The EPCC seminars listed here are open to everybody, and take place from 2–3pm unless otherwise advertised.
If you would like to present at one of our seminars please contact Arno Proeme.
Graph-based problems and the SpiNNaker neural HPC architecture
Dr Alan Stokes (Advanced Processor Technologies group, School of Computing Science, University of Manchester)
Wed September 6th 2017, James Clerk Maxwell Building, room 4325A
This talk highlights two of the many issues high performance computers will have to tackle to reach an exascale machine - power and data communication - and how these problems are starting to be solved. We discuss how software applications will need to be adapted for the solutions to these problems and then describe the SpiNNaker hardware platform and its synergies with the solution for HPCs. We then walk though a simple application mapped from standard C code onto SpiNNaker, and its performance. We end with options on how to acquire access to SpiNNaker hardware and training.
SpiNNaker is a novel computer architecture inspired by the working of the human brain. A SpiNNaker machine is a massively parallel computing platform, targeted towards three main areas of research:
• Neuroscience. Understanding how the brain works is a Grand Challenge of 21st century science. We will provide the platform to help neuroscientists to unravel the mystery that is the mind. The largest SpiNNaker machine will be capable of simulating a billion simple neurons, or millions of neurons with complex structure and internal dynamics.
• Robotics. SpiNNaker is a good target for researchers in robotics, who need mobile, low power computation. A small SpiNNaker board makes it possible to simulate a network of tens of thousands of spiking neurons, process sensory input and generate motor output, all in real time and in a low power system.
• Computer Science. SpiNNaker breaks the rules followed by traditional supercomputers that rely on deterministic, repeatable communications and reliable computation. SpiNNaker nodes communicate using simple messages (spikes) that are inherently unreliable. This break with determinism offers new challenges, but also the potential to discover powerful new principles of massively parallel computation.
MONC: an LES for cloud and atmospheric modelling
Dr Nick Brown (EPCC, University of Edinburgh)
Wed August 30th 2017, James Clerk Maxwell Building, room 4325A
For the past three years I have been working with the Met Office on the Met Office NERC Cloud model (MONC.) This replaces a thirty year old model which has been a crucial tool for UK weather and climate communities but which exhibited significant issues around performance, scalability and the code itself. Our replacement, MONC, has been written from scratch, maintaining the science of the previous model but with modern software engineering and parallelisation techniques. The aim has been to enable the scientists to study vastly larger systems, at far higher accuracy over many cores. In addition to computation, scientists also desire to perform analysis the on raw data in order to generate higher level information. This is a challenge because the raw data is very large in size (many TBs) so it is not realistic to write it out to file and analyse offline. Instead this is performed in-situ on the data as it is generated, which raised several challenges that we had to solve. I will talk about both these aspects of MONC, as well as some of the offshoot work that we have looked at such as porting and evaluating aspects of the model on GPUs and KNLs.
Experiences from EPCC's first MOOC: Supercomputing
Dr David Henty (EPCC, University of Edinburgh)
Wed July 26th 2017, James Clerk Maxwell Building room 4325A
As part of PRACE (Partnership for Advanced Computing Europe), EPCC ran its first ever MOOC (Massive Open Online Course) in March this year. The 5-week course used the FutureLearn platfrom - www.futurelearn.com/courses/supercomputing - which hosts many other Edinburgh MOOCs including the Higgs course from SoPA. In this short informal talk I will cover the history of the course, the process of designing our first MOOC, features of the FutureLearn platform and experiences from the first run in March. I will also compare and contrast MOOCs with other online teaching such as the HPC distance-learning courses we run as part of the DSTI (Data ScienceTech Institute) MSc programme. *Note: the next run of the MOOC starts August 28th - register now!*
Solar Panel detection in Satellite Images using Deep Learning
Marc Sabate (EPCC, University of Edinburgh)
Wed July 12th 2017, James Clerk Maxwell Building room 4325A
Deep Learning models have become very popular with the release of libraries such as Tensorflow, Torch, or Theano, allowing to train deep networks in a reasonable amount of time. In this talk I will present how a Convolutional Neural Network can be used to detect solar panels in Satellite Images.
This talk will start with a brief overview of binary classification problems using Logistic Regression. We will see how Logistic Regression models are built under the assumption that classes are linearly separable, and how Neural Networks can overcome this limitation. I will provide a defitinion of Convolutional Neural Networks, a particular type of Neural Network specifically designed for Image Processing problems, and I will finally present a network that successfully detects solar panels in satellite images from four cities in California.
It is all still an ExaHyPE
Dr Tobias Weinzierl (Department of Computer Science, Durham University)
Wed June 28th 2017, James Clerk Maxwell Building room 4325A
ExaHyPE (http://www.exahype.eu) is a H2020 project where an international consortium of scientists writes a simulation engine for hyperbolic equation system solvers based upon the ADER-DG paradigm. Two grand challenges are tackled with this engine: long-range seismic risk assessment and the search for gravitational waves emitted by rotating binary neutron stars. The code itself is based upon a merger of flexible spacetree data structures with highly optimised compute kernels for the majority of the simulation cells. It provides a very simple and transparent domain specific language as front-end that allows to rapidly set up parallel PDE solvers discretised with ADER-DG or Finite Volumes on dynamically adaptive Cartesian meshes.
This talk starts with a brief overview of ExaHyPE and demonstrates how ExaHyPE codes are programmed, before it sketches the algorithmic workflow of the underlying ADER-DG
scheme. We rephrase steps of this workflow in the language of tasks.
We then focus on a few methodological questions: how can we deploy these tasks to manycores, what execution patterns do arise, and are the new OpenMP task features of any use? How can we rearrange ADER-DG's workflow such that we reduce accesses to the memory, i.e. weaken the pressure on the memory subsystem? How can we reprogram
the most expensive tasks such that they exploit the wide vector registers coming along with the manycores? A brief outlook on MPI parallelisation wraps up this methodological talk.
We focus on results obtained on Intel KNL nodes provided by the RSC Group, on Intel Broadwell results from Durham's supercomputer Hamilton, and on results from the SuperMUC phase 2 supercomputer at Leibniz Supercomputing Centre.
This is joint work with groups from Frankfurt's FIAS, the University of Trento, as well as Ludwig-Maximilians-University Munich and Technical University of Munich.