Section: New Results

Parallelism

Processor in Memory

Participants: C. Deltel, D. Lavenier

The concept of PIM (Processor In Memory) aims to dispatch the computer power near the data. Together with the UPMEM company, which is currently developing a DRAM enhanced with computing units, we investigate the parallelization of several bioinformatics algorithms for this new types of memory. The first results show that blast-like algorithms or mapping algorithms can highly benefit of such memory. But the core algorithms must be revisited in order to better suite the PIM architecture.

Alignment search tools on cloud

Participants: S. Brillet, D. Lavenier, I. Petrov

PLAST is an alternative version of Blast to target intensive sequence comparison (bank-to-bank comparison). The multicore version offers a speed from 5 to 10 compared to Blast. In 2015, we deploy PLAST in the IFB cloud infra-structure (French Bioinformatics Institute) and demonstrate that an Hadoop implementation provides a very good scalability [34] .

Bioinformatics Workflow

Participants: D. Lavenier, F. Moorews

Bioinformatics workflows play an important role in the development of new methodologies for analyzing sequencing data. Optimizing this activity brings the questions of how workflow can be efficiently captured and how technical tasks integration can be simplified. Thus, we define an expressive graphic worfklow language, adapted to the quick capture of workflows. This graphical input is then interpreted by a workflow engine based on a new model of computation with high performances obtained by the use of multiple levels of parallelism. A Model-Driven design approach is associated to facilitate the data parallelism generation and the production of suitable implementations for different execution contexts. In the case of the cloud model Container as a Service (CaaS), a workflow specification intrinsically re-executable and readily disseminatable has been developed. The adoption of this kind of model could lead to an acceleration of exchanges and a better availability of data analysis workflows [25] [31] [13] .

Graph processing

Participants: D. Lavenier, R. Andonov

In the paper [20] we present a new approach for solving the all-pairs shortest-path (APSP) problem for planar graphs that exploits the massive on-chip parallelism available in today's Graphics Processing Units (GPUs). We describe two new algorithms based on our approach. Both algorithms use Floyd-Warshall method, have near optimal complexity in terms of the total number of operations, while their matrix-based structure is regular enough to allow for efficient parallel implementation on the GPUs. By applying a divide-and-conquer approach, we are able to make use of multi-node GPU clusters, resulting in more than an order of magnitude speedup over fastest known Dijkstra-based GPU implementation and a two-fold speedup over a parallel Dijkstra-based CPU implementation.

Analytical models and Optimization for GPUs

Participants: R. Andonov

In [28] we develop a methodology for modeling the energy efficiency of tiled nested-loop codes running on a graphics processing unit (GPU) and use it for energy efficiency optimization. We use the polyhedral model, and we assume that a highly optimized and parametrized version of a tiled nested – loop code, either written by an expert programmer or automatically produced by a polyhedral compilation tool – is given to us as an input. We then model the energy consumption as an analytical function of a set of parameters characterizing the software and the GPU hardware. Our approach develops analytical models based on (i) machine and architecture parameters, (ii) program size parameters as found in the polyhedral model and (iii) tiling parameters, such as those that are chosen by auto-or manual tuners. Our model therefore allows efficient optimization of the energy efficiency with respect to a set of parameters of interest.