INDIVIDUAL CONTRIBUTION (Hardware Acceleration)

Home‎ > ‎

INDIVIDUAL CONTRIBUTION

Project Milestone Date: Nov/11/2009

Team Member	Contribution
Hua Cheng	Read the below paper to understand the mathematics and algorithms of AMG - Falgout, Robert D., Van Emden Henson, Jim E. Jones, and Ulrike Meier Yang, "BoomerAMG: A Parallel Implementation of Algebraic Multigrid," presented by Van Emden Henson at the 9th SIAM Conference on Parallel Processing, San Antonio, TX, March 22, 1999. UCRL-MI-133583 Create and maintain project webpage Helped compile AMG benchmark on herry2 and CELLBE Compile AMG benchmark with gprof option and mpiP benchmark Helped to identify the AMG benchmark computation hotpot with different processor size and argument topolgy Initiate the process of compile the CML on cell - identify and resolve the compile error Helped to compile the AMG benchmark program with test ppu and spu programs on CELLBE - build 3 layers of compiling hyachy: host, ppu, spu - link test spu with ppu layer - working on compile ppu layer with host program Worked on parallizing the hypre_CSRMatrixMatvec function on the ppu and spu layer - identify the parallel region - create data struct to do dma transfer
Brandon Dwiel	Read the following papers and presentations to understand the mathematics and algorithms of AMG ▪ Robert D. Falgout and Panayot Vassilevski, "On Generalizing the AMG Framework." Presented at Schnelle Löser für Partielle Differentialgleichungen, Oberwolfach, Germany, June, 2003. UCRL-PRES-150807 ▪ Robert D. Falgout, "Adaptive Algebraic Multigrid." Presented at the Seventh European Multigrid Conference, Hohenwart Forum, Germany, October, 2002. UCRL-PRES-149422 ▪ A. Kejariwal and C. Cascaval, “Parallelization Spectroscopy: Analysis of Thread-level Parallelism in HPC Programs.” In The 2nd Workshop on Parallel Execution of Sequential Programs on Multi-core Architectures (PESPMA) 2009, pages 30-39, June 2009. ◦ Helped profile the AMG benchmark for different processor numbers and input sizes to determine the hot spots ▪ Determined how to profile all tasks and collectively gather the results ◦ Helped to try and compile the CML ▪ Investigated and solved errors ◦ Worked on parallelizing the hypreBoomerAMGBuildCoarseOperator function ▪ Identified the loops that will likely show the most benefit from being ported to the SPE • Classified all variables as either private, read-only shared or read-write shared ◦ Helped develop a test program to test using MPI on the CellBE with both the PPE and SPE ◦ Helped develop a test program that will used as a template for us to call our SPE kernels from inside the benchmark ▪ This will allow us to call a function from anywhere in the AMG benchmark that will run on the PPE that will initialize the required data and call the SPE and return the necessary values back.
Unnati Ojha	• Read the following papers and presentations to understand the mathematics of AMG Robert D. Falgout and Panayot Vassilevski, "On Generalizing the AMG Framework." Presented at Schnelle Löser für Partielle Differentialgleichungen, Oberwolfach, Germany, June, 2003. UCRL-PRES-150807 Robert D. Falgout, "Adaptive Algebraic Multigrid." Presented at the Seventh European Multigrid Conference, Hohenwart Forum, Germany, October, 2002. UCRL-PRES-149422 Robert D. Falgout, "HYPRE: High Performance Preconditioners." Presented at the DAM/DOE Meeting on Solvers, CEA/DAM Bruyères le Châtel, France, February, 2004. UCRL-PRES-140609-REV-1 Falgout, Robert D., Van Emden Henson, Jim E. Jones, and Ulrike Meier Yang, "BoomerAMG: A Parallel Implementation of Algebraic Multigrid," presented by Van Emden Henson at the 9th SIAM Conference on Parallel Processing, San Antonio, TX, March 22, 1999. UCRL-MI-133583 Van Emden Henson, "An Algebraic Multigrid Tutorial," by Van Emden Henson. Presented at the Ninth Copper Mountain Conference on Multigrid Methods, Copper Mountain, CO, April 10, 1999. UCRL-MI-Jim E. Jones, "A Parallel Multigrid Tutorial," by Jim E. Jones. Presented at the Ninth Copper Mountain Conference on Multigrid Methods, Copper Mountain, CO, April 10, 1999. UCRL-MI-133748 • Helped in Profiling the AMG benchmark using the gprof to find the hotspots • Helped in trying to compile the CML • Started working in the hypreBoomerAMGRelax function Relax function has a huge switch case statement that selects the techniques to relax among different options (for e.g. Jacobi, Gauss-Siedel). There are 9 such options. The default option is 6 which uses the hybrid: SSOR-symmetric gauss siedel off-processor, SSOR on-processor with outer relaxation parameters • Identified the loops that needed to be ported to cell • Identified private and shared variables