Programme

The programme is as follows:

9:00-9:20 Welcome 

9:20-9:30 Welcome presentation for the 1st HAMM workshop by C. Prud'homme

9:30-11:10 Invited speakers

• Joel Falcou, Univ. Orsay

  Designing Architecture-aware Library using Boost.Proto

  Embedded Domain Specific Languages (EDSL) are great tools to design both efficient and expressive library. One of the various way to do so is to build C++ library based on the Expression templates idiom. However, many projects using expression templates rely on a manual rewrite of the expression system, leading to poorly scalable, hard to retarget code.
  To try to go further, we decided to use Boost.Proto - an EDSL toolkit - to redesign NT2, a C++ scientific computation library similar to MATLAB in term of interface, while taking architecture into account at the very beginning. This talks is about how this redesign helped leveraging multi-architectures support, how it helped introducing a simple openCL back-end and how it helped makign NT2 more open and aware of potential future systems 
• Vivien Clauzon, Numtech

  GPGPU computing @ Numtech

  

  Since 2009, Numtech performs benchmarks of several GPGPU applications aiming at developing a fast operational decision support system for crisis management. Because return time is of main interest for this type of applications, GPGPU appears to be a lightweight and reliable solution. Two of these benchmarks will be presented :
  

  •  a 2D/3D interactive compressible flow solver (Euler's equations) on structured grid using HLLC Riemann solver and a MUSCL type method. User can interact with the simulation adding walls and « splashes ». This application can be viewed as a prototype of a shock wave propagation method in urban-like geometries. This application performs 60 times faster (around 65GFlops using a GTX275) than its single threaded counterpart on a Core i7 920 (using small sized grid, 512x512)

  •  a 2D interactive incompressible flow solver (Navier-Stokes equations) on structured grid using a semi-lagrangian scheme, implicit diffusion and a conjugated gradient solver to solve pressure Poisson equation. 

  This application also include user interaction (velocity addition, obstacle, smoke, …), some other physical effects like buoyancy or reaction rate, and some additionnal numerical scheme such as Mac Cormack advection and vorticity confinement. This benchmark performs around 40GFlops on a Tesla C1060 (using a small sized grid : 512x512), which is 10 times better than the OpenMP version on a Core i7 920. Shared, texture and constant memories are used to avoid concurrent and/or non-coalesent access. Both application are running entirely on the GPU althought some tasks like I/O lie on the CPU (I/O are made asynchonous anyway). These benchmarks appear to be memory bounded. Those projects are still under development and part of the OpenGPU french project.

  
  

  Numtech 1	YouTube Video
  Numtech 3	Numtech 4
  Numtech 5

  
  

11:10-11:30 Coffee Break

11:30-12:20 Invited speakers

• MTL4: massive parallel computations and towards hybrid computations by Peter Gottschling (U. Dresden/Simunova)

  MTL4: Massive parallel computations and towards hybrid computations

  The Matrix Template Library 4 is designed for efficient and most notably productive development in scientific computing (www.simunova.com/pmtl4).  To achieve the productivity the library provides an intuitive interface with concise notation using operators as far as possible. To overcome the limited number of operators, their semantics is overloaded for a variety of types establishing a domain-specific embedded language (DSEL). For instance, operator[] can be used to access matrix elements, row and column vectors or sub-matrices.  The parallel edition of MTL4 naturally extends its sequential counter-part.  Distributed matrices and vectors provide the same interface (except for prohibitively expensive operations) so that applications can be parallelized with little or no modifications. For instance, the iterative Krylov sub-space methods are identical in the sequential and parallel MTL4. Together with the easy parallel initialization, a distributed linear system can be set up in few lines and solved with the latest iterative solvers.  The library possesses a simple interface to ParMetis. One can write a scientific application with migration in less than an hour.  Benchmarks on up to 1000 cores have shown that the solver of PMTL4 was slightly faster than PETSc. An implementation of MTL4 with GPUs is in progress. The generic design of the library will allow for smoothly merging the GPU and parallel version in order to support GPU clusters.

