Advanced Hybrid MPI/OpenMP Parallelization Paradigms for Nested Loop Algorithms onto Clusters of SMPs Nikolaos Drosinos and Nectarios Koziris National.

Advanced Hybrid MPI/OpenMP Advanced Hybrid MPI/OpenMP Parallelization Paradigms for Parallelization Paradigms for Nested Loop Algorithms onto Nested Loop Algorithms onto

Clusters of SMPsClusters of SMPs

Nikolaos Drosinos and Nectarios Koziris

National Technical University of Athens

Computing Systems Laboratory

{ndros,nkoziris}@cslab.ece.ntua.grwww.cslab.ece.ntua.gr

2/10/2003 EuroPVM/MPI 2003 2

OverviewOverview

Introduction Pure MPI Model Hybrid MPI-OpenMP Models

• Hyperplane Scheduling• Fine-grain Model• Coarse-grain Model

Experimental Results Conclusions – Future Work

2/10/2003 EuroPVM/MPI 2003 3

IntroductionIntroduction

Motivation:

• SMP clusters• Hybrid programming models

Mostly fine-grain MPI-OpenMP paradigms Mostly DOALL parallelization

2/10/2003 EuroPVM/MPI 2003 4


Contribution:

• 3 programming models for the parallelization of nested loops algorithms

• pure MPI• fine-grain hybrid MPI-OpenMP• coarse-grain hybrid MPI-OpenMP

• Advanced hyperplane scheduling• minimize synchronization need• overlap computation with communication

2/10/2003 EuroPVM/MPI 2003 5


Algorithmic Model:

FOR j0 = min0 TO max0 DO

…

FOR jn-1 = minn-1 TO maxn-1 DO

Computation(j0,…,jn-1);

ENDFOR

…

ENDFOR

Perfectly nested loops Constant flow data dependencies

2/10/2003 EuroPVM/MPI 2003 6


Target Architecture: SMP clusters

2/10/2003 EuroPVM/MPI 2003 7

OverviewOverview




2/10/2003 EuroPVM/MPI 2003 8

Pure MPI ModelPure MPI Model

Tiling transformation groups iterations into atomic execution units (tiles) Pipelined execution Overlapping computation with communication Makes no distinction between inter-node and intra-node communication

2/10/2003 EuroPVM/MPI 2003 9


Example:

FOR j1=0 TO 9 DO FOR j2=0 TO 7 DO A[j1,j2]:=A[j1-1,j2] + A[j1,j2-1]; ENDFORENDFOR

10

01D

2/10/2003 EuroPVM/MPI 2003 10


j1

j2

CPU1

CPU0

CPU1

CPU0

NODE1

NODE0

4 MPI nodes

2/10/2003 EuroPVM/MPI 2003 11


j1

j2

CPU1

CPU0

CPU1

CPU0

NODE1

NODE0

4 MPI nodes

2/10/2003 EuroPVM/MPI 2003 12


tile0 = nod0;…tilen-2 = nodn-2;FOR tilen-1 = 0 TO DO

Pack(snd_buf, tilen-1 – 1, nod); MPI_Isend(snd_buf, dest(nod)); MPI_Irecv(recv_buf, src(nod)); Compute(tile); MPI_Waitall; Unpack(recv_buf, tilen-1 + 1, nod);END FOR

1

11 minmax

n

nn

x

2/10/2003 EuroPVM/MPI 2003 13

OverviewOverview




2/10/2003 EuroPVM/MPI 2003 14

Hyperplane SchedulingHyperplane Scheduling

Implements coarse-grain parallelism assuming inter-tile data dependencies Tiles are organized into data-independent subsets (groups) Tiles of the same group can be concurrently executed by multiple threads Barrier synchronization between threads

2/10/2003 EuroPVM/MPI 2003 15


j1

j2

CPU1

CPU0

CPU1

CPU0

NODE1

NODE0

2 MPI nodes

x

2 OpenMP threads

2/10/2003 EuroPVM/MPI 2003 16

j1

j2


CPU1

CPU0

CPU1

CPU0

NODE1

NODE0

2 MPI nodes

x

2 OpenMP threads

2/10/2003 EuroPVM/MPI 2003 17

Hyperplane SchedulingHyperplane Scheduling#pragma omp parallel{ group0 = nod0; … groupn-2 = nodn-2; tile0 = nod0 * m0 + th0; … tilen-2 = nodn-2 * mn-2 + thn-2; FOR(groupn-1){ tilen-1 = groupn-1 - ;

if(0 <= tilen-1 <= ) compute(tile); #pragma omp barrier }}

tnn 11 minmax

2

0

n

iitile

2/10/2003 EuroPVM/MPI 2003 18

OverviewOverview




2/10/2003 EuroPVM/MPI 2003 19

Fine-grain ModelFine-grain Model

Incremental parallelization of computationally intensive parts Relatively straightforward from pure MPI Threads (re)spawned at computation Inter-node communication outside of multi-threaded part Thread synchronization through implicit barrier of omp parallel directive

