Matthieu Dorier, a postdoctoral appointee in Argonne’s Mathematics and Computer Science Division, has won second place in the competition for the Gilles Kahn Prize. The prize is given every year to at most the three best Ph.D. theses in computer science in France and is jointly delivered by the Société Informatique de France and the French Academy of Science. The candidates are judged on all aspects of their Ph.D. work, from fundamental contributions to industrial transfers, and from publication impacts to teaching, mentoring, and scientific dissemination activities.
Dorier’s thesis, titled “Addressing the Challenges of I/O Variability in Post-Petascale HPC Simulations,” was completed at Ecole Normale Supérieure de Rennes and at IRISA Rennes, under the supervision of Dr. Gabriel Antoniu and Prof. Luc Bougé. It tackles the problem of performance variability in data management tasks. This variability significantly impacts application performance at scale and its predictability over time. In his thesis, Dorier focused on variability caused by I/O contention between processes in the same nodes and across independent applications. One of his solutions, Damaris, involves using dedicated cores to overlap I/O with computation and gather data into large files while avoiding synchronization between cores.
The new approach successfully hides the variability, as well as all I/O-related costs, thus making the simulation’s performance predictable. Moreover, it increases the sustained throughput by a factor of 15 compared with standard approaches, and it allows almost perfect scalability of the simulation up to over 9,000 cores.
Dorier also used Damaris to tackle the problem of efficient in situ analysis and visualization and to study the tradeoffs between performance and energy consumption in high-performance computing simulations. Using an atmospheric model on the Grid’5000 testbed, he implemented three approaches in Damaris: dedicated cores, time partitioning, and dedicated nodes. The results revealed significant differences in performance and energy consumption.
“Choosing the most energy-efficient approach for a particular simulation on a particular machine can be a daunting task,” said Dorier. He therefore formulated a model that helps scientists estimate the energy consumption of their application on a particular platform and under different data management approaches.
Dorier also addressed the challenges of coordinating multiple independent applications to prevent I/O interference and predicting the I/O behaviors of high-performance computing (HPC) applications.
Dorier’s work was done in the context of the Joint Laboratory for Petascale Computing (now JLESC), a collaboration involving INRIA, Argonne, the University of Illinois at Urbana-Champaign, Jülich Supercomputing Centre, Barcelona Supercomputing Center, and Riken. Within this context, Dorier visited Argonne several times for internships.
As part of his research at Argonne, Dorier has continued to explore new approaches to in situ visualization as well as new ways to predict and model I/O behaviors in HPC applications.
Dorier’s dissertation is available on the web.