Author
Weaver, Jesse
Other Contributors
Hendler, James A.; Carothers, Christopher D.; Fox, Peter A.; Mizell, David W.;
Date Issued
2013-05
Subject
Computer science
Degree
PhD;
Terms of Use
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute, Troy, NY. Copyright of original work retained by author.;
Abstract
The problem is first considered from the perspective of the operational semantics of inference with production rules. The question is asked, under what conditions is embarrassingly parallel inference guaranteed to be correct? Sufficient conditions are determined and proven at both a fine-grained level close to the basic operational semantics and a more coarse-grained level that applies directly to rules. The conditions are placed on the relationship between rules and distribution schemes, that is, the way in which data is assigned to processors.; Finally, an evaluation is performed that tests these theoretical findings for restricted versions of RDFS and OWL2RL inference on two large, well-known datasets exceeding a billion triples: LUBM10K and BTC2012. The LUBM10K dataset represents an optimistic case, meaning that if performance is poor with LUBM10K, then it will likely be poor on many datasets. On the other hand, the BTC2012 dataset represents a pessimistic case, meaning that if performance is good with BTC2012, then it is likely that performance will be good with other datasets. While the usual scalability metrics are used (speedup, efficiency, etc.), the Karp-Flatt metric reveals that inference is almost entirely parallel for LUBM10K data, demonstrating the practical feasibility of the theoretical findings. However, for BTC2012, it must be ensured that there is sufficient memory and load-balancing to achieve this high level of scalability on distributed memory architectures. Regardless, for feasible cases, very low times are achieved for LUBM10K (seconds) and BTC2012 (minutes).; Then, a special class of distribution schemes is considered called replication schemes. Replication schemes require that individual data either be replicated to all processors or placed arbitrarily on some processor(s). The aforementioned conditions are then reformulated to consider replication schemes which reveals that testing the conditions for replication schemes is reducible to satisfiability (SAT), and not only SAT but 2SAT. An augmented version of this reduction which is a reduction to 3SAT also accounts for the possibility to eliminate some rules in order to improve parallelization. These reductions along with a proposed methodology for restricting rules are used to derive restricted versions of the RDFS and OWL2RL rules that are amenable to parallel inference.; This thesis considers the problem of scaling rule-based inference to large quantities of RDF data found on the Semantic Web. The general approach is one of data parallelism, that is, dividing data among processors such that the collective results of each processor's individual inference is the same as though inference was performed sequentially. In this way, theoretically speaking, more processors can be added to accommodate more data.;
Description
May 2013; School of Science
Department
Dept. of Computer Science;
Publisher
Rensselaer Polytechnic Institute, Troy, NY
Relationships
Rensselaer Theses and Dissertations Online Collection;
Access
Restricted to current Rensselaer faculty, staff and students. Access inquiries may be directed to the Rensselaer Libraries.;