Trade-Offs in Automatic Provenance Capture

Abstract

Abstract-Automatic provenance capture from arbitrary applications is a challenging problem. Different approaches to tackle this problem have evolved, most notably a. system-event trace analysis, b. compile-time static instrumentation, and c. taint flow analysis using dynamic binary instrumentation. Each of these approaches offers different trade-offs in terms of the granularity of captured provenance, integration requirements, and runtime overhead. While these aspects have been discussed separately, a systematic and detailed study, quantifying and elucidating them, is still lacking. To fill this gap, we begin to explore these trade-offs for representative examples of these approaches for automatic provenance capture by means of evaluation and measurement. We base our evaluation on UnixBench-a widely used benchmark suite within systems research. We believe this approach will make our results easier to compare with future studies.

Fine-Grained Provenance for Matching & ETL

“…Our implementation enables the UDFs to specify what items in an object or a group were “sub-selected,” while also capturing the relationship to the broader object or group. In contrast to SubZero’s [28] or to event logging [22], [22], [27], our model captures equivalences among computations (including equivalences that hold for particular datatypes and UDFs). PROVision’s query optimizer exploits these to “trace” provenance and aid in troubleshooting.…”

Capturing Provenance for Runtime Data Analysis in Computational Science and Engineering Applications

“…Associating them to results can improve both fine-tuning and data analyses at runtime. Despite the several solutions available for making applications provenance-aware [5][6][7], capturing provenance data in CSE applications is still an open issue. The challenges are mainly related to performance and provenance granularity.…”

“…The challenges are mainly related to performance and provenance granularity. Stamatogiannakis et al [5] evaluated tradeoffs in provenance capture mechanisms. They consider that solutions that are easy to deploy collect provenance in a very fine grain and present a significant overhead, while solutions that are based on function calls present low overhead and granularity is controlled by the code instrumentation.…”

Análise de Hiperparâmetros em Aplicações de Aprendizado Profundo por meio de Dados de Proveniência

“…na análise e escolha dos hiperparâmetros para o treinamento. Existem diversas abordagens de captura de dados de proveniência[Stamatogiannakis et al 2016]. Abordagens de captura automática de dados possuem granularidade muito fina, gerando um overhead significativo na execução de scripts, principalmente os de larga escala.…”