PMSCS 657_Parallel and Distributed processing

1
Parallel
and Distributed processing
PMSCS 657

2
Syllabus: Part 1
• Introduction; motivation; types of parallelism;
data decomposition; uses of parallel computers;
classification of machines.
• SPMD programs; memory models; shared and
distributed memory; OpenMP.
• Example of summing numbers; interconnection
networks; network metrics; Gray code mappings.
• Classification of parallel algorithms; Speedup and
efficiency.
• Scalable algorithms; Amdahl's law; sending and
receiving messages; programming with MPI.

3
Syllabus: Part2
• Collective communication; integration example;
regular computations and a simple example.
• Regular two-dimensional problems and an
example.
• Dynamic communication and the molecular
dynamics example.
• Irregular computations; the WaTor simulation.
Load balancing strategies.
• Message passing libraries; introduction to PVM.
• Review lectures.

4
Books
• “Parallel Programming,” B. Wilkinson and M. Allen, published
by Prentice Hall, 1999. ISBN 0-13-671710-1.
• “Parallel Computing: Theory and Practice,” M. Quinn,
published by McGraw-Hill, 1994.
• “Solving Problems on Concurrent Processors, Volume 1,” Fox,
Johnson, Lyzenga, Otto, Salmon, and Walker, published by
Prentice-Hall, 1988.
• “Using MPI,” Gropp, Lusk, and Skjellum, published by MIT
Press, 1994.
• “Parallel Programming with MPI,” Peter Pacheco, published
by Morgan Kaufmann, 1996.

5
What is Parallelism?
• Parallelism refers to the simultaneous
occurrence of events on a computer.
• An event typically means one of the
following:
– An arithmetical operation
– A logical operation
– Accessing memory
– Performing input or output (I/O)

6
Types of Parallelism 1
• Parallelism can be examined at several
levels.
– Job level: several independent jobs
simultaneously run on the same computer
system.
– Program level: several tasks are performed
simultaneously to solve a single common
problem.

7
Types of Parallelism 2
– Instruction level: the processing of an instruction, such
as adding two numbers, can be divided into sub-
instructions. If several similar instructions are to be
performed their sub-instructions may be overlapped
using a technique called pipelining.
– Bit level: when the bits in a word are handled one after
the other this is called a bit-serial operation. If the bits
are acted on in parallel the operation is bit-parallel.
In this parallel processing course we shall be mostly
concerned with parallelism at the program level.
Concurrent processing is the same as parallel
processing.

8
Scheduling Example
• Average utilisation is 83.3%
• Time to complete all jobs is 7 time units.
Time Jobs running Utilisation
1 S, M 75%
2 L 100%
3 S, S, M 100%
4 L 100%
5 L 100%
6 S, M 75%
7 M 50%

9
A Better Schedule
• A better schedule would allow jobs to be taken out of order
to give higher utilisation.
S M L S S M L L S M M
• Allow jobs to “float” to the front to the queue to maintain
high utilisation.
Time Jobs running Utilisation
1 S, M, S 100%
2 L 100%
3 S, M, S 100%
4 L 100%
5 L 100%
6 M, M 100%

10
Notes on Scheduling Example
• In the last example:
– Average utilisation is 100%.
– Time to complete all jobs is 6 time units.
• Actual situation is more complex as jobs
may run for differing lengths of time.
• Real job scheduler must balance high
utilisation with fairness (otherwise large
jobs may never run).

11
Parallelism Between Job Phases
• Parallelism also arises when different independent jobs
running on a machine have several phases, e.g.,
computation, writing to a graphics buffer, I/O to disk or
tape, and system calls.
• Suppose a job is executing and needs to perform I/O
before it can progress further. I/O is usually expensive
compared with computation, so the job currently running
is suspended, and another is started. The original job
resumes after the I/O operation has completed.
• This requires special hardware: I/O channels or special I/O
processor.
• The operating system controls how different jobs are
scheduled and share resources.

12
Program Level Parallelism
This is parallelism between different parts of the
same job.
Example
A robot has been programmed to look for electrical
sockets when it runs low on power. When it finds
one it goes over to it and plugs itself in to recharge.
Three subsystems are involved in this - the vision,
manipulation, and motion subsystems. Each
subsystem is controlled by a different processor, and
they act in parallel as the robot does different things.

13
Robot Example
Task Vision Manipulation Motion
1. Looking for
electrical socket
 
2. Going to
electrical socket
 
3. Plugging into
electrical socket
 

14
Notes on Robot Example
• The subsystems are fairly independent, with
the vision subsystem guiding the others.
• There may also be a central “brain”
processor.
• This is an example of task parallelism in
which different tasks are performed
concurrently to achieve a common goal.

15
Domain Decomposition
A common form of program-level parallelism arises
from the division of the data to be programmed
into subsets.
• This division is called domain decomposition.
• Parallelism that arises through domain
decomposition is called data parallelism.
• The data subsets are assigned to different
computational processes. This is called data
distribution.
• Processes may be assigned to hardware processors
by the program or by the runtime system. There
may be more than one process on each processor.

16
Data Parallelism
• Consider an image
digitized as a square array
of pixels which we want
to process by replacing
each pixel value by the
average of its neighbours.
• The domain of the
problem is the two-
dimensional pixel array.
               
               
               
               
               
               
               
               
               
               
               
               
               
               
               
               

17
Domain Decomposition
• Suppose we decompose
the problem into 16
subdomains
• We then distribute the
data by assigning each
subdomain to a process.
• The pixel array is a
regular domain because
the geometry is simple.
• This is a homogeneous
problem because each
pixel requires the same
amount of computation
(almost - which pixels
are different?).
               
               
               
               
               
               
               
               
               
               
               
               
               
               
               
               

18
Why Use Parallelism?
• Better utilisation of resources. Want to keep
hardware busy.
• Want to run programs faster by spreading work
over several processors.
• Ways to measure speed:
– Floating point operations per second. 1 Mflop/s is one
million floating point operations per second.
• High performance workstation  10-20 Gflop/s
• Current best supercomputer  1 Pflop/s
– Transactions per second. 1 Tps is one transaction per
second.
– Instructions per second. 1 Mips is one million
instructions per second.

19
Parallelism and Memory
• Want more memory to solve bigger or more complex
problems.
• Typical workstations have 1 Gbytes of RAM,
expandable to 32 Gbytes. Can fit an 65,536  65,536
array into 32 Gbytes of memory.
• The IBM “Roadrunner” parallel computer at Lawrence
Livermore National Lab has 122,400 cores with a total
of 98 Tbytes of RAM. Can fit a 3,670,000  3,670,000
array into memory. See
http://www.top500.org/system/9485.

21
Parallelism and Supercomputing
• Parallelism is exploited on a variety of high
performance computers, in particular massively
parallel computers (MPPs) and clusters.
• MPPs, clusters, and high-performance vector
computers are termed supercomputers.
• Currently supercomputers have peak performance
in the range of 100-1000 Tflop/s, and memory of
10 to 100 Tbytes. They cost about 20-50 million
pounds.
• Supercomputers are leading to a new methodology
in science called computational science joining
theoretical and experimental approaches.

22
Uses of Parallel Supercomputers
• Weather forecasting. Currently forecasts are usually accurate up to
about 5 days. This should be extended to 8 to 10 days over the next
few years. Researchers would like to better model local nonlinear
phenomena such as thunderstorms and tornadoes.
• Climate modelling. Studies of long-range behaviour of global
climate. This is relevant to investigating global warming.
• Engineering. Simulation of car crashes to aid in design of cars.
Design of aircraft in “numerical wind tunnels.‘”
• Material science. Understanding high temperature superconductors.
Simulation of semiconductor devices. Design of lightweight, strong
materials for construction.
• Drug design. Prediction of effectiveness of drug by simulation. Need
to know configuration and properties of large molecules.

23
More Uses of Parallelism
• Plasma physics. Investigation of plasma fusion
devices such as tokamaks as future source of
cheap energy.
• Economics. Economic projections used to guide
decision-making. Prediction of stock market
behavior.
• Defense. Tracking of multiple missiles. Event-
driven battlefield simulations. Code cracking.
• Astrophysics. Modeling internal structure of stars.
Simulating supernova. Modeling the structure of
the universe.

24
Classification of Parallel
Machines
• To classify parallel machines we must first
develop a model of computation. The approach we
follow is due to Flynn (1966).
• Any computer, whether sequential or parallel,
operates by executing instructions on data.
– a stream of instructions (the algorithm) tells the
computer what to do.
– a stream of data (the input) is affected by these
instructions.

25
Classification of Parallel
Machines
• Depending on whether there is one or several of
these streams we have 4 classes of computers.
– Single Instruction Stream, Single Data Stream: SISD
– Multiple Instruction Stream, Single Data Stream: MISD
– Single Instruction Stream, Multiple Data Stream: SIMD
– Multiple Instruction Stream, Multiple Data Stream: MIMD

26
SISD Computers
This is the standard sequential computer.
A single processing unit receives a single stream of
instructions that operate on a single stream of data
Control Processor Memory
Instruction
stream stream
Data
Example:
To compute the sum of N numbers a1,a2,…,aN the processor needs to
gain access to memory N consecutive times. Also N-1 additions are
executed in sequence. Therefore the computation takes O(N)
operations
Algorithms for SISD computers do not contain any process parallelism
since there is only one processor.

27
MISD Computers
N processors, each with its own control unit,
share a common memory.
Data
stream
Memory
Control 1
Control 2
Control N
Processor 1
Processor 2
Processor N
Instruction
streams
1
2
N

28
MISD Computers (continued)
• There are N streams of instructions
(algorithms/programs) and one stream of data.
Parallelism is achieved by letting the processors
do different things at the same time to the same
data.
• MISD machines are useful in computations where
the same input is to be subjected to several
different operations.

29
MISD Example
• Checking whether a number Z is prime. A simple
solution is to try all possible divisions of Z.
Assume the number of processors is N=Z-2. All
processors take Z as input and each tries to divide
it by its associated divisor. So it is possible in one
step to check if Z is prime. More realistically, if
N<Z-2 then a subset of divisors is assigned to each
processor.
• For most applications MISD computers are very
awkward to use and no commercial machines exist
with this design.

30
SIMD Computers
• All N identical processors operate under the
control of a single instruction stream issued by a
central control unit.
• There are N data streams, one per processor, so
different data can be used in each processor.
Processor 2 Processor N
Control
Shared memory or interconnection network
Processor 1
Data
streams
1 2 N
Instruction stream

31
Notes on SIMD Computers
• The processors operate synchronously and a global
clock is used to ensure lockstep operation, i.e., at
each step (global clock tick) all processors execute
the same instruction, each on a different datum.
• Array processors such as the ICL DAP,
Connection Machine CM-200, and MasPar are
SIMD computers.
• SIMD machines are particularly useful at
exploiting data parallelism to solve problems
having a regular structure in which the same
instructions are applied to subsets of data.

32
SIMD Example
The same instruction is issued to all 4 processors
(add two numbers), and all processors execute the
instructions simultaneously. It takes one step to add
the matrices, compared with 4 steps on a SISD
machine.
a11 a12
a21 a22
b11 b12
b21 b22
c11 c12
c21 c22
+ =
Problem: add two 22 matrices on 4 processors.

33
Notes on SIMD Example
• In this example the instruction is simple, but in
general it could be more complex such as merging
two lists of numbers.
• The data may be simple (one number) or complex
(several numbers).
• Sometimes it may be necessary to have only a
subset of the processors execute an instruction, i.e.,
only some data needs to be operated on for that
instruction. This information can be encoded in the
instruction itself indicating whether
– the processor is active (execute the instruction)
– the processor is inactive (wait for the next instruction)

34
MIMD Computers
This is the most general and most powerful of our
classification. We have N processors, N streams of
instructions, and N streams of data.
Processor 2 Processor N
Control 1
Shared memory or interconnection network
Processor 1
Data
streams
1 2 N
Instruction
streams
Control 2 Control N
1 2 N

35
Notes on MIMD Computers
• The processors can operate asynchronously,
i.e., they can do different things on different
data at the same time.
• As with SIMD computers, communication
of data or results between processors can be
via shared memory or an interconnection
network.

36
Notes on SIMD and MIMD
• In most problems to be solved on SIMD and MIMD
computers it is useful for the processors to be able to
communicate with each other to exchange data or results.
This can be done in two ways
– by using a shared memory and shared variables, or
– using an interconnection network and message passing
(distributed memory)
• MIMD computers with shared memory are known as
multiprocessors. An example is the Onyx 300 produced by
Silicon Graphics Inc.
• MIMD computers with an interconnection network are
known as multicomputers. An example is the E6500
produced by Sun Microsystems.
• Clusters are multicomputers composed of off-the-shelf
components

37
Potential of the 4 Classes
SISD
SIMD
+
A B
+
A B
C D
A+B
A+B
C+D
MIMD
MISD
+
*
A B
+
*
A B
C D
A+B
A+B
A*B
C*D

38
Single Program Multiple Data
• An MIMD computer is said to be running in SPMD mode if
the same program is executing on each process.
• SPMD is not a hardware paradigm, so it is not included in our
4 classifications.
• It is a software paradigm for MIMD machines.
• Each processor executes an SPMD program on different data
so it is possible that different branches are taken, leading to
asynchronous parallelism. The processors no longer do the
same thing (or nothing) in lockstep as they do on an SIMD
machine. They execute different instructions within the same
program.

39
SPMD Example
• Suppose X is 0 on processor 1, and 1 on processor
2. Consider
• Then processor 1 executes S1 at the same time that
processor 2 executes S2.
• This could not happen on an SIMD machine.
IF X = 0
THEN S1
ELSE S2

40
Interprocessor Communication
• Usually a parallel program needs to have some
means of sharing data and results processed by
different processors. There are two main ways of
doing this
1. Shared Memory
2. Message passing
• Shared memory consists of a global address
space. All processors can read from and write
into this global address space.

41
Global Shared Memory
GLOBAL
MEMORY
PROCESSOR
PROCESSOR
PROCESSOR
PROCESSOR
PROCESSOR
PROCESSOR

42
Shared Memory Conflicts
The shared memory approach is simple but can lead
to problems when processors simultaneously access
the same location in memory.
Example:
Suppose the shared memory initially holds a variable
