/
The Charm++ Programming Model and NAMD The Charm++ Programming Model and NAMD

The Charm++ Programming Model and NAMD - PowerPoint Presentation

mitsue-stanley
mitsue-stanley . @mitsue-stanley
Follow
384 views
Uploaded On 2017-03-29

The Charm++ Programming Model and NAMD - PPT Presentation

Abhinav S Bhatele Department of Computer Science University of Illinois at UrbanaChampaign httpcharmcsuiucedu Email bhateleillinoisedu Feb 13th 2009 Abhinav S Bhatele 2 University of Illinois ID: 530887

abhinav bhatele 2009 13th bhatele abhinav 13th 2009 feb pme marenostrum namd illinois communication charm processors computation mpi vps bonded problem number

Share:

Link:

Embed:

Download Presentation from below link

Download Presentation The PPT/PDF document "The Charm++ Programming Model and NAMD" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.


Presentation Transcript

Slide1

The Charm++ Programming Model and NAMD

Abhinav S BhateleDepartment of Computer ScienceUniversity of Illinois at Urbana-Champaign

http://charm.cs.uiuc.edu

E-mail: bhatele@illinois.eduSlide2

Feb 13th, 2009

Abhinav S Bhatele

2

University of Illinois

Urbana-Champaign,

Illinois, USASlide3

Processor Virtualization

User View

System View

Programmer: Decomposes the computation into objects

Runtime: Maps the computation on to the processors

Feb 13th, 2009

3

Abhinav S BhateleSlide4

Benefits of the Charm++ Model

Software EngineeringNumber of VPs independent of physical processorsDifferent sets of VPs for different computationsDynamic MappingLoad balancing

Change the set of processors usedMessage-driven executionCompositionalityPredictability

Feb 13th, 2009

4

Abhinav S BhateleSlide5

Charm++ and CSE Applications

Feb 13th, 2009

5

Abhinav S Bhatele

Gordon Bell Award, SC 2002

Fine-grained CPAIMD

Cosmological SimulationsSlide6

Adaptive MPI (AMPI)

Feb 13th, 2009Abhinav S Bhatele

6

MPI “Processes” - Implemented as virtual “processes” (light-weight user level

migratable

threads)Slide7

Load Balancing

Based on Principle of PersistenceRuntime instrumentationMeasures communication volume and computation time

Measurement based load balancersUse the instrumented data-base periodically to make new decisionsMany alternative strategies can use the database

Centralized

vs

distributed

Greedy improvements

vs

complete reassignments

Taking communication into account

Feb 13th, 2009

Abhinav S Bhatele

7Slide8

NAMD

Feb 13th, 2009Abhinav S Bhatele

8

NAnoscale

Molecular Dynamics

Simulates the life of bio-molecules

Simulation window broken

down into a large number of time steps (typically 1

fs

each)

Forces

on each atom calculated every step

Positions and velocities updated and atoms migrated to their new positionsSlide9

Parallelization of NAMD

Hybrid DecompositionFeb 13th, 2009

Abhinav S Bhatele

9Slide10

Control Flow

Feb 13th, 2009Abhinav S Bhatele

10Slide11

Feb 13th, 2009

Abhinav S Bhatele

11

Non-bonded Work

Bonded Work

Integration

PME

CommunicationSlide12

Optimizations on MareNostrum

Fixes to the CVS versionCrash due to a xlc compiler bugPerformance analysisUsing Projections performance analysis tool

Performance tuningConfiguration parameters for NAMD to enable high parallelization

Feb 13th, 2009

Abhinav S Bhatele

12Slide13

Performance Tuning

Increasing the number of objectstwoAway{X, Y, Z} optionsChoice of PME implementationSlab (1D) vs. Pencil (2D) decompositionOffloading processors doing PME

Use of spanning tree for communicationFeb 13th, 2009

Abhinav S Bhatele

13Slide14

DD on MareNostrum

Feb 13th, 2009

14

Abhinav S BhateleSlide15

DD with Slab PME

Feb 13th, 2009Abhinav S Bhatele

15Slide16

DD with Pencil PME

Feb 13th, 2009Abhinav S Bhatele

16Slide17

Problem?

Feb 13th, 2009Abhinav S Bhatele

17

1.2 msSlide18

Feb 13th, 2009

Abhinav S Bhatele

18Slide19

Nucleosome on MareNostrum

Feb 13th, 2009

19

Abhinav S BhateleSlide20

Membrane on MareNostrum

Feb 13th, 2009

20

Abhinav S BhateleSlide21

Similar Problem?

Feb 13th, 2009Abhinav S Bhatele

21

16 ms

Bandwidth on

Myrinet

: 250 MB/s

Expected Latency for 6-10 KB messages < 0.1 msSlide22

Optimizations on the Xe Cluster

Getting the CVS version running (over mpi)Trying a development version of charmRuns directly over

Infiniband (ibverbs)Currently unstable

Feb 13th, 2009

Abhinav S Bhatele

22Slide23

DD on Xeon Cluster

Feb 13th, 2009

23

Abhinav S BhateleSlide24

Thank You!