Performance and reliability study for distributed video servers: mirroring or parity?

Abstract

We study video server performance and reliability. We classify several reliability schemes based on the redundancy technique used (mirroring vs. parity) and of the distribution granularity of redundant data. Then, we propose for each scheme its adequate data layout. To calculate the server reliability we apply discrete modeling based on Markov chains. Further we focus on the trade-off between achieving high reliability and low per stream cost. Our results show that, in contrast to intuition, for the same degree of reliability mirroring-based schemes always outperform parity-based schemes in terms of per stream cost and also restart latency after disk failure. Our results also show that a mirroring scheme that copies original data of a single disk onto a subset of all disks significantly improves the server reliability and slightly increases the per stream cost as compared to the classical interleaved mirroring scheme.