Yufei Lin

A Parallel RBF Mesh Deformation Method with Multi-greedy Algorithm in OpenFOAM

Radial Basis Function(RBF) mesh deformation method has been widely used in CFD simulations with moving boundaries due to its high robustness and accuracy. The original implementation of the RBF mesh deformation method in OpenFOAM(a widely used CFD software) is purely serial with relatively low computational performance. To reduce the time cost of the mesh motion in large-scale simulations, this paper proposes a parallel RBF mesh deformation method with multi-greedy algorithm in OpenFOAM. The proposed multi- greedy method could reduce the control points used by the RBF interpolation on both the moving boundary and the static boundary, which is more applicable than the previous typical greedy algorithm. Based on a master-worker algorithm, the computation of the mesh deformation is highly parallelized. Tests on the benchmark of a three-dimensional moving fish show that with an error tolerance of 1e-4, the interpolation time of the internal mesh motion using our multi-greedy method is about 10.2 times faster than the original one, and with a parallelism of 132, the time cost of the whole mesh motion is greatly reduced with a speedup of 37.

A Hybrid Parallel Algorithm for Solving Eeuler Equation Using Explicit RKDG Method Based on OpenFOAM

OpenFOAM is a framework of the open source C CFD toolbox for flexible engineering simulation, which uses finite volume method (FVM) in the discretization of partial differential equations (PDEs). The problem solving procedure in OpenFOAM consists in equations dicretization stage, equations solving stage and field limiting stage. The best parallelism is limited by the equation solving stage, which contains communications. Compared to FVM, discontinuous Galerkin (DG) method is a high-order discretization method, which can accelerate the convergence of the residuals over same mesh scale and has higher resolution of the flow. Based on OpenFOAM with DG method, the ratio of overhead in equations discretization stage increases, especially when solving Euler equations using an explicit method. The equations discretization stage has a better potential parallelism than the other two stages due to no existence of communication. In this paper, we will analysis the difference of time cost in these three stages between original OpenFOAM and OpenFOAM with DG method. By decoupling these three stages, a hybrid parallel algorithm for solving PDEs is proposed and implemented based on OpenFOAM with DG method. The experimental results show that the simulation time is reduced by 16%, and the relative speedup of the hybrid parallel algorithm is up to 2.88 compared to the original parallel algorithm with the same degree of parallelism.

Practical Performance Models for High-Order CFD Simulation

This paper discusses the performance model for practical computational fluid dynamics (CFD) simulation. Based on the analysis of the conventional qualitative model and empirical knowledge in the simulation, a set of practical models for the linear solution in high-order finite element method (FEM) simulation are established: a linear function model for the storage consumption, and a power function model for the solution time. The unknown parameters in the models are estimated by linear fitting and nonlinear fitting, respectively. Tests on a typical high-order FEM platform show that the proposed models are highly accurate and practical. In addition, it has been tested that nonlinear fitting is better than linear fitting in our case.

SMP-SIM: An event-driven performance simulator for SMP

Performance prediction for the high performance computer system is of great importance for designing, implementing, and optimizing system. As a widely used technique for predicting performance, simulation method attracts more and more attention from the research community. Based on analyzing the problems in the current performance simulation techniques, we present a key idea of the performance simulator for SMP system based on event-driven. We propose the framework of SMP-SIM and implement it based on MPICH2. The simulation results show that, our simulation technique has the advantages of high accuracy and simulation performance.

Communication-aware scalability analysis based on critical path

With the scale of parallel computing system growing up, the influence of the communication overhead cannot be ignored. Parallel speedup is the most commonly used metric for evaluating scalability in parallel computing, and we propose communication-aware speedup based on the traditional speedup to quantitatively analyze the system scalability. This paper presents the definition of communication-aware scalability, discusses the communication-aware scalability by using critical path, and then quantitatively classifies the systems according to their communication-aware scalabilities. Through case studies, we analyze the communication-aware scalabilities of some parallel systems. A significant conclusion of this paper is that the system is scalable, if the communication overhead increases at a rate with lower exponent than parallel workload.

SaPM: Switch-aware process mapping model for parallel computing

The problem of assigning processes of a parallel program to processors of a parallel system, namely process mapping problem, has major impact on the resulting performance. The previous works only concern about the computation and communication cost, while ignoring the process switching cost. For a large-scale and communication-intensive application, more than one process may be placed onto a same processor. When the application is running, it will take long time to switch the processes. In this paper, we propose and introduce SwitchCAL, a method to calculate the process switching cost inside a processor, into the process mapping problem. Then we propose a switch-aware process mapping model - SaPM, and integrate the SaPM model with existing process mapping algorithm to obtain optimized process mapping scheme. The experiments show that our approach has high effectiveness.

Workload-Awared HPC System Scalability Analysis Based on Electronic Technology Developing

As the electronic technology develops, the integration levels of CPUs and memories keep growing, and the speeds of communication devices are improved. The high-performance computing (HPC) systems consist of processing nodes and communication network, and their sizes are advanced by the development of electronic technology. Then the scalability of a large-scale parallel computing system, i.e. whether the computing performance is increased with the system size, becomes a major goal pursued by designers of parallel algorithms and high-performance parallel machines. Parallel speedup is a popular way to measure the scalability. This paper proposes the definition of HPC system scalability based on speedup first, and then analyzes the influence of function G(P), which describes how the workload changes with processor number, on the system scalability. Through case studies, we analyze some typical programs based on our scalability theorems and the results show that our analysis approach is correct.

PS-SIM: An Execution-Driven Performance Simulation Technology Based on Process-Switch

Nowadays, the performance of large-scale parallel computer system improves continuously, and the system scale becomes extremely large. Performance prediction has become an important approach to guide system design, implementation and optimization. Simulation method is the most widely used performance prediction technology for large-scale parallel computer system. In this paper, after analyzing the extant problems, we proposed a novel execution-driven performance simulation technology based on process-switch. We designed a simulation framework named PS-SIM, and implemented a prototype system based on MPICH2. Finally, we verified the proposed approach by experiments. Experimental results show that the approach has high accuracy and simulation performance.

A Task-Mapping Algorithm for Performance Simulation Based on Simulated Annealing

Nowadays, the scale of parallel computer systems is increasing, and simulation technology has become an important tool for performance prediction in the system development process. Task mapping approach is an important aspect affecting the performance of simulation. In this paper, in order to solve the task mapping problem in performance simulation, a task mapping algorithm based on simulated annealing is proposed, and we verified the correctness and effectiveness of the algorithm by experiments. Experimental results show that the algorithm has high efficiency, and can solve the large-scale problem with lower time cost.