Introduction to Parallel Programming

Linda Woodard
CAC
19 May 2010

Introduction to Parallel Computing on Ranger

5/18/2010 www.cac.cornell.edu 1
What is Parallel Programming?
Using more than one processor or computer to complete a task

– Each processor works on its section of the problem (functional parallelism)

– Each processor works on its section of the data (data parallelism)

– Processors can exchange information

Grid of Problem to be solved

CPU #1 works on this area CPU #2 works on this area

of the problem of the problem
y

CPU #3 works on this area CPU #4 works on this area

of the problem of the problem

5/18/2010 www.cac.cornell.edu 2
Forest Inventory with Ranger Bob

5/18/2010 www.cac.cornell.edu 3
Why Do Parallel Programming?
• Limits of single CPU computing
– performance
– available memory

• Parallel computing allows one to:

– solve problems that don’t fit on a single CPU
– solve problems that can’t be solved in a reasonable time

• We can solve…
– larger problems
– faster
– more cases

5/18/2010 www.cac.cornell.edu 4
Forest Inventory with Ranger Bob

5/18/2010 www.cac.cornell.edu 5
 Terminology (1)
• serial code is a single thread of execution working on a single data item at any
one time

• parallel code has more than one thing happening at a time. This could be
– A single thread of execution operating on multiple data items simultaneously
– Multiple threads of execution in a single executable
– Multiple executables all working on the same problem
– Any combination of the above

• task is the name we use for an instance of an executable. Each task has its own
virtual address space and may have multiple threads.
5/18/2010 www.cac.cornell.edu 6
 Terminology (2)

• node: a discrete unit of a computer system that typically runs its own instance
of the operating system

• core: a processing unit on a computer chip that is able to support a thread of

execution; can refer either to a single core or to all of the cores on a particular
chip

• cluster: a collection of machines or nodes that function in someway as a single

resource.

• grid: the software stack designed to handle the technical and social challenges
of sharing resources across networking and institutional boundaries. grid also
applies to the groups that have reached agreements to share their resources.

5/18/2010 www.cac.cornell.edu 7
Limits of Parallel Computing
• Theoretical Upper Limits
– Amdahl’s Law

• Practical Limits
– Load balancing (waiting)
– Conflicts (accesses to shared memory)
– Communications
– I/O (file system access)

5/18/2010 www.cac.cornell.edu 8
Theoretical Upper Limits to Performance
• All parallel programs contain:
– parallel sections (we hope!)
– serial sections (unfortunately)

• Serial sections limit the parallel effectiveness

serial portion parallel portion
1 task
2 tasks

4 tasks

• Amdahl’s Law states this formally

5/18/2010 www.cac.cornell.edu 9
Amdahl’s Law
• Amdahl’s Law places a strict limit on the speedup that can be
realized by using multiple processors.
– Effect of multiple processors on run time

t n = (f p / N + f s )t 1

– Where
• fs = serial fraction of code
• fp = parallel fraction of code
• N = number of processors
• t 1 = time to run on one processor

5/18/2010 www.cac.cornell.edu 10
Limit Cases of Amdahl’s Law
• Speed up formula:

S = 1 / (fs + fp / N)

Where
• fs = serial fraction of code
• fp = parallel fraction of code
• N = number of processors

Case:
1. fs = 0, fp = 1, then S = N
2. N  infinity: S = 1/fs; if 10% of the code is sequential, you will never
speed up by more than 10, no matter the number of processors.

5/18/2010 www.cac.cornell.edu 11
Ilustration of Amdahl’s Law
250

fp = 1.000
200
fp = 0.999
fp = 0.990
S 150 fp = 0.900

100

50

0
0 50 100 150 200 250
Number of processors

5/18/2010 www.cac.cornell.edu 12
More Terminology
• synchronization: the temporal coordination of parallel tasks. It involves
waiting until two or more tasks reach a specified point (a sync point) before
continuing any of the tasks.

• parallel overhead: the amount of time required to coordinate parallel

tasks, as opposed to doing useful work, including time to start and terminate
tasks, communication, move data.

• granularity: a measure of the ratio of the amount of computation done in a

parallel task to the amount of communication.
– fine-grained (very little computation per communication-byte)
– coarse-grained (extensive computation per communication-byte).

