Affordable Access

Publisher Website

Markov prefetching in multi-block particle tracing

Authors
Publisher
Elsevier B.V.
Identifiers
DOI: 10.1016/b978-044453035-6/50006-7
Disciplines
  • Computer Science
  • Mathematics

Abstract

Publisher Summary In general, postprocessing of large-scale data sets is a time-consuming task. The prefetcher is used for predicting data requests in multiblock particle tracing. By the exploitation of auxiliary multiblock topology meta-data, the particle-integration algorithm diminishes the amount of blocks to be loaded. The prefetcher takes care of an optimum overlapping of I/O and computation. The remaining I/O waiting time is reduced considerably when using the runtime evolved Markov prefetcher. The multiblock structure contains connection windows for every neighboring block. The graph's edges are labeled with the transition probability proportional to the size of the connecting windows. To be efficient even when the pos-processing framework is started or new data sets are selected, different approaches of external Markov graph initializations are integrated; this yields a substantial improvement in comparison to uninitialized Markov prefetching or sequential prefetching strategies. One shortcoming of the applied multiblock (MB) meta-data is that it merely considers topology information on time levels separately; this can occasionally result in inadequate Markov initializations for time-variant algorithms. Tree-based indexing structures for fast loading of cells or meta-cells instead of blocks allow for a more granular access.

There are no comments yet on this publication. Be the first to share your thoughts.