2/10/2003 EuroPVM/MPI 2003 20

Fine-grain ModelFine-grain Model

FOR(groupn-1){ Pack(snd_buf, tilen-1 – 1, nod); MPI_Isend(snd_buf, dest(nod)); MPI_Irecv(recv_buf, src(nod)); #pragma omp parallel { thread_id=omp_get_thread_num(); if(valid(tile,thread_id,groupn-1)) Compute(tile); } MPI_Waitall; Unpack(recv_buf, tilen-1 + 1, nod);}

2/10/2003 EuroPVM/MPI 2003 21

OverviewOverview




2/10/2003 EuroPVM/MPI 2003 22

Coarse-grain ModelCoarse-grain Model

SPMD paradigm Requires more programming effort Threads are only spawned once Inter-node communication inside multi-threaded part (requires MPI_THREAD_MULTIPLE) Thread synchronization through explicit barrier (omp barrier directive)

2/10/2003 EuroPVM/MPI 2003 23

Coarse-grain ModelCoarse-grain Model#pragma omp parallel{ thread_id=omp_get_thread_num(); FOR(groupn-1){ #pragma omp master{ Pack(snd_buf, tilen-1 – 1, nod); MPI_Isend(snd_buf, dest(nod)); MPI_Irecv(recv_buf, src(nod)); } if(valid(tile,thread_id,groupn-1)) Compute(tile); #pragma omp master{ MPI_Waitall; Unpack(recv_buf, tilen-1 + 1, nod); } #pragma omp barrier }}

2/10/2003 EuroPVM/MPI 2003 24

Summary: Fine-grain vs Summary: Fine-grain vs Coarse-grainCoarse-grain

Fine-grain Coarse-grainThreads re-spawning Threads are only

spawned onceInter-node MPI communication outside of multi-threaded region

Inter-node MPI communication inside multi-threaded region, assumed by master thread

Intra-node synchronization through implicit barrier (omp parallel)

Intra-node synchronization through explicit OpenMP barrier

2/10/2003 EuroPVM/MPI 2003 25

OverviewOverview

Introduction Pure MPI model Hybrid MPI-OpenMP models



2/10/2003 EuroPVM/MPI 2003 26

Experimental ResultsExperimental Results

8-node SMP Linux Cluster (800 MHz PIII, 128 MB RAM, kernel 2.4.20) MPICH v.1.2.5 (--with-device=ch_p4, --with-comm=shared) Intel C++ compiler 7.0 (-O3 -mcpu=pentiumpro -static) FastEthernet interconnection ADI micro-kernel benchmark (3D)

2/10/2003 EuroPVM/MPI 2003 27

Alternating Direction Implicit Alternating Direction Implicit (ADI)(ADI)

Unitary data dependencies 3D Iteration Space (X x Y x Z)

X

Y

Z

2/10/2003 EuroPVM/MPI 2003 28

ADI – 4 nodesADI – 4 nodes

Pure MPI Hybrid

I: MPII: OpenMP

XX

Y Y

2/10/2003 EuroPVM/MPI 2003 29


X < Y

X > Y

Pure MPI HybridX

Y

X

Y

Pure MPI HybridXX

Y Y

2/10/2003 EuroPVM/MPI 2003 30

ADI X=512 Y=512 Z=8192 – 4 ADI X=512 Y=512 Z=8192 – 4 nodesnodes

2/10/2003 EuroPVM/MPI 2003 31


2/10/2003 EuroPVM/MPI 2003 32


2/10/2003 EuroPVM/MPI 2003 33


Pure MPI Hybrid

I: MPII: OpenMP

XX

Y Y

2/10/2003 EuroPVM/MPI 2003 34


X < Y

X > Y

Pure MPI HybridXX

Y Y

Pure MPI HybridXX

Y Y

2/10/2003 EuroPVM/MPI 2003 35


2/10/2003 EuroPVM/MPI 2003 36


2/10/2003 EuroPVM/MPI 2003 37


2/10/2003 EuroPVM/MPI 2003 38


2/10/2003 EuroPVM/MPI 2003 39


2/10/2003 EuroPVM/MPI 2003 40


Computation Communication

2/10/2003 EuroPVM/MPI 2003 41


Computation Communication

2/10/2003 EuroPVM/MPI 2003 42

OverviewOverview

Introduction Pure MPI model Hybrid MPI-OpenMP models



2/10/2003 EuroPVM/MPI 2003 43

ConclusionsConclusions

Nested loop algorithms with arbitrary data dependencies can be adapted to the hybrid parallel programming paradigm Hybrid models can be competitive to the pure MPI paradigm Coarse-grain hybrid model can be more efficient than fine-grain one, but also more complicated Programming efficiently in OpenMP not easier than programming efficiently in MPI

2/10/2003 EuroPVM/MPI 2003 44

Future WorkFuture Work

Application of methodology to real applications and benchmarks Work balancing for coarse-grain model Performance evaluation on advanced interconnection networks (SCI, Myrinet) Generalization as compiler technique

2/10/2003 EuroPVM/MPI 2003 45

Questions?Questions?

http://www.cslab.ece.ntua.gr/~ndros

Advanced Hybrid MPI/OpenMP Parallelization Paradigms for Nested Loop Algorithms onto Clusters of SMPs Nikolaos Drosinos and Nectarios Koziris National.

Documents

pure mpi model tile

tile n

pure mpi model example

destnod mpi

endfor slide

mpi nodes x

forgroup n

srcnod computetile mpi