parallel programming on the iucaa clusters sunu engineer

Parallel Programming On the IUCAA Clusters

Sunu Engineer

IUCAA Clusters

The Cluster – Cluster of Intel Machines on LinuxHercules – Cluster of HP ES45 quad processor

nodes

References: http://www.iucaa.ernet.in/

The Cluster

Four Single Processor Nodes with 100 Mbps Ethernet interconnect.

1.4 GHz, Intel Pentium 4 512 MB RAM Linux 2.4 Kernel (Redhat 7.2 Distribution) MPI – LAM 6.5.9 PVM – 3.4.3

Hercules

Four quad processor nodes with Memory Channel interconnect

1.25 GHz Alpha 21264D RISC Processor 4 GB RAM Tru64 5.1A with TruCluster software Native MPI LAM 7.0 PVM 3.4.3

Expected Computational Performance

Intel Cluster Processor - 512/590 System GFLOPS ~ 2 Algorithm/Benchmark

Used – Specint/float/HPL

ES45 Cluster Processor ~ 679/960 System GFLOPS ~ 30 Algorithm/Benchmark

Used – Specint/float/HPL

Parallel Programs

Move towards large scale distributed programs Larger class of problems with higher resolution Enhanced levels of details to be explored …

The Starting Point

Model Single Processor Program Multi Processor Program

Model Multiprocessor Program

Decomposition of a Single Processor Program

Temporal Initialization Control Termination

Spatial Functional Modular Object based

Multi Processor Programs

Spatial delocalization – Dissolving the boundary Single spatial coordinate - Invalid Single time coordinate - Invalid

Temporal multiplicity Multiple streams at different rates w.r.t an external

clock.

In comparison

Multiple points of initialization Distributed control Multiple points and times of termination Distribution of the activity in space and time

Breaking up a problem

Yet Another way

And another

Amdahl’s Law

Degrees of refinement

Fine parallelism Instruction level Program statement level Loop level

Coarse parallelism Process level Task level Region level

Patterns and Frameworks

Patterns - Documented solutions to recurring design problems.

Frameworks – Software and hardware structures implementing the infrastructure

Processes and Threads

From heavy multitasking to lightweight multitasking on a single processor

Isolated memory spaces to shared memory space

Posix Threads in Brief

pthread_create(pthread_t id, pthread_attr_t attributes, void *(*thread_function)(void *), void * arguments)

pthread_exit pthread_join pthread_self pthread_mutex_init pthread_mutex_lock/unlock Link with –lpthread

Multiprocessing architectures

Symmetric Multiprocessing Shared memory

Space Unified Different temporal streams

OpenMP standard

OpenMP Programming

Set of directives to the compiler to express shared memory parallelism

Small library of functions Environment variables. Standard language bindings defined for

FORTRAN, C and C++

Open MP example

#include <stdio.h>#include <omp.h> int main(int argc, char ** argv) {#pragma omp parallel { printf(“Hello World from

%d\n”,omp_get_thread_num());

}return(0);}

C An openMP program program openmp

!$OMP PARALLEL print *, “Hello world from”, omp_get_thread_num()

!$OMP END PARALLELstop

end

Open MP directivesParallel and Work sharing

OMP Parallel [clauses] OMP do [ clauses] OMP sections [ clauses] OMP section OMP single

Combined work sharingSynchronization

OMP parallel do OMP parallel sections OMP master OMP critical OMP barrierOMP atomicOMP flushOMP orderedOMP threadprivate

OpenMP Directive clauses

shared(list) private(list)/threadprivate firstprivate/lastprivate(list) default(private|shared|none) default(shared|none) reduction (operator|intrinsic : list) copyin(list) if (expr) schedule(type[,chunk]) ordered/nowait

Open MP Library functions

omp_get/set_num_threads() omp_get_max_threads() omp_get_thread_num() omp_get_num_procs() omp_in_parallel() omp_get/set_(dynamic/nested)() omp_init/destroy/test_lock() omp_set/unset_lock()

OpenMP environment variables

OMP_SCHEDULE OMP_NUM_THREADS OMP_DYNAMIC OMP_NESTED

OpenMP Reduction and Atomic Operators

Reduction : +,-,*,&,|,&&,|| Atomic : ++,--,+,*,-,/,&,>>,<<,|

Simple loops

do I=1,N z(I) = a * x(I) + y end do

!$OMP parallel do do I=1,N z(I) = a * x(I) + y end do

Data Scoping

Loop index private by default Declare as shared, private or reduction

Private variables

!$OMP parallel do private(a,b,c) do I=1,m

do j =1,n b=f(I) c=k(j) call abc(a,b,c) end do end do#pragma omp parallel for private(a,b,c)

Dependencies

Data dependencies (Lexical/dynamic extent) Flow dependencies Classifying and removing the dependencies Non removable dependenciesExamples

Do I=2,na(I) =a(I)+a(I-1)

end doDo I=2,N,2 a(I)= a(I)+a(I-1)End do

Making sure everyone has enough work

Parallel overhead – Creation of threads, synchronization vs. work done in the loop

$!OMP parallel do schedule(dynamic,3) schedule type – static, dynamic, guided,runtime

Parallel regions – from fine to coarse parallelism

$!OMP Parallel threadprivate and copyin Work sharing constructs

do, sections, section, singleSynchronization critical, atomic, barrier, ordered, master

To distributed memory systems

MPI, PVM, BSP …

Existing parallel libraries and toolkits include: PUL, the Parallel Utilities Library from EPCC. The Multicomputer Toolbox from Tony Skjellum and

colleagues at LLNL and MSU. The Portable, Extensible, Toolkit for Scientific

computation from ANL. ScaLAPACK from ORNL and UTK. ESSL, PESSL on AIX PBLAS, PLAPACK, ARPACK

Some Parallel Libraries