EN FR
EN FR




Section: New Results

Scalable I/O and visualization for post-petascale HPC simulations

Participants : Matthieu Dorier, Gabriel Antoniu.

In the context of the Joint INRIA/UIUC Laboratory for Petascale computing (JLCP), we are addressing the new challenges related to I/O, data analysis and visualization for extreme-scale simulations. As HPC resources approaching millions of cores become a reality, a growing challenge in maintaining high performance is the presence of high variability in the effective throughput of codes performing I/O operations. Since I/O is mainly performed for the purpose of subsequent data analysis and visualization, another way to limit the impact of I/O performance on scientific discovery consists in enabling in-situ visualization. This brings again new challenges such as how to efficiently couple large-scale simulations with visualization software.

We started the development of Damaris, a middleware targeting multicore SMP nodes to efficiently address the problems mentioned above. Damaris has been evaluated with the CM1 atmospheric simulation  [43] , one of the targeted application for the BlueWaters project. To show the capability of our approach to efficiently hide I/O jitter and related costs, experiments have been carried on the French Grid'5000, on a Power5 cluster at NCSA and on the Kraken Cray XT5 supercomputer (currently 11th in the Top500) with up to 9K cores. By gathering data into large files while avoiding synchronization between processes, our solution brings several benefits:

  1. it increases the sustained write throughput of the simulation by a factor of almost 15;

  2. it provides almost 70 % overall application speedup on 9,000 cores;

  3. it fully hides I/O-related costs;

  4. it enables a 600 % compression ratio without any additional overhead, leading to a major reduction of storage requirements.

A poster [16] presenting some of these results has been accepted at ICS'11 and awarded the second price at the ACM Student Research Competition. All these results are presented in a research report [25] pending for publication.

Current work addresses the efficient coupling of large-scale simulations and visualization tools through Damaris. We have been able to get access to the Jaguar supercomputer hosted at ORNL and we are planning very-large scale experiments on up to 100,000 cores to show the benefits of our Damaris approach.