120 likes | 234 Views
High Performance Geo-Simulations. Saulo Henrique Cabral Silva Tiago Garcia de Senna Carneiro* TerraLAB – Laboratory for Earh System Modeling and Simulation Computer Science Department Federal University of Ouro Preto GeoInfo 2013. Challenges.
E N D
High Performance Geo-Simulations Saulo Henrique Cabral Silva Tiago Garcia de Senna Carneiro* TerraLAB – Laboratory for Earh System ModelingandSimulation ComputerScienceDepartment Federal Universityof Ouro Preto GeoInfo 2013
Challenges • Modelers have few programming expertise • Problem partition into several processes? • Processes communication? • Processes synchronization? • What is the design goal: Scalability or Performance? • Global scheduling scalabitity • Local scheduling high performance • How to parallelize/distribute workload? • Model architecture can be defined by: • Application domain area (the phenomenon being modeled) • Chosen modeling paradigm: CA, multiagente, DEVS, etc • Which load balancing strategy to use? • Agents clusters • Dynamic and adaptative • How to verify models with no influence of concurrency control rules?
Flexibility through annotation: Prey-predator model Annotations - API: HPA PARALLEL HPA JOINfunction_name HPA JOINALL HPA ACQUIREresource_name HPA RELEASEresource_name function execAgentPred(i) predatores[i]:execute() end function execAgentPrey(i) preys[i]:execute() end for i = 1,#predatores do --HPA PARALLEL execAgentPred(i) end --HPA JOINALL for i = 1,#preys do --HPA PARALLEL execAgentPrey(i) end --HPA JOINALL
Discussion • TerraME HPA scales well when the modeler designs efficient parallel solutions • 90% of linear speedup with 8 processors (prey predator model) • 35% of linear speedup with 8 processors (TROLL model)
EXP1: Spatially-explicit prey predator model • Species compete, evolve and disperse, in order to seek resources to support their struggle for their existence • Each cell has a population of prey and a population of predators • Model executes in two steps: • First, predators are simulated in parallel • Then, prey are simulated in parallel
EXP2: 3D tropical rain forest growth model - TROLL • TROLL model simulates a 3D environment where: • Individual trees compete for light • Seed dispersal by wind • Trees interacts falling over others and forming gaps • … • Adopted partitioning strategy: each layer (h, h-1, h-2, …) is simulated in parallel • Higher layers are simulated before lower layers