ParBiBit: Parallel tool for binary biclustering on modern distributed-memory systems

Abstract

Biclustering techniques are gaining attention in the analysis of large-scale datasets as they identify two-dimensional submatrices where both rows and columns are correlated. In this work we present ParBiBit, a parallel tool to accelerate the search of interesting biclusters on binary datasets, which are very popular on different fields such as genetics, marketing or text mining. It is based on the state-of-the-art sequential Java tool BiBit, which has been proved accurate by several studies, especially on scenarios that result on many large biclusters. ParBiBit uses the same methodology as BiBit (grouping the binary information into patterns) and provides the same results. Nevertheless, our tool significantly improves performance thanks to an efficient implementation based on C++11 that includes support for threads and MPI processes in order to exploit the compute capabilities of modern distributed-memory systems, which provide several multicore CPU nodes interconnected through a network. Our performance evaluation with 18 representative input datasets on two different eight-node systems shows that our tool is significantly faster than the original BiBit. Source code in C++ and MPI running on Linux systems as well as a reference manual are available at https://sourceforge.net/projects/parbibit/.

ScalaParBiBit: scaling the binary biclustering in distributed-memory systems

“…However, all the previously mentioned works are focused on quantitative datasets and they are not able to take advantage of the binary representation of the information. Up to our knowledge, ParBiBit [6] is the only MPIbased biclustering tool specifically designed for binary data and publicly available to download. It is a parallel version of BiBit [24], a tool that obtains accurate results for gene expression data, especially on cases with many large biclusters [18].…”

“…Furthermore, the thread-based parallelization of this third step in ParBiBit is not able to fully exploit the resources of a one node because it requires many thread synchronizations to add biclusters to the C++ set, which is a non thread-safe structure. More information about the ParBiBit approach and its bottlenecks is available in [6].…”

“…For this, we had to choose the right programming tools for the development of the application and then we had to focus on stages 3 and 4, as they are the ones that consume the overwhelming majority of the runtime of the application, which is in fact the reason why they were the steps parallelized in [6]. From the results in that paper it was also clear that while stage 4 enjoyed a scalable parallelization able to exploit all the resources in a multicore cluster, this was not the case for stage 3, which despite consuming a relevant portion of the runtime, could only be performed by a single MPI process.…”

“…This new strategy gives place to interactions between the threading and the MPI components of the application that did not occur in ParBiBit. The reason is that it introduces overlaps between these elements, which was not the case in [6]. Indeed, in ParBiBit all the bicluster initializations are first performed by multiple threads in the master, and only once they finish the MPI communications are performed by the master thread of each process.…”

“…In this paper we present ScalaParBiBit, a parallel tool to accelerate binary biclustering in multicore clusters that exploits both thread-level and process-level parallelism in order to provide high scalability when increasing the number of nodes. Our software package is derived from ParBiBit [6], a high performance tool also targeting these kinds of clusters, but offering much lower performance and scalability while having much greater memory limitations, as our experiments show. As we will see, the development of ScalaParBiBit not only involved typical sequential and parallel optimizations, but also a strong redesign of the algorithm.…”