks
have less storage capacity than outer tracks. Due to this uneven track-size problem, the disk
scheduler in most continuous-media server designs retrieves media units in fixed-size blocks
instead of tracks. While reading the entire track can eliminate the rotational latency, the amount
of buffer required to maintain a balanced disk schedule is often prohibitively large [6].
Unlike the retrieval process, the rebuild process is a non-real-time process that does not
require data retrieval at a constant rate. Consequently, we can employ track-based retrieval for
the rebuild process to improve rebuild performance and keep using block-based retrieval for
the streaming process to maintain low buffer requirement.

5.5.1 Rebuild Algorithm
Figure 5.5 depicts the track-based rebuild algorithm. In reading data from the data disks,
playback data are still retrieved in fixed-size blocks but rebuild data are retrieved in tracks.
This allows the elimination of rotational latency during rebuild data retrieval.
Specifically, in block-based retrieval, the disk head must wait for the required disk sec-
tor to rotate to beneath the disk head before data transfer can begin. In the worst case
where the required sector has just passed over the disk head after searching is com-
plete, the disk will have to wait for one complete round of rotation before beginning data
transfer.
By contrast, under track-based retrieval, the disk head can start data transfer as soon as
seeking is completed because the entire track is to be retrieved. Clearly, the reading time is