Algorithms for Server Rebuild - Scalable Continuous Media Streaming Systems - page 252

no human intervention is required and transparent in the sense that existing streaming sessions
are not affected in any way.
For simplicity, we consider only single-server failure in the rest of the chapter but the analysis
can readily be extended to cover multiple-simultaneous server failures. Let N S be the number
of servers in the system, U be the storage capacity of each server, and S S be the effective server
transfer capacity. The effective server transfer capacity S S is the maximum data rate at which
a server can transfer data to/from the network. That is, if a server has an effective transfer
capacity of S S = 600Mbps, and the server is sending data at a rate of 200Mbps, then the server

will only be able to receive data up to a rate of 400Mbps.

14.2 Disk Migration
It is possible that a server failure may not be caused by (unrecoverable) disk failures. In this
case, we may simply disconnect the disk subsystem from the failed server and connect it to the
spare server. This process is simple but requires human intervention to first identify the source
of the failure, and then to migrate the disk subsystem to the spare server. It is also possible to
use an electronic wiring network to connect a disk subsystem to two or more servers, including
a spare server (e.g., using twin-tailed disks). The migration of disk units can then be done
electronically.

14.3 Reloading Data from Back-up
If the media data in the system are also stored in a back-up storage, then we can simply load the
required media data from the back-up media to the spare server. This can be done automatically
without affecting the remaining active servers. The rebuild rate will then be equal to the transfer
capacity of the back-up device or the spare server, whichever is smaller.
While this scheme is simple and efficient, the mechanisms needed to automate the process
(selection and loading of media data) are not inexpensive, such as large juke-boxes or robotic
tape libraries. Moreover, management of the back-up data will be more complex as old media
data are removed and new ones uploaded to the servers from time to time. Finally, if the media
data streams are backed up sequentially, then considerable searching will be needed when
loading the media data into a spare server due to the server stripping scheme used.
For comparison with other rebuild algorithms, we assume the back-up device has infinite
throughput and capacity. Therefore, the rebuild rate is bounded by the server throughput, or

R reload = S S (14.1)

which is also the maximum rate that can be achieved by any data rebuild algorithm.

14.4 Baseline Rebuild
The twomethods discussed earlier both require extra hardware/software and/or human operator
support. This section presents the first automatic algorithm - baseline rebuild, that requires
neither extra hardware/software support nor human intervention. The principle is to utilize the
idle capacities in the remaining servers to send data/redundant units to the spare server, which