8
BG/Q vs BG/P—Applications Perspective from Early Science Program Timothy J. Williams Argonne Leadership Computing Facility 2013 MiraCon Workshop Monday 3/4/2013 Session: 3:45-4:30pm

BG/Q vs BG/P—Applications Perspective from Early Science Program

  • Upload
    carys

  • View
    44

  • Download
    0

Embed Size (px)

DESCRIPTION

BG/Q vs BG/P—Applications Perspective from Early Science Program. Timothy J. Williams Argonne Leadership Computing Facility 2013 MiraCon Workshop Monday 3/4/2013 Session: 3:45-4:30pm. BG/P applications should run, unchanged, on BG/Q — faster. - PowerPoint PPT Presentation

Citation preview

Page 1: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

BG/Q vs BG/P—Applications Perspective from Early Science

Program

Timothy J. WilliamsArgonne Leadership Computing Facility

2013 MiraCon WorkshopMonday 3/4/2013Session: 3:45-4:30pm

Page 2: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

2

BG/P applications should run, unchanged, on BG/Q — faster

Page 3: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

3

16 projects– Large target allocations– Postdoc

Proposed runs between Mira acceptance and start of production

2 billion core-hours to burn in a few months

First in Mira Queue: Early Science Program

http://esp.alcf.anl.gov

Page 4: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

4

16 ESP ProjectsAlgorithms/MethodsStructured GridsUnstructured GridsFFTDense Linear AlgebraSparse Linear AlgebraParticles/N-BodyMonte Carlo

7 National Lab PIs

9 University PIs

Science AreasAstrophysicsBiologyCFD/AerodynamicsChemistryClimateCombustionCosmologyEnergyFusion PlasmaGeophysicsMaterialsNuclear Structure

Page 5: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

5

Next 2 slides, efforts characterized as S=small, M=medium, L=large– S : zero – few days of effort, modifications to 0% - 3% of existing lines of code– M : few weeks of effort, modifications to 3% - 10% of existing lines of code– S : few months of effort, modifications beyond 10% of existing lines of code

Ranking based on estimates by people who actually did the work

How Much Effort to “Port” to BG/Q?

Page 6: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

6

How Much Effort?PI/affiliation Code(s) Magnitude

of Changes Nature of Changes

Balaji/GFDL HIRAM L Improve OpenMP implementation, reformulate divergence-damping

Curtiss/ANL QMCPACK M S to port, L to use QPX in key kernels; plan: nested OpenMP

Frouzakis/ETH Nek5000 S Optimized small matrix-matrix multiply using QPX

Gordon/Iowa State GAMESS M 64-bit addressing, thread integral kernels with OpenMP

Habib/ANL, UC HACC M Short-range-force only: tree code

Harrison/ORNL MADNESS S Threading runtime tuning Kernel tuning to use QPX

Jansen/U Colorado PHASTA S Unchanged MPI-only performs well; OpenMP threaded in testing

Jordan/USC AWP-ODC, SORD S, M None, Threading

Page 7: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

7

How Much Effort?PI/affiliation Code(s) Magnitude

of Changes Nature of Changes

Khoklov/UC HSCD S Tune OpenMP parameters, link optimized math libs

Lamb/UC FLASH/RTFlame S OpenMP threading

Mackenzie/Fermilab

MILC, Chroma, CPS L

Full threading, QPX intrinsics/assembler, kernel on SPI comm.

Moser/UTexas PSDNS S Compile dependency libs, add OpenMP directives for threading

Pieper/ANL GFMC S Tune no. threads & ranks.

Roux/UC NAMD, Charm++ L Threads, PAMI implementation of Charm++

Tang/Princeton GTC S Improve OpenMP implementation

Voth/UC, ANL NAMD, LAMMPS, RAPTOR M OpenMP threads & serial

optimizations in RAPTOR/LAMMPS

Page 8: BG/Q  vs  BG/P—Applications Perspective from Early Science Program

8

Threads Communications

– One-sided– Beneath MPI

Kernel optimizations– QPX – Code restructuring

Parallel I/O Algorithms targeting Blue Gene architecture BG/Q Tuned libraries

– Linear algebra– Math functions– FFTs

Areas of Effort