Section: Research Program

Research axis 3: I/O management, in situ visualization and analysis on HPC systems at extreme scales

Over the past few years, the increasing amounts of data produced by large-scale simulations have motivated a shift from traditional offline data analysis to in situ analysis and visualization. In situ processing started by coupling a parallel simulation with an analysis or visualization library, to avoid the cost of writing data on storage and reading it back. Going beyond this simple pairwise tight coupling, complex analysis workflows today are graphs with one or more data sources and several interconnected analysis components.

Collaboration.

This axis is worked out in close collaboration with Rob Ross (ANL), Tom Peterka (ANL), Matthieu Dorier (ANL), Toni Cortes (BSC), Bruno Raffin (Inria). Some additional collaborations are in discussion with other members of JLESC, and with CEA and Total.

Relevant groups with similar interests include the following ones.

• The group of Manish Parashar at Rutgers University, USA (I/O management for HPC systems, in situ processing).

• The group of Scott Klasky at Oak Ridge National Lab, USA (I/O management for HPC systems, in situ processing).

• The CNRS IPSL laboratory (Sébastien Denvil, Pôle de modélisation du climat) in Paris, France (in situ data analytics).

Toward a joint optimized architecture for in situ visualization and advanced processing

From Inria and ANL, four tools at least have emerged to address some challenges of coupling simulations with visualization packages or analysis workflows. Each of them focused on some particular aspect:

Damaris

(Inria, [5], [4]) exploits dedicated cores to enable jitter-free I/O and in situ visualization;

Decaf

(ANL, [33]) implements a coupling service for workflows;

FlowVR

(Inria, [45]) connects workflow components for in situ processing;

Swift

(ANL, [48]) focuses on implicitly parallel data flows and was optimized for Big Data processing.

Our plan is to explore how these tools could best leverage their respective strengths in a joint optimized architecture for in situ visualization and advanced processing in the HPC area. We published a preliminary study describing the lessons learned from using these tools in production environments with real applications [7]. Such a joint architecture will contribute to address the data volume and velocity challenges raised by data-intensive workflows, including complex data-intensive analytics phases. It may also impact, in a subsequent step, future data analysis pipelines for converged Big Data and HPC architectures.