SFHG

From HP-SEE Wiki

Latest revision as of 09:40, 5 March 2013

General Information

• Application's name: Self Avoiding Hamiltonian Walk on Gaskets
• Application's acronym: SFHG
• Virtual Research Community: Computational Physics Applications
• Scientific contact: Sreten Lekic, slekic@blic.net
• Technical contact: Sreten Lekic, slekic@blic.net
• Developers: Sreten Lekic, Faculty of Mech. Engineering, University of Banja Luka (UoBL), Bosnia - Herzegovina, Igor Sevo, Mihajlo Savic Faculty of Electrical Engineering, University of Banja Luka (UoBL), Bosnia - Herzegovina
• Web site: http://wiki.hp-see.eu/index.php/SFHG

Short Description

Hamiltonian self avoiding walks on fractals (Sierpinsky gaskets) are one of perspective models for long polymers (DNA, RNA and other bio polymers) behavior description.

We have developed a program for counting self-avoiding Hamiltonian walks to run on multiple processors in a parallel mode. We study Hamiltonian walks (HWs) on the family of two-dimensional modified Sierpinski gasket fractals, as a simple model for compact polymers in nonhomogeneous media in two dimensions. We apply an exact recursive method which allows for explicit enumeration of extremely long Hamiltonian walks of different types: closed and open, with end-points anywhere in the lattice, or with one or both ends fixed at the corner sites. The leading term n is characterized by the value of the connectivity constant 1, which depends on fractal type, but not on the type of HW.

Problems Solved

New serial C++ code produced. Parallelization of serial C++ code done in OpenMP.

Scientific and Social Impact

Collaborations

Beneficiaries

Faculty of Science. Dept of Physics

Number of users

3

Development Plan

• Concept: 2012-06-01
• Start of alpha stage: 2012-10-01
• Start of beta stage: 2013-01-01
• Start of testing stage: 2013-01-15
• Start of deployment stage: 2013-02-15
• Start of production stage: 2013-03-01

Resource Requirements

• Number of cores required for a single run: up to 64
• Minimum RAM/core required: <1GB
• Storage space during a single run: <1GB
• Long-term data storage: <1GB
• Total core hours required: <32000

Technical Features and HP-SEE Implementation

• Primary programming language: C/C++
• Parallel programming paradigm: OpenMP
• Main parallel code: Nested OpenMP
• Pre/post processing code: N/A
• Application tools and libraries: libgomp, gcc, gprof

Usage Example

Infrastructure Usage

• Home system: PARADOX
  • Applied for access on: 2010-09-01
  • Access granted on: 2010-09-01
  • Achieved scalability: 8
• Accessed production systems: .

1. PARADOX
   • Applied for access on: 2010-09-01
   • Access granted on: 2010-09-01
   • Achieved scalability: 8
2. Pecs SC
   • Applied for access on: 2013-01-18
   • Access granted on: 2013-01-21
   • Achieved scalability: 48
3. Szeged SC
   • Applied for access on: 2013-01-18
   • Access granted on: 2013-01-21
   • Achieved scalability: 24

• Porting activities: Emulated implementation of OpenMP functions not implemented in older versions of gcc/libgomp
• Scalability studies: Scalability studies performed at PARADOX (up to 8 CPU cores), BA-01-ETFBL (up to 16 CPU cores) and Pecs SC (up to 48 CPU cores). Detailed results available in Deliverable D8.4.

Running on Several HP-SEE Centres

Application is currently running at PARADOX, Pecs SC and Szeged SC.

• Benchmarking activities and results: Performed scalability study for D8.4. Achieved 207k Walks/s for level 8 and 132k Walks/s for level 9.
• Other issues: Issues with incomplete or missing support for specific OpenMP features with Open64 and Portland Group compilers.

Achieved Results

• We have calcualted exact number of walks for previously unavailable levels 8 and 9 for all needed types of walks.
• We have created a novel approach to solution of the problem without resorting to brute-force walk counting.

Publications

Foreseen Activities

• Optimization of parallelization approach to tasks - partially completed.
• Creating MPI code.
• Creating hybrid MPI+OpenMP code.
• Performing all types of walks at the same time (most likely via MPI).