Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • harvard1
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Social Simulations Accelerated: Large-Scale Agent-Based Modeling on a GPU Cluster
University of Borås, Swedish School of Library and Information Science.
2013 (English)Conference paper, (Other academic)
Abstract [en]

Pandora is an agent-based modeling framework designed for the Humanities and Social Sciences. Simulation is gaining ground in these fields, as such models naturally capture how knowledge is acquired and accumulated in human societies, and they provide a more natural fit than abstract mathematical models. Real-world terrain, climate data, precipitation, and many other parameters can be included. Emergent behavior can typically observed only in large-scale simulations, and hence Pandora uses MPI and OpenMP to run simulations on a high-performance cluster. This poster is a case study in which we accelerated the critical computational loop of a simulation scenario by up to 7x using Thrust and a Tesla GPU. The acceleration is crucial to verify a larger number of alternative simulation scenarios, as an extensive range of parameter combinations are valid and may lead to fundamentally different conclusions. To achieve the speedup, we had to completely redesign the data access pattern, which was originally optimized for latency-oriented CPUs with sizable cache memory and a cache coherence protocol. The "horizon" of an individual agent causes the majority of irregular memory accesses. Moreover, there are thousands of agents in a simulation, each with its own horizon. We introduced a stencil to arrive at a rectangular memory region, and we perform a 2D stencil-based reduction to update an agents knowledge. While the acceleration naturally maps to a GPU cluster, further work is necessary to achieve a consistent speedup outside the critical region. Currently only the computationally most demanding region, the update of an agent's knowledge is accelerated, leading to a very low, 4-6 % utilization of the GPU. By performing all actions of an agent on the GPU, multiple agents could be executed simultaneously, leading to a much higher utilization.

Place, publisher, year, edition, pages
2013.
Keyword [en]
GPU, Agent-Based Modelling
National Category
Computer and Information Science
Research subject
Library and Information Science
Identifiers
URN: urn:nbn:se:hb:diva-7166Local ID: 2320/13342OAI: oai:DiVA.org:hb-7166DiVA: diva2:887874
Conference
GPU Technology Conference 2013, San José, CA, USA
Note

Sponsorship:

The second author is supported by the SimulPast Project (CSD2010-00034), funded by the CONSOLIDER-INGENIO2010 program of the Ministry of Science and Innovation -- Spain. This work is also supported by AWS in Education Machine Learning Grant award.

Available from: 2015-12-22 Created: 2015-12-22

Open Access in DiVA

No full text

Search in DiVA

By author/editor
Wittek, Peter
By organisation
Swedish School of Library and Information Science
Computer and Information Science

Search outside of DiVA

GoogleGoogle Scholar

Total: 51 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • harvard1
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf