Benchmarking of triple stores scalability for MPSoC trace analysis

International audienceA Multi Processor System-on-Chip (MPSoC) is a complex embedded system used in consumer electronic devices, such as smartphones, tablets and set-top boxes. In order to cope with the complexity of MPSoC architectures, software developers rely on post-mortem trace analysis for application de-bugging or optimization. The traces are explored to localize expected and unexpected programs behaviors. However, the low semantic value of low-level trace events make the trace exploration difficult. We propose to perform trace exploration through an ontology which adds semantics to events and provides a declarative language for querying data. Because traces can be huge, such an ontology contains a large number of instances stored as RDF triples. Because analysts need fast results on classical computer, an efficient system for query answering is preferred. Therefore, saturating , loading and querying those triples pose a scalabil-ity challenge to state-of-the-art knowledge base repositories (KBR). In this paper, we have conducted a benchmark of 7 KBRs: Jena, Sesame-native, Sesame-memory, tdb, sdb, rdf-3x and vertical-mdb, to test their scalability in a non-distributed environment close to analyst environment. We used these KBRs to analyze real traces through VIDECOM, an ontology we designed for trace analysis of applications on MPSoC. Results show that vertical-mdb has a loading rate 3 times faster than the others. It is the only KBR able to saturate the biggest trace of our dataset without exceeding system memory and to run complex queries on it in an acceptable time. Other approaches failed, due to memory limitation or inefficient join implementation