Session 11
Session 11
11.1 Project Talk: Enhancing Asynchronous Parallelism in OmpSs with Argobots ”Improving Hybrid MPI/OmpSs applications using Argobots”
Jesus Labarta (BSC)
”This talk will present the latest work done on the ””Enhancing Asynchronous Parallelism in OmpSs with Argobots”” project, where we have implemented the OmpSs tasking model on top of the Argobots threading library. In this talk we will explain our latest experiments that show how the integration between the Argobots threading library and the MPICH MPI library can improve both programmability and performance of hybrid MPI and OmpSs applications. Moreover, we also propose and compare this close integration of Argobots and MPICH with an alternative method based on MPI call interception.”
11.2 Individual Talk: Exploring RDMA libraries for PGAS language
Tetsuya Odajima (RIKEN)
Recently, there are many communication libraries. Especially, an MPI-3 is widely used in many of HPC applications. Similarly, some PGAS languages employ MPI-3 as communication layer. However, there are not enough evidence that MPI-3 is optimal for runtime of PGAS language. In this project, we explore RDMA library which has high performance and high usability for developer of languages. In this talk, we evaluate preliminary performance of these libraries to find and optimal library for PGAS runtime system.
11.3 Break Out session: HPC challenge of LQCD and related
Char: Yoshifumi Nakamura (RIKEN)
Speakers:
Stefan Krieg (JSC) ”QCD software development at JSC”
Hiroya Suno (RIKEN)”Algorithm and code development in lattice QCD using the K computer” Giorgio Silvi (JSC)”Qlua optimization for many core architectures”
Yoshifumi Nakamura (RIKEN)”Towards high performance Lattice QCD simulations on Exascale computers”
We propose a break-out session for elementary particle physics applications with a focus on Lattice QCD. Lattice QCD (LQCD) is an approach to solving the quantum chromodynamics (QCD) which describs interaction for quark and gluon, is basic theorem of hadron physics, and application to understand particle collisions and state of matter in extreme environments like early universe, neutron stars and so on. LQCD is defined on 4 dimensional (space and time) lattices. The domain decomposition is used with MPI and Openmp. The hot spot is at solving a Dirac equation by using Krylov subspace methods. Theoretical required B/F in matrix vector multiplication is about 2 in double precision. When program is highly optimized on L2 cache, costs for global reduction and boundary exchange between processes become dominant. So this application is one of most challenging applications on extreme scale computers The aim of the breakout session is to reveal issues of elementary particle physics applications, mainly LQCD, on extreme scale computers and to explore the possibility of new collaboration to tackle them.