National University of Singapore

HPC Challenge 2009 Results and Highlights

HPC Challenge 2009 was launched in June 2009 with the objectives of promoting awareness and effective use of High Performance Computing (HPC) resources among the NUS research community.  Participants were invited to demonstrate their parallel computing skills and use of parallel computing resources such as the multi-core clusters and grid environment at NUS.  At its close in mid October 2009, the Challenge attracted 15 entries in research domains spanning the Life and Physical Sciences, the Engineering Sciences, Building and cross disciplinary areas such as Bio- and Nano-technology.  Many real world problems were addressed in Process, Structure and Building design, Fluid-Structure interactions, Bio-systems, Microscopy and Cancer research, and Financial Engineering. Techniques used include Fluid dynamics modeling, Molecular simulations, Bioinformatics, and Optimization searches.

Training workshops in parallel programming were conducted and Challenge participants got down to parallelizing their code with techniques such as MPI (Message Passing Interface), OpenMP, parallel algorithms, and use of parallel applications and tools. The winning entries demonstrated the researchers’ ability to make their parallel programs run optimally.  For example, a speedup of about 14 times using 15 processors was obtained.  Entries in the Grid category have demonstrated very effective use of Grid computing resources.  Our NUS Grid or TCG@NUS harvests the idle cycles of up to 1500 Windows and Linux based computers across campus. Beside data intensive applications, the Grid has proven effective for parameter estimation studies as shown in the entries in this year’s Challenge.  One project achieved a speedup of over 80 times for a single computation job.  Another entry reduced the job completion time from 386 days to under 8 days.

A prize award presentation and seminar by winners was held on 24th November.  Please find on this page information on the Challenge winners and some details of their winning projects. 

Serial to Parallel Category

Gold award: Dr. Benny Raphael
Project title: Computation of Lighting Energy Consumption using Parallelized Day Lighting Simulation

Abstract:
Lighting simulation and energy computation are very practical tasks and of increasingly importance due to effects of global warming and climate change.  Architects are interested in evaluating the energy consumption of building systems to select the best design options. This is possible only if simulations can be completed fast. Lighting simulation takes much time when the model contains many surfaces, such as when objects like trees are included. In this research, the computation time was reduced by developing a parallel algorithm and implementing it using Message Passing Interface (MPI) library.  An existing problem that took about 8 hours to complete on a windows machine now completed in under an hour on the parallel computing cluster. The maximum speedup was 10.8 using 12 processors and average efficiency was 0.76. Runs using different number of processors show linear scaling.
Full Report

Silver award: Dr. Yu Peng
Project title: A Hybrid Meshfree-Cartesian Scheme to Simulate Complex Moving Body Problems in Three-Dimensional Space

Abstract:
Flows involving moving bodies/boundaries are frequently encountered in engineering applications and everyday living.  Examples are flows in turbomachines, artificial heart valves, extrusion processes, fish swimming and bird/insect flight. These are unsteady flows in which the fluid boundaries are constantly changing. The project aims to develop a code to simulate 3D moving boundary problems, which combines the advantages of existing approaches. A SVD-GFD based hybrid meshfree-Cartesian grid scheme has been successfully implemented to simulate several moving boundary problems. It is speeded up by utilizing OpenMP technique with more than 95% of the code parallelized. By systemically considering coverage, granularity, load balancing, locality and synchronization effects, the code achieves speedup of around 10 for 16 CPU-cores, with a parallel efficiency of 61%. Simulations which take two months to finish on a single core can now complete within a week on 16 CPU cores.
Full Report  Presentation

Merit award: Mr. Shyam Sundar Dhanabalan
Project title: Parallel Computing and Dynamic Load Balancing for DG-WENO schemes

Abstract:
With high performance computers, many realistic Computational Fluid Dynamics (CFD) simulations such as flow around an aircraft, combustion in internal combustion engines etc. are now realizable. However, obtaining an accurate solution representing all the flow characteristics is still a challenge. Such flow solvers should be able to solve problems of different scales of fluctuations, ranging from strong shocks to acoustic noise generation and propagation. In this aspect, the higher order Weighted Essentially Non Oscillatory (WENO) schemes and Discontinuous Galerkin (DG) methods were developed.  In this work, a higher order DG scheme, formulated with a new higher order compact WENO scheme for problems involving strong shocks, is reported for the first time.  The DGWENO code is parallelized using MPI libraries and tested on two different clusters.  Significant improvement in computational time is achieved with average parallel efficiency of around 93% and 97%.
Full Report

Grid Category

Gold award: Ms. Hu Yongli
Project title: Accelerating Evolutionary Molecular Phylogenetic Analyses on the NUS TCG Grid

