INTEROP 2016 Workshop

Abstract

Distributed compute resources are necessary for compute-intensive information extraction tasks processing large collections of heterogeneous documents (e.g. patents). For optimal usage of such resources, the breaking down of complex workflows and document sets into independent smaller units is required. The UIMA framework facilitates implementation of modular workflows, which represents an ideal structure for parallel processing. Although UIMA-AS already includes parallel processing functionality, we tested two other approaches for distributed computing. First, we integrated UIMA workflows into the grid middleware UNICORE, which allows high performance distributed computing using control structures like loops or branching. While good distribution management and performance is a key requirement, portability, flexibility, interoperability and easy usage are also desired features. Therefore, as an alternative, we deployed UIMA applications in a microservice architecture that supports all these aspects. We show that UIMA applications are well-suited to run in a microservice architecture while using an event-based asynchronous communication method. These applications communicate through a standardized STOMP message protocol via a message broker. Within this architecture, new applications can easily be integrated, portability is simple, and interoperability also with non-UIMA components is given. Markedly, a first test shows an increase of processing performance in comparison to the UNICORE-based HPC solution.

Keywords: UIMA, Microservice, Text Mining, Distributed Computing, Interoperability