1. Parallel Landscape Fish Model for South Florida Ecosystem Simulation Dali Wang 1, Michael W. Berry 2, Eric A. Carr 1, Louis J. Gross 1 The ATLSS Hierarchy of Models Computational Environments The Structure of ALFISH Model Parallel Computational Model Simulation Results Selected References  D. Wang, M. W. Berry, E. Carr, L. J. Gross, Design and Implementation of Dynamic Fish Model in South Florida on Parallel Architecture, Proceedings of 37th Hawaii International Conference on System Sciences, 2004  H.Gaff, D. DeAngelis, L. Gross, R. Salinas, M. Shorrosh, A Dynamic Landscape Model for Fish in the Everglades and its Application to Restoration, Ecological Modeling, 127 (2000), pp. 33--53.  H. Gaff, J. Chick, J. Trexler, D. L DeAngelis, L. J. Gross, R. Salinas, Evaluation of and Insights from ALFISH: a Spatial-explicit, Landscape-level Simulation of Fish Populations in the Everglades, Hydrobiologia (to appear).  D. Wang, M. W. Berry, E. Carr, L. J. Gross, A Parallel Fish Landscape Model for Ecosystem Modeling on a Computing Grid, Parallel and Distributed Computing Practices (in review) Conclusions 1 The Institute for Environmental Modeling, 569 Dabney Hall, University of Tennessee, Knoxville, TN 37996 wang@tiem.utk.edu, carr@tiem.utk.edu, gross@tiem.utk.edu 2 Department of Computer Science, University of Tennessee, Knoxville, TN 37996 berry@cs.utk.edu An ecological multimodel designed to assess the effects on key biota of alternative water management plans for the regulation of water flow across the Everglades landscape. ALFISH is an Intermediate Trophic Level Functional Groups model which includes two main subgroups (small planktivorous fish and large piscivirous fish), structured by size. 1Gbs Ethernet Network with dedicated switches System bus Memory Proc System bus Memory Proc System bus Memory Proc 16 nodes 3.2 Gbs System Interconnection 10 G Main Memory 8M Ext Cache Proc 2(16K) Cache Proc 2(16K) Cache Proc 2(16K) Cache 14 CPUs Processors 14(400MHz) Architecture UltraSPARC L1 Cache: 16-KB / 16-KB Ext Cache: 8 MB System Interconnect 3 GB/sec Main Memory 10 GB Storage 1.5 TB Processors 2(450MHz) Architecture UltraSPARC L1 Cache: 16-KB / 16-KB Ext Cache: 4 MB Main Memory 0.5 GB Storage 30 GB System Connect 1 GB/sec Initialization Update water and lower trophic data Density-independent fish movement Diffusive fish movement Output Final output End of run? yes no Mortality/Aging/Growth/Reproduction Initialization Update water and lower trophic data Density-independent fish movement Diffusive fish movement Output End of run? no Mortality/Aging/Growth/Reproduction Initialization Final output yes Receive Date Repartition the landscape (MPI_SENDRECV) End of run? yes Move data between boundaries (MPI_Send) Finalization (MPI_Broadcast) no Master ProcessorComputational Processors Initialization Update lower trophic data Update water data from memory Density-independent fish movement Diffusive fish movement Save output from previous step Mortality/Aging/Growth/Reproduction Final output yes (Barrier_wait) (pthread_cond_wait) (Barrier_wait) End of run? (pthread_signal) no Master ThreadComputational Threads Create threads Update water data for next step Save date to memory (Barrier_wait) Join threads repartition boundary (pthread_cond_wait) (pthread_signal) This research has been supported by the National Science Foundation under grant No. DEB-0219269. This research used the resources of the Scalable Intracampus Research Grid (SInRG) Project at the University of Tennessee, supported by the National Science Foundation CISE Research Infrastructure Award EIA-9972889. The sequential implementation was developed with support from the U.S. Geological Survey, through Cooperative Agreement No. 1445-CA09-95-0094 with the University of Tennessee, and the Department of Interior's Critical Ecosystem Studies Initiative. Grid-based partitioning can be highly effective for age- and size-structured spatially explicit landscape fish models. Adapting dynamic partitioning, the fast model turnaround time is about 2.5 hours yielding a speedup factor of 12. The MPI implementation requires substantial (main) memory and is suitable for execution on clusters with high speed interconnection. Compared to the MPI implementation, a Pthread implementation of the model requires less (main) memory and more scalable, but is more workload sensitive. Objectives To compare, in a spatially explicit manner, the relative effects of alternative hydrologic scenarios on fresh-water fish densities across South Florida. To provide a measure of dynamic, spatially-explicit food resources available to wading birds. Speedup model performance to allow real-time data generation and information sharing between models. October 1January 1April 1 Performance Performance of Static and Dynamic Landscape Partitioning Each Node Whole System Spatial averaged (over 31 years) fish density map in Everglades. No discernable differences are observed in spatial outputs form the different parallel implementations. 8M Ext Cache Radio-telemetry Tracking Tools Abiotic Conditions Models Spatially-Explicit Species Index Models Linked Cell Models Process Models Age/Size Structure Models Individual-Based Models High Resolution Hydrology High Resolution Topography Disturbance Cape Sable Seaside Sparrow Snail Kite Long-legged Wading Birds Short-legged Wading Birds White-tailed Deer Alligators Lower Trophic Level Components Vegetation Fish Functional Groups AlligatorsReptiles and Amphibians White-tailed Deer Florida Panther Snail Kite Wading Birds © TIEM / University of Tennessee Cape Sable Seaside Sparrow Static PartitionDynamic partition