Difference between revisions of "PhyloSoC: DIM SUM 2 GPU computing for individual based simulator of movement and demography"

From Phyloinformatics
Jump to: navigation, search
(Project Plan)
Line 7: Line 7:
 
* [http://www.phyleauxgenetics.org/ Jeremy Brown]
 
* [http://www.phyleauxgenetics.org/ Jeremy Brown]
 
==Project Plan==
 
==Project Plan==
===Week 1 (May 23 - May 30):===
+
===Week 1 (May 23 - May 29):===
 
* Performance analysis of DIM SUM.
 
* Performance analysis of DIM SUM.
 
* Split populateAndMigrateThreaded into smaller function to better understand it.
 
* Split populateAndMigrateThreaded into smaller function to better understand it.
===Week 2 (May 30 - June 6):===
+
===Week 2 (May 30 - June 5):===
 
* Replace the ArrayLists of Objects in DispersalSettings, which are needed for the computations in the populateAndMigrateThreaded, with data structure, which can be copied in the graphic memory. The only acceptable data structures for CUDA are arrays of primitives. That means e.g. that all ArrayList<PFunction> must be replaced by a two dimension array for the outcomes and a two dimensional array for the probabilities.
 
* Replace the ArrayLists of Objects in DispersalSettings, which are needed for the computations in the populateAndMigrateThreaded, with data structure, which can be copied in the graphic memory. The only acceptable data structures for CUDA are arrays of primitives. That means e.g. that all ArrayList<PFunction> must be replaced by a two dimension array for the outcomes and a two dimensional array for the probabilities.
===Week 3 (June 6 - June 13):===
+
===Week 3 (June 6 - June 12):===
* Start replacing the ArrayLists, which stores the generation with CUDA compatible data structures. Even thought this are only one or two ArrayLists this task might be more time consuming than the first one, because there are elements added and deleted to this data structure during the simulation and I think the GPU can't resize an array.
+
* Replace the ArrayList, which stores the generation with a CUDA compatible data structure. Even thought it is only one ArrayList this task isn't trivial consuming because elements are removed from this array during the computation, which isn't possible on the GPU.
===Week 4 (June 13 - June 20):===
+
===Week 4 (June 13 - June 19):===
 
* Replace lldata with an Array, which stores the double values.
 
* Replace lldata with an Array, which stores the double values.
===Week 5 (June 20 - June 27):===
+
===Week 5 (June 20 - June 26):===
* Finish replacing the ArrayLists, which stores the generation with CUDA compatible data structures.
+
* Replace the Nodes with a data structures, which can be used on the graphics card.
===Week 6 (June 27 - July 4):===
+
* Remove variables from the Nodes, which are only needed for debugging.
 +
===Week 6 (June 27 - July 3):===
 
* Port the functions in XYFunction and DispersalSettings,which are needed by the migrate function, to CUDA.
 
* Port the functions in XYFunction and DispersalSettings,which are needed by the migrate function, to CUDA.
===Week 7 (July 4 - July 11):===
+
===Week 7 (July 4 - July 10):===
 
* Port the migrate function to CUDA. Even through this function is not used anymore. Porting this function to CUDA is a good preparation for porting populateAndMigrateThreaded. The parallelization of this function should be relative easy. The basic Idea is that the computation per child are independent from each other, so they can be executed in parallel. I think the only problem is that elements are removed from the children array, which can be solved by marking the elements (e.g. in a boolean array) and remove them on the CPU.
 
* Port the migrate function to CUDA. Even through this function is not used anymore. Porting this function to CUDA is a good preparation for porting populateAndMigrateThreaded. The parallelization of this function should be relative easy. The basic Idea is that the computation per child are independent from each other, so they can be executed in parallel. I think the only problem is that elements are removed from the children array, which can be solved by marking the elements (e.g. in a boolean array) and remove them on the CPU.
===Week 8 (July 11 - July 18):===
+
===Week 8 (July 11 - July 17):===
 
* Finish to port all the functions in DispersalSettings, Pfunction and XYFunction, which are needed during the computations, to CUDA.
 
* Finish to port all the functions in DispersalSettings, Pfunction and XYFunction, which are needed during the computations, to CUDA.
===Week 9 (July 18 - July 25):===
+
===Week 9 (July 18 - July 24):===
 
* Port all the functions in lldata to CUDA.
 
