The SIPSim implicit parallelism model and the SkelGIS library

Abstract

Scientific simulations give rise to complex codes where data size and computation time become very important issues, and sometimes a scientific barrier. Thus, parallelization of scientific simulations becomes a significant work. Many time and human efforts are deployed to produce efficient parallel programs. But still, many simulations could not be parallelized because of lack of time to learn parallel programming or lack of human resources. Therefore, aiding parallelization through abstracted parallelism or implicit parallelism has become a main topic in computer science. Many implicit parallelism solutions have been proposed such as algorithmic skeletons libraries, domain-specific languages or specific libraries. In this paper is introduced a new type of solution to give a totally transparent access to parallel programming for non-computer scientists of the domain of numerical simulations. This solution is an implicit parallelism model, called Structured Implicit Parallelism on scientific Simulations (SIPSim). After a description of the SIPSim model, this paper presents the implementation of the model, as a C++ templated library called SkelGIS, for two different cases of simulations: simulations on Cartesian meshes and simulations of two physical phenomena linked through a network. For each case, the implementation of the SIPSim components are described, and a simple simulation example is given. SkelGIS is then evaluated on two real cases, one for each case, first on the resolution of shallow water equations and second on an arterial blood flow simulation. To clearly state on SkelGIS performance and its ease of programming, different experiments on both cases are evaluated. THE SIPSIM IMPLICIT PARALLELISM MODEL AND THE SKELGIS LIBRARY 2123The data parallel version of Algorithm 1 is given in the Algorithm 2. This algorithm seems very close to the sequential one; however, it requires the decomposition of the mesh into p parts that are distributed among the available processors. In the time loop, a communication phase is performed so that each processor obtains the whole needed information to perform the new iteration. These communications may be tricky to implement for non-computer scientists.This paper presents an implicit parallelism model that proposes a systematic way to write implicit parallelism solutions for numerical simulations. The model aims at providing efficient parallel programs as the Algorithm 2, while exposing a sequential program as the Algorithm 1. In order to let as much freedom as possible to the programmer, the main time iteration loop as well as the loops on the physical border and the mesh element will not be hidden in the sequential program. Related workThe implicit parallelism model presented in this paper, and implemented as a library, namely SkelGIS, totally hides parallelization of programs but preserves the programming freedom and the flexibility of a sequential programming style.Different kinds of implicit parallelism solutions can be found in the literature. One of the most recent ...

Towards transparent combination of model management execution strategies for low-code development platforms

“…For instance, data may need to be merged into a single result. is computation strategy is the one followed by the parallel algorithmic skeletons [16] on data structures [19,35].…”

Automatic Energy Efficient HPC Programming: A Case Study

“…As an example of production numerical simulation, we consider FullSWOF2D 4 [5] (denoted FS2D), developed at the MAPMO laboratory, University of Orléans, France. FS2D consists in solving the Shallow Water equations (two dimensional Navier-Stokes equations) using finite volumes methods especially chosen for hydrodynamic purposes (transitions between wet and dry areas, small water heights, etc.).…”

“…In MSL these techniques are implemented by using the Message Passing Interface (MPI) and the OpenMP Application Programming Interface. The four code versions produced by MSL are: (1) MpiBase, where data parallelization is applied by domain decomposition and by using MPI; (2) MpiOmpFor, where data parallelization is introduced at two different levels, first, by domain decomposition with MPI, and second, by using parallel loops of OpenMP; (3) MpiOmpForkJoin, where both data and task parallelization techniques are combined, and where the adopted task parallelization technique is a static fork/join scheduling implemented using OpenMP; and finally (4) MpiOmpDyn, where both data and task parallelization techniques are also combined, but where the adopted task parallelization technique is the dynamic scheduling of tasks introduced in OpenMP 4.5 5 .…”

New advances in High Performance Computing and simulation: parallel and distributed systems, algorithms, and applications

“…Coullon et al [15], in their paper Implicit parallelism on 2D meshes using SkelGIS, tackle the issue of overcoming restrictions in parallelization of scientific simulations because of the complexity of functional concepts and specific features. Parallelization of scientific simulations requires a lot of efforts and field-specific knowledge to produce efficient parallel programs.…”