The DEEP-ER way towards Exascale I/O and resilience

Suarez, Estela

Conference Presentation (Invited)

FZJ-2015-04864

The DEEP-ER way towards Exascale I/O and resilience

Suarez, E. (Corresponding author)FZJ*

2015

NorduGrid Conference 2015, Bern, Switzerland, 4 Jun 2015 - 5 Jun 2015

Please use a persistent id in citations: http://hdl.handle.net/2128/8962

Abstract: Resiliency and I/O are two major challenges faced in High Performance Computing when considering building, operating and working on an Exascale system. The European Exascale project "DEEP-ER" addresses both topics in a holistic approach that covers the full system stack from hardware to applications, and takes into account also aspects as scalability, concurrency, and power consumption.The largest effort is invested in DEEP-ER in software and application development. Making file systems more scalable, improving the performance of parallel I/O libraries, and developing resiliency techniques to reduce the impact of hardware failures on applications are some of the key contributions of the DEEP-ER project. Additionally, DEEP-ER will build a hardware prototype to extend the hybrid "Cluster-Booster architecture" – which was first introduced in the predecessor EU-Exascale project DEEP – integrating new memory technologies and using them to improve I/O and checkpointing performance. Hardware and software designs are driven in DEEP-ER taking application requirements into account. For this purpose, seven HPC codes representing very different scientific fields are being analysed, optimised and ported. The experience gathered with them will serve also to evaluate the performance and usability of the DEEP-ER developments against other alternatives and approaches.An overview of the DEEP-ER I/O and resiliency approaches will be given, with focus on the current project status and recent achievements.

Contributing Institute(s):