Monte Carlo Simulation of Molecular Systems - PowerPoint PPT Presentation

1 / 20
About This Presentation
Title:

Monte Carlo Simulation of Molecular Systems

Description:

Quality of random numbers. Random Number Generators. Linear Congruential Generator ... RNG Name. The L'Ecuyer paper -- http://www.iro.umontreal.ca/~simardr ... – PowerPoint PPT presentation

Number of Views:118
Avg rating:3.0/5.0
Slides: 21
Provided by: coursesEc
Category:

less

Transcript and Presenter's Notes

Title: Monte Carlo Simulation of Molecular Systems


1
Monte Carlo Simulation of Molecular Systems
  • Faycal Benmlih, John Sartori, Alan Kaatz
  • in conjunction with
  • John Stone, Peter Freddolino

2
Overview
  • Monte Carlo Simulation of Protein Conformation
  • Deviation From Original Plan
  • Implementation
  • Energy Calculation
  • Random Number generation and Perturbation
  • CPU Orchestration
  • Final Results and Summary

3
Introduction
  • What is conformation anyway?
  • Why is it useful?

pharm.wayne.edu
4
Deviation From Original Plan
5
Deviation From Original Plan
6
Who is doing what?
7
Energy Calculation Breakdown
  • Block size energy kernel
  • Inter-block energy kernel
  • Reduction kernel

8
(No Transcript)
9
Optimizations
  • Reduce the register usage
  • Minimize global memory accesses.
  • Move redundant computations outside of loops.
  • Use a single array for both block size energy and
    inter-block energy calculations.
  • Eliminates the need of combining the two array
    before reducing them.

10
(No Transcript)
11
Monte Carlo Perturbation
  • Agitate molecules

12
Random Number Generation Implementations
  • Generate the random numbers on the CPU and
    transfer them to the GPU
  • Have a dedicated kernel generate numbers and
    store them in global memory
  • Have each perturbing thread generate random
    numbers as needed

13
GPU RNG Issues
  • Loading/Saving RNG state vector
  • Initial seeding of every RNG thread
  • RNG computational complexity
  • Quality of random numbers

14
Random Number Generators
  • Linear Congruential Generator
  • Linear Feedback Shift Register

15
Random Number Generation
16
Random Number Generators
The LEcuyer paper -- http//www.iro.umontreal.ca/
simardr/testu01/tu01.html
17
(No Transcript)
18
(No Transcript)
19
GPU vs. CPU Speedup Bottlenecks?
  • Lack of inter-block or global memory
    synchronization requires MC to have multiple
    kernels per simulation step
  • Must send some information periodically back to
    host or have one-thread kernel
  • Cannot mask all global memory transfers in energy
    calculation and perturbation

20
(No Transcript)
Write a Comment
User Comments (0)
About PowerShow.com