* Port all the functions in lldata to CUDA.
===Week 10 (July 25 - August 1):===
+
===Week 10 (July 25 - July 31):===
 
* Start porting the populateAndMigrateThreaded function to CUDA. The function is already parallelized. Everything which is executed in parallel in DIM SUM should now be executed on the GPU. That means that the run function in DispersalThread has to be ported to the GPU. Like with the migrate function the data structures are the main concern.
 
* Start porting the populateAndMigrateThreaded function to CUDA. The function is already parallelized. Everything which is executed in parallel in DIM SUM should now be executed on the GPU. That means that the run function in DispersalThread has to be ported to the GPU. Like with the migrate function the data structures are the main concern.
===Week 11 (August 1 - August 8):===
+
===Week 11 (August 1 - August 7):===
 
* Finish porting the populateAndMigrateThreaded function to CUDA. The function is already parallelized. Everything which is executed in parallel in DIM SUM should now be executed on the GPU. That means that the run function in DispersalThread has to be ported to the GPU. Like with the migrate function the data structures are the main concern.
 
* Finish porting the populateAndMigrateThreaded function to CUDA. The function is already parallelized. Everything which is executed in parallel in DIM SUM should now be executed on the GPU. That means that the run function in DispersalThread has to be ported to the GPU. Like with the migrate function the data structures are the main concern.
===Week 12 (August 8 - August 15):===
+
===Week 12 (August 8 - August 14):===
 
* Bug-fixing.
 
* Bug-fixing.
 
* Code cleanup.
 
* Code cleanup.
 
* Improve documentation.
 
* Improve documentation.

Revision as of 16:55, 6 June 2011

Abstract

DIM SUM is a population demography and individual migration simulation. The goal of the project is to use the graphics card to improve the speed of DIM SUM significantly, so it is possible to simulate scenarios with much larger populations and larger landscapes.

Student

Mentors

Project Plan

Week 1 (May 23 - May 29):

  • Performance analysis of DIM SUM.
  • Split populateAndMigrateThreaded into smaller function to better understand it.

Week 2 (May 30 - June 5):

  • Replace the ArrayLists of Objects in DispersalSettings, which are needed for the computations in the populateAndMigrateThreaded, with data structure, which can be copied in the graphic memory. The only acceptable data structures for CUDA are arrays of primitives. That means e.g. that all ArrayList<PFunction> must be replaced by a two dimension array for the outcomes and a two dimensional array for the probabilities.

Week 3 (June 6 - June 12):

  • Replace the ArrayList, which stores the generation with a CUDA compatible data structure. Even thought it is only one ArrayList this task isn't trivial consuming because elements are removed from this array during the computation, which isn't possible on the GPU.

Week 4 (June 13 - June 19):

  • Replace lldata with an Array, which stores the double values.

Week 5 (June 20 - June 26):

  • Replace the Nodes with a data structures, which can be used on the graphics card.
  • Remove variables from the Nodes, which are only needed for debugging.

Week 6 (June 27 - July 3):

  • Port the functions in XYFunction and DispersalSettings,which are needed by the migrate function, to CUDA.

Week 7 (July 4 - July 10):

  • Port the migrate function to CUDA. Even through this function is not used anymore. Porting this function to CUDA is a good preparation for porting populateAndMigrateThreaded. The parallelization of this function should be relative easy. The basic Idea is that the computation per child are independent from each other, so they can be executed in parallel. I think the only problem is that elements are removed from the children array, which can be solved by marking the elements (e.g. in a boolean array) and remove them on the CPU.

Week 8 (July 11 - July 17):

  • Finish to port all the functions in DispersalSettings, Pfunction and XYFunction, which are needed during the computations, to CUDA.

Week 9 (July 18 - July 24):

  • Port all the functions in lldata to CUDA.

Week 10 (July 25 - July 31):

  • Start porting the populateAndMigrateThreaded function to CUDA. The function is already parallelized. Everything which is executed in parallel in DIM SUM should now be executed on the GPU. That means that the run function in DispersalThread has to be ported to the GPU. Like with the migrate function the data structures are the main concern.

Week 11 (August 1 - August 7):

  • Finish porting the populateAndMigrateThreaded function to CUDA. The function is already parallelized. Everything which is executed in parallel in DIM SUM should now be executed on the GPU. That means that the run function in DispersalThread has to be ported to the GPU. Like with the migrate function the data structures are the main concern.

Week 12 (August 8 - August 14):

  • Bug-fixing.
  • Code cleanup.
  • Improve documentation.