12:20-13:40 Lunch Break

13:40-15:00

• OpenCL and the SSE vectorisation in BigDFT by Brice Videau (LIG/INRIA)

  OpenCL and the SSE Vectorisation in BigDFT

  Determining material properties is necessary in order to develop nanotechnologies.  BigDFT is an electronic simulation software that uses a wavelet implementation of the Density Functional Theory to study those properties. The base operations of BigDFT are small 3-dimensional convolutions. We present optimization strategies applied to these convolutions in order to direct their integration in an autotuning library.  In order to yield acceptable performances in a rich HPC environment BigDFT needs to be optimized for several architectures. Two target architectures are studied, namely processor with vector units and GPU acceleration using OpenCL, and different optimization levels are considered. Especially, the different code transformations necessary to yield performance gains on each architecture are presented. Several performance studies are conducted, from kernel benchmarking to BigDFT running in a hybrid environment with different brand of GPUs.
• The european project montblanc (what does it mean for HAMM) by JF Mehaut (UJF/HAMM)

  The european project MontBlanc, what does it mean for HAMM

  In this talk, I will present an overview of the Mont-Blanc project. The main motivation of this project is energy efficiency. The Mont-Blanc approach is to investigate the use of low power CPUs (ARM core) and embedded GPU for energy efficient HPC. These technologies have mainly developed and for smart phones and tablets and have never beeen used in HPC systems before, leading to a number of significant challenges.
  
  The Mont-Blanc project has three objectives:
  
  

  1. To develop a fully functional energy-efficient HPC prototype using low-power commercially available embedded technology
  2. To design a next-generation HPC system together with a range of embedded technologies in order to overcome the limitations identified in the prototype system
  3. To develop a portfolio of Exascale applications to be run on this new generation of HPC systems.
  4. I will describe the advances made in Mont-Blanc and the possible collaborations with the HAMM project.

• Advances in Feel++/HAMM by V. Chabannes (UJF/HAMM), A. Samake (UJF/HAMM), S. Veys (UJF/HAMM), Jean-Marc Gratien (IFPEN/HAMM) and C. Prud'homme (UJF/HAMM)

  Advances in Feel++/HAMM

  We present an overview of the advances made in Feel++/HAMM in various areas namely (i) parallel computing using MPI and PETSc (ii) domain decomposition methods such as Schwarz methods, mortar or threefields methods, (iii) reduced basis methods  framework and finally (iv) the domain specific languages that bind these contributions. Some applications will be presented to illustrate the work.

• GPUAlgebraFramework by JM Gratien (IFPEN/HAMM)

  GPUAlgebraFramework, a library to perform algebra operations on multi independent linear systems on GP-GPU cards

  Usually muti-scale methods consist in  finding the solution of a PDE on a fine grid, defining a set of appropriate basis functions that enable to define a coarse problem on a coarse grid and to recover the fine solution from the solution of the coarse problem. The construction of theses basis functions often leads to build and solve a set of independent linear systems. GPUAlgebraFramework has been developed within the HAMM Project to solve such systems on GPGPU cards. We detail some technical implementation issues, present some performance results and finally discuss on interesting perspectives for this framework.
  
• Coupling of a particle solver with flow solver by J.M. Etancelin and G.H. Cottet (UJF/HAMM)

Coupling of a particle solver with flow solver

Particles methods are powerful and versatile methods to approximate solutions of partial differential equations. In our work we focus on two aspects of the method. First, our goal is to build a parallel particle solver coupling with a LES solver on distributed architectures for large Schmidt number. The goal is to avoid the CFL constraint on the transport equations.

In a second part, we look at the parallelization using opencl to build a versatile and portable solver on hybrid platforms.

Sponsors

The Feel++ Users day is sponsored by the ANR Project HAMM and the Région Rhône-Alpes hanks to the project CHPID/ISLE.