Postdoctoral Position : Agent-Based Modeling of Social Complexity in Ancient Egypt : University of Cape Town, South Africa
UCT CS Colloquium by Marshini Chetty of the University of Maryland, USA
CS Colloquia and Seminars and other talks
Document Actions
Colloquium - GPU Programming for Simulation
| What | Colloquium |
|---|---|
| When |
2009-07-30 13:00
2009-07-30 13:45
2009-07-30 from 13:00 to 13:45 |
| Where | CS LT302 |
| Contact Name | James Gain |
| Contact Email | jgain@cs.uct.ac.za |
| Add event to calendar |
vCal
iCal
|
Speaker: Ian Tunbridge
Date: 30 July 2009
Time: 13h00-13h45
Venue: CS LT302
Abstract:
Graphics processing unit (GPU) architectures provide the means to migrate algorithms from the SISD paradigm, synonymous with CPU architectures, to the SIMD paradigm. The benefit of commodity multicore hardware with general programmability is significant speedups for migrated codes. We describe the implementation of a coarse-grained model for molecular simulation on a GPU, using a Replica Exchange Monte Carlo simulation method. Coarse-grained molecular models provide reduced complexity when compared to the traditional, computationally expensive, all-atom models. However, while the coarse-grain model is much less computationally expensive than the all-atom approach, pair-wise energy calculations required at each iteration of the algorithm continue to cause a computational bottleneck in serial implementations. Our implementation of coarse-grained models on the highly parallel GPU hardware vastly increases the size and time scales accessible to molecular simulation. We describe the nontrivial process of migrating the algorithm to a GPU and the effect of various GPU techniques on algorithm speedup. Benchmarking and profiling shows that the GPU implementation scales favourably compared to a CPU implementation. While simulation times for our CPU implementation grow quadratically with respect to problem size, times from our GPU implementation grow linearly. In effect, the parallel nature of the GPU architecture transforms the performance bottleneck in the algorithm from O(N²) to O(N).
Bio:
Ian Tunbridge will be discussing his work as an MSc student in the Computer Science department