Abstract:
Molecular phylogenetics is the science of estimating the evolutionary past with sequence information. However with the advancement of sequencing technologies, tremendously large amount of sequence data is generated and they can no longer be analyzed within an acceptable time frame. Therefore, to accelerate and enable the analysis of large protein sequence datasets in phylogenetic studies, a version of PHYLIP has been successfully placed on the NUS TCG grid as an application called meta-PHYLIP. Harvesting idle cycle of grid clients, analyses of 14 different dataset have yielded a significant speedup of about 14 to 65 times. The grid-enabled meta-PHYLIP will alleviate the bottle-neck in large scale phylogenetic analyses and renders previously unanalyzed protein sequences analyzable.
Full Report  Presentation

Silver award: Mr. Liu Bing
Project title:  Statistical Analysis of Partially Calibrated Bio-pathway Models

Abstract:
Understanding the functioning of complex biological systems is a major challenge. A standard formalism used to model bio-pathways is a system of ordinary differential equations (ODEs). Bio-pathways usually involve a large number of molecular species and bio-chemical reactions. The values of parameters (rate constants) are often unknown because they are very difficult to be measured. Thus, we have to resort to computational methods and estimate them from experimental data by collective fitting. As the noisy data are often insufficient to constrain all parameters to have a unique solution, we propose to perform parameter estimation multiple times with different starting points. We then can gain biological insights by statistically analyzing the estimated parameters. Although this process is very time-consuming, it can be easily parallelized. We tested our approach on Complement system and achieved speedup of 82.1 by enabling our application on the Grid.
Full Report  Presentation

Merit award: Mr. Omid Khanmohamadi
Project title:  GARCH Models, Noise, and Particle Filters: in Pursuit of Parameters

Abstract:
The problem dealt with is parameter estimation of a price model. The underlying model is of general autoregressive conditional heteroscedasticity (GARCH) type, allowing the variance of the current error term to depend on those of previous time points.  Presence of noise and nonlinearity of the model necessitate using a particle filter (of sequential Monte Carlo class), since Kalman filter cannot be used to estimate a nonlinear model unless it is linearized around some equilibrium state, which would lead to erroneous results as it is known that prices can diverge from equilibrium values due to microstructure noises.  The implementation is a combination of C and MATLAB. The particle filter is implemented in C and, through a mexFunction gateway, the compiled file is ported to the MATLAB part of the program. Results show that a particular run of the application completes in 4 days compared with CPU time of over 112 days.
Full Report

Parallel Speedup Category

Gold award: Dr. Yang Lifeng
Project title: DACH-Platinum-DNA Interactions: An Investigation into the Basis for Binding and Recognition by Parallel Molecular Modelling

Abstract:
Stereoisomers of the DACH group which formed 1,2-GG intrastrand cross-links with DNA were studied by molecular dynamics (MD) simulations using AMBER. The calculated binding energies between R,R-DACH-Pt, S,S-DACH-Pt and cis-DACH-Pt moieties and DNA revealed a strong correlation with antitumour activities. The result provided more clues to understand the biological interactions of chiral platinum drugs. DNA structure analysis indicated that DNA tolerated the distortion resulted in the different Pt-DNA adducts and various local and global structure distortions were found. In this work, we design a protocol for the estimation of the free energy of molecular conformations of DACH-Pt and DNA using the MM/PBSA approach. Our goal was to compare the free binding energies of the corresponding R,R-DACH-Pt-DNA, S,S-DACH-Pt-DNA and cis-DACH-Pt-DNA pairs. These results suggest that MM/PBSA free energy estimation is good for Pt/DNA moieties.  A typical simulation in a 10ns time scale for one complex can run for 3 days using 16 CPUs. We observed linear speedup of the Amber 9 code, showing that it is ideally parallelized for this algorithm.

Silver award: Dr. Tho Kee Kiat
Project title: Numerical Analysis of Riser Seabed Interaction

Merit award: Dr. Zeng Yan
Project title: Computational Fluid Dynamics Analysis on Drag Reduction for Grid-Fin Configurations

Abstract:  
A normal choke usually forms at the back of the lattice grid cells at transonic Mach numbers, causing a significant increase in drag force. The present study proposed an improved sweptback grid fin to reduce the transonic choking, so as to reduce the drag. Viscous computational fluid dynamic (CFD) simulations were performed to investigate flows over an ogive cylinder with baseline and sweptback grid fins in transonic and supersonic flows of Mach number 0.8-2.0, at zero angle of attack. The present simulations show that the choke can be reduced by using the sweptback grid fins which are found to reduce drag by approximately 5% for the overall drag and nearly 12% on the fin part, for all the Mach numbers investigated.  Parallel processing was used to speed up the computation.  For the present cases which usually need 5-7 days using 8-CPU cores, 25-35 days may be needed to finish one case for a single CPU running.
Full Report

 

Sponsors

Platinum

Silver