Towards an Exaflop
The largest supercomputers today can perform over a Petaflop of
calculations per second on real scientific applications. However exascale
systems are planned for 2016-2018, performing an exaflop is a million
million million calculations per second, a thousand times faster. To
put this in context, 1 exasecond is 32 billion years, or more than twice
the approximate age of the universe.
To achieve this, the trend in HPC systems is towards a massive increase
in parallelism and exascale systems may have billions of cores.
Clearly there is an immense challenge in utilising these systems, with
many innovations in supercomputer hardware, software and applications
needed over the next decade.
Exascale computing will be vital for European competitiveness in many
areas of daily life including global numerical weather prediction,
sustainable energy development and medical advances. By being at the
forefront we hope to enjoy the benefits of this vital technology.
Projects:
Exascale Technology Centre
EPCC established the Exascale Technology Centre at Edinburgh jointly
with Cray and as part of Cray’s European Exascale Research Initiative.
With a team jointly funded by both organisations, the Exascale
Technology Centre in Edinburgh is exploring new ideas and new
technologies to meet the challenge of delivering an exaflop within the
next decade. In particular we are focussed on new and evolving
programming models and accelerated node design and programmability. To
scale to the extreme, programming models will need, for example, to be
able to provide good data locality management, lightweight asynchronous
communications and improved synchronisation. The centre is therefore
looking at hybrid, single-sided and PGAS models. Exascale
architectures will meanwhile contain large numbers of multi-socket,
multi-core nodes. To increase node performance, accelerators will be an
important component and developing efficient programming models for
these essential. The centre is therefore exploring programming model
for accelerators. The project is funded by the University of Edinburgh
and by Cray. Contact: Lorna Smith, l.smith@epcc.ed.ac.uk
Novel Asynchronous Algorithms and Software for Large Sparse Systems
The solution of large sparse systems, both linear and nonlinear is a
key numerical technology underpinning many areas of computational
science and engineering, including climate and environmental modelling,
nuclear fusion, materials science and computational chemistry. The
reliance of these and other application domains on sparse system
solution means that they all face difficulties in achieving extreme
scalability, since the underlying algorithms are highly synchronous.
This project aims to develop more scalable numerical methods through
the use of asynchronous iterative algorithms. In asynchronous
iterations, the order in which components of the solution are updated
is arbitrary and the past values of components that are used in the
updates are also selected arbitrarily. This is a model for parallel
computation in which different processors work independently and have
access to data values in local memory. The use of asynchronous methods
allows one to overcome many of the communication, load balancing and
fault tolerance issues we now face and which limit our ability to scale
to the extreme. This project is funded by EPSRC. Contact: Mark
Bull, m.bull@epcc.ed.ac.uk
Numerical Algorithms and Intelligent Software (NAIS)
Many of the most common underlying algorithms are already performance
limited on current tera and petascale platforms. These limitations will
only increase, and in some cases become untenable, at exascale. For
example, achieving a reasonable computation/communication balance
without a major increase in problem size, coping with increasing memory
latencies, avoiding global synchronisation points and load imbalance
are all important considerations. In addition, exploiting the complex
heterogeneous nature of the underlying architecture will become key. A
consortium of three Scottish universities, led by Edinburgh, has been
awarded a grant to set up The Centre for Numerical Algorithms and
Intelligent Software (NAIS) where mathematicians, computational
scientists and HPC software developers can work together on a new
generation of scalable numerical algorithms. In particular, NAIS is
focussed on developing new systems of code annotation, new compilers
and efficient implementations for application-oriented computational
methods such as adaptive finite elements, multi-scale modelling,
molecular simulation and optimisation. These will be applied to real
applications which require exascale performance. NAIS is funded by
EPSRC and SFC. Contact: George Beckett, g.beckett@epcc.ed.ac.uk
The TEXT Project
TEXT builds on the belief that the programming model is the key
component to support high productivity and efficient use of an exaflop
system. A hybrid model consisting of MPI + Barcelona’s SMP SuperScalar
(SMPSs) can be demonstrated today and can show the way to exascale.
The SMPSs model, which uses directives to describe tasks and their data
dependencies, provides the necessary support for asynchrony and
heterogeneity as well as enabling incremental parallelisation,
modularity and portability of applications. By combining in with MPI we
can propagate SMPSs characteristics to the global application level.
This is also a way to leverage and provide a smooth migration path for
the huge number of applications today written in MPI. The focus of the
TEXT project is to install the MPI + SMPSs environment at the HPC
facilities of several partners and demonstrate how it can be used to
improve seven real and relevant applications/libraries. TEXT is funded
by the European Community’s Seventh Framework Programme. Contact: Mark
Bull, m.bull@epcc.ed.ac.uk