IBM and GENCI drive supercomputing closer to exascale
IBM and GENCI, the French national supercomputing agency, have announced a collaboration aimed at speeding up the path to exascale computing – giving GENCI access to some of the most advanced high performance computing technologies stemming from the rapidly expanding OpenPOWER foundation.
As part of the collaboration, GENCI will examine the impact and requirements of POWER’s open architecture on scientific applications, intending to foster a deeper understanding of application requirements as the computing industry advances towards the era of exascale computing.
The collaboration will attempt to take full advantage of the impact of OpenPOWER-based innovations such as the connection of Nvidia GPUs to POWER processors through Nvidia's high-speed NVLink interconnect. In addition Mellanox EDR 100Gbps InfiniBand switches can exploit IBM’s Coherent Application Processor Interface (CAPI) to dramatically improve solution performance.
‘If we want to continue to address the challenges of the French scientists and engineers, we need to anticipate the rise of new high performance computing architectures that bring us closer to exascale,’ stated Catherine Riviere, CEO of GENCI.
Working closely with supercomputing experts from IBM, GENCI will have access to some of IBM's most advanced high performance computing technologies being developed by, or in conjunction with, the OpenPOWER ecosystem. OpenPOWER has been growing rapidly, it now consists of more than 140 OpenPOWER foundation members and thousands of developers worldwide. The collaboration, planned to run for at least 18 months, focuses on readying complex scientific applications for systems currently in development. These systems are expected to be in excess of more than 100 petaflops, the next logical step towards exascale.
Experts from GENCI and French research organisations, in collaboration with IBM, plan to work on understanding the evolution of programming models, considering MPI and OpenMP as a first step for shared memory multiprocessing programming. Alternative application program interfaces will also be considered, given potential changes may be required as systems move closer toward exascale.
Currently the fastest systems in the world perform between 10 and 33 petaflops, or 10 to 33 million billion calculations per second – roughly one to three per cent the speed of exascale. Although some experts believe that raw FLOPs performance may not be the best way to derive a supercomputers true application performance - as reported in 'Exascale: expect poor performance' - it remains the deafult standard for the HPC industry .
IBM will provide dedicated technical experts to support application porting and optimisation efforts as well as organising, along with GENCI, education and porting sessions. This collaboration will be supported by the newly created POWER Acceleration and Design Center in Montpellier as part of the partnership established with both Nvidia and Mellanox. The centre will provide technical expertise around scientific applications, programming models and systems, as well as early access to forthcoming 2016 platforms and the latest innovative technologies (Nvidia NVLink, IBM CAPI), the IBM high performance computing software stack and the Nvidia Tesla Accelerated Computing Platform.
'The work we are doing with GENCI – bringing together some of the best minds in science and information technology – is a collaborative effort on a grand scale involving not just GENCI and IBM, but thousands of developers contributing to the rapidly expanding OpenPOWER ecosystem worldwide,’ said Michel Teyssedre, CTO of IBM France. ‘We fully expect our collaborative efforts will produce innovations capable of moving the supercomputing industry that much closer to exascale.’