5/18/2010 www.cac.cornell.edu 13
 Practical Limits: Amdahl’s Law vs. Reality
• Amdahl’s Law shows a theoretical upper limit || speedup
• In reality, the situation is even worse than predicted by Amdahl’s Law due to:
• – Load balancing (waiting)
• – Scheduling (shared processors or memory)
• – Communications
80
• – I/O f = 0.99
p
70

60
50
Amdahl's Law
S 40 Reality
30
20
10
0
0 50 100 150 200 250
Number of processors

5/18/2010 www.cac.cornell.edu 14
Forest Inventory with Ranger Bob

5/18/2010 www.cac.cornell.edu 15
 Is it really worth it to go Parallel?
• Writing effective parallel applications is difficult!!
– Load balance is important
– Communication can limit parallel efficiency
– Serial time can dominate

• Is it worth your time to rewrite your application?

– Do the CPU requirements justify parallelization? Is your problem really `large’?
– Is there a library that does what you need (parallel FFT, linear system solving)
– Will the code be used more than once?

5/18/2010 www.cac.cornell.edu 16
Types of Parallel Computers (Flynn's taxonomy)
Data Stream
Single Multiple
Single Instruction Single Instruction
Single Single Data Multiple Data

SISD SIMD
Instruction Stream
Multiple Instruction Multiple Instruction
Single Data Multiple Data
Multiple
MISD MIMD

5/18/2010 www.cac.cornell.edu 17
Types of Parallel Computers (Memory Model)
• Nearly all parallel machines these days are multiple instruction, multiple
data (MIMD)

• A much more useful way to classify modern parallel computers is by their

memory model
– shared memory
– distributed memory

5/18/2010 www.cac.cornell.edu 18
Shared and Distributed Memory Models
P P P P P P
P P P P P P

M M M M M M
Bus
Memory
Network

Shared memory: single address space. All Distributed memory: each processor
processors have access to a pool of shared has its own local memory. Must do
memory; easy to build and program, good message passing to exchange data
price-performance for small numbers of between processors. cc-NUMA enables
processors; predictable performance due to larger number of processors and shared
UMA .(example: SGI Altix) memory address space than SMPs; still
easy to program, but harder and more
Methods of memory access : expensive to build. (example: Clusters)
- Bus
- Crossbar Methods of memory access :
- various topological interconnects
5/18/2010 www.cac.cornell.edu 19
Ranger

5/18/2010 www.cac.cornell.edu 20
Shared Memory vs. Distributed Memory
• Tools can be developed to make any system appear to look like a
different kind of system
– distributed memory systems can be programmed as if they have shared
memory, and vice versa
– such tools do not produce the most efficient code, but might enable
portability

• HOWEVER, the most natural way to program any machine is to use

tools and languages that express the algorithm explicitly for the
architecture.

5/18/2010 www.cac.cornell.edu 21
Programming Parallel Computers
• Programming single-processor systems is (relatively) easy because
they have a single thread of execution and a single address space.

• Programming shared memory systems can benefit from the single

address space

• Programming distributed memory systems is the most difficult due to

multiple address spaces and need to access remote data

• Both shared memory and distributed memory parallel computers can

be programmed in a data parallel, SIMD fashion and they also can
perform independent operations on different data (MIMD) and
implement task parallelism.
5/18/2010 www.cac.cornell.edu 22
 Data vs. Functional Parallelism
• Partition by Data (data parallelism)
Each process does the same work on a unique piece of data
– First divide the data. Each process then becomes responsible for whatever
work is needed to process its data.
– Data placement is an essential part of a data-parallel algorithm
– Usually more scalable than functional parallelism
– Can be programmed at a high level with OpenMP, or at a lower level
(subroutine calls) using a message-passing library like MPI.

• Partition by Task (functional parallelism)

Each process performs a different "function" or executes a different code section
– First identify functions, then look at the data requirements
– Commonly programmed with message-passing libraries

5/18/2010 www.cac.cornell.edu 23
 Data Parallel Programming Example
One code will run on 2 CPUs
Program has array of data to be operated on by 2 CPUs so array is split in two.
CPU A CPU B
program: program: program:
… … …
if CPU=a then low_limit=1 low_limit=51
low_limit=1
upper_limit=50 upper_limit=100
upper_limit=50
elseif CPU=b then do I= low_limit, do I= low_limit,
low_limit=51 upper_limit upper_limit
upper_limit=100 work on A(I) work on A(I)
end if end do end do
do I = low_limit, … …
upper_limit end program end program
work on A(I)
end do
...
end program

5/18/2010 www.cac.cornell.edu 24
Forest Inventory with Ranger Bob

5/18/2010 www.cac.cornell.edu 25
Single Program, Multiple Data (SPMD)
SPMD: dominant programming model for shared and distributed
memory machines.
– One source code is written
– Code can have conditional execution based on which processor is
executing the copy
– All copies of code are started simultaneously and communicate and
sync with each other periodically

5/18/2010 www.cac.cornell.edu 26
SPMD Programming Model

source.c

source.c source.c source.c source.c

Processor 0 Processor 1 Processor 2 Processor 3

5/18/2010 www.cac.cornell.edu 27
Task Parallel Programming Example
One code will run on 2 CPUs
Program has 2 tasks (a and b) to be done by 2 CPUs
CPU A CPU B
program.f:
… program.f: program.f:
initialize … …
...
if CPU=a then initialize initialize
do task a … …
elseif CPU=b then
do task b do task a do task b
end if … …
….
end program end program end program

5/18/2010 www.cac.cornell.edu 28
Forest Inventory with Ranger Bob

5/18/2010 www.cac.cornell.edu 29
Shared Memory Programming: OpenMP
• Shared memory systems (SMPs and cc-NUMAs) have a single
address space:
– applications can be developed in which loop iterations (with no
dependencies) are executed by different processors
– shared memory codes are mostly data parallel, ‘SIMD’ kinds of codes
– OpenMP is the new standard for shared memory programming
(compiler directives)
– Vendors offer native compiler directives

5/18/2010 www.cac.cornell.edu 30
Accessing Shared Variables
• If multiple processors want to write to a shared variable at the same time,
there could be conflicts :
– Process 1 and 2
– read X
– compute X+1
– write X
Shared variable X in
• Programmer, language, and/or memory
architecture must provide ways
of resolving conflicts
X + 1 in proc1 X + 1 in proc2

5/18/2010 www.cac.cornell.edu 31
OpenMP Example #1: Parallel Loop
!$OMP PARALLEL DO
do i=1,128
b(i) = a(i) + c(i)
end do
!$OMP END PARALLEL DO

The first directive specifies that the loop immediately following should be
executed in parallel. The second directive specifies the end of the parallel
section (optional).

For codes that spend the majority of their time executing the content of
simple loops, the PARALLEL DO directive can result in significant parallel
performance.
5/18/2010 www.cac.cornell.edu 32
OpenMP Example #2: Private Variables
!$OMP PARALLEL DO SHARED(A,B,C,N) PRIVATE(I,TEMP)
do I=1,N
TEMP = A(I)/B(I)
C(I) = TEMP + SQRT(TEMP)
end do
!$OMP END PARALLEL DO

In this loop, each processor needs its own private copy of the variable
TEMP. If TEMP were shared, the result would be unpredictable since
multiple processors would be writing to the same memory location.

5/18/2010 www.cac.cornell.edu 33
Distributed Memory Programming: MPI
Distributed memory systems have separate address spaces for
each processor

– Local memory accessed faster than remote memory

– Data must be manually decomposed

– MPI is the standard for distributed memory programming (library of

subprogram calls)

5/18/2010 www.cac.cornell.edu 34
Data Decomposition
For distributed memory systems, the ‘whole’ grid or sum of particles
is decomposed to the individual nodes
– Each node works on its section of the problem
– Nodes can exchange information

Grid of Problem to be solved

CPU #1 works on this area CPU #2 works on this area

of the problem of the problem
y

CPU #3 works on this area CPU #4 works on this area

of the problem of the problem

5/18/2010 www.cac.cornell.edu 35
MPI Example
#include <. . . .>
#include "mpi.h"
main(int argc, char **argv)
{
char message[20];
int i, rank, size, type = 99;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if (rank == 0) {
strcpy(message, "Hello, world");
for (i = 1; i < size; i++)
MPI_Send(message, 13, MPI_CHAR, i, type, MPI_COMM_WORLD);
}
else
MPI_Recv(message, 20, MPI_CHAR, 0, type, MPI_COMM_WORLD, &status);
printf( "Message from process = %d : %.13s\n", rank,message);
MPI_Finalize();
}

5/18/2010 www.cac.cornell.edu 36
MPI: Sends and Receives
MPI programs must send and receive data between the processors
(communication)

The most basic calls in MPI (besides the initialization, rank/size, and
finalization calls) are:
– MPI_Send
– MPI_Recv

These calls are blocking: the source processor issuing the

send/receive cannot move to the next statement until the target
processor issues the matching receive/send.

5/18/2010 www.cac.cornell.edu 37
Programming Multi-tiered Systems
• Systems with multiple shared memory nodes are becoming common
for reasons of economics and engineering.

• Memory is shared at the node level, distributed above that:

– Applications can be written using OpenMP + MPI
– Developing apps with only MPI usually possible

5/18/2010 www.cac.cornell.edu 38