Randomized data allocation for real-time disk I/O

Continuous media such as video or audio from databases that are disk resident require real-time disk I/O support. Video on demand systems have been widely studied and most proposed designs take advantage of the (largely) predictable nature of the I/O stream to provide both guaranteed upper bounds on delay and reasonably high resource utilizations. For disk based 3D interactive systems the problem is very different since the user's actions determine the future deadlines for model data at the display. The data layout we propose and evaluate in this paper abandons the idea of a careful layout of data for a completely randomized layout. We consider large, multidisk systems in which the 3D model data is partitioned into granules which are the logical unit of data that gets transferred an and out of the active scene graph. Granules form parity groups as is familiar from disk arrays. In the proposed system, the use of redundancy in the form of parity is used under normal conditions for load balancing (since for a parity group of size G, any G-1 blocks can be read to obtain the data). Preliminary simulation results suggest that for moderate increase in storage for parity, one can obtain high degree of disk bandwidth utilization combined with a "nearly guaranteed" maximum delay, i.e., with probability approaching 1. The simulations show the nature of the tradeoffs between utilization and the probability that deadlines are met.