Information for "ANAPSID: An Adaptive Query Processing Engine for SPARQL Endpoints"

We report on runtime performance, which corresponds to the user time produced by the _ 􀀀_ command of the Unix operation system. Experiments in RDF-3X were run in both cold and warm caches; to run cold cache, we executed the same query five times by dropping the cache just before running the first iteration of the query. Each query executed by ANAPSID and SPARQL endpoints was run ten times, and we report on the average time.

We empirically analyze the performance of the proposed query processing techniques and report on the execution time of plans comprised of ANAPSID operators versus queries posed against SPARQL endpoints, and state-of-the-art RDF engines.

Three sets of queries were considered (Table of Figure 5(b)); each sub-query was executed as a query against its corresponding endpoint. Benchmark 1 is a set of 10 queries against LinkedSensorData-blizzards; each query can be grouped into 4 or 5 sub-queries. Benchmark 2 is a set of 10 queries over linkedCT with 3 or 4 subqueries. Benchmark 3 is a set of 10 queries with 4 or 5 sub-queries executed against linkedCT and DBPedia endpoints.

Experiments were executed on a Linux CentOS machine with an Intel Pentium Core2 Duo 3.0 GHz and 8GB RAM.

Following the design rules of Linked Data, the number of available SPARQL endpoints that support remote query processing is quickly growing; however, because of the lack of adaptivity, query executions may frequently be unsuccessful. First, fixed plans identified following the traditional optimize-then execute paradigm, may timeout as a consequence of endpoint availability. Second, because blocking operators are usually implemented, endpoint query engines are not able to incrementally produce results, and may become blocked if data sources stop sending data. We present ANAPSID, an adaptive query engine for SPARQL endpoints that adapts query execution schedulers to data availability and run-time conditions. ANAPSID provides physical SPARQL operators that detect when a source becomes blocked or data traÆc is bursty, and opportunistically, the operators produce results as quickly as data arrives from the sources. Additionally, ANAPSID operators implement main memory replacement policies to move previously computed matches to secondary memory avoiding duplicates. We compared ANAPSID performance with respect to RDF stores and endpoints, and observed that ANAPSID speeds up execution time, in some cases, in more than one order of magnitude.

In the future we plan to extend ANAPSID with more powerful and lightweight operators like Eddy and MJoin, which are able to route received responses through different operators, and adapt the execution to unpredictable delays by changing the order in which each data item is routed.

Lightweight wrappers translate SPARQL sub-queries into calls to endpoints as well as convert endpoint answers into ANAPSID internal structures. Mediators maintain information about endpoint capabilities, statistics that describe their content and performance, and the ontology used to describe the data accessible through the endpoint.

The Local As View (LAV) approach is used to describe endpoints in terms of the ontology used in the endpoint dataset. Further, mediators implement query rewriting techniques, decompose queries into sub-queries against the endpoints, and gather data retrieved from the contacted endpoints. Currently, only SPARQL queries comprised of joins are considered; however, the rewriting techniques have been extended to consider all SPARQL operators,

but this piece of work is out of the scope of this paper. Finally, mediators hide delays, and produce answers as quickly as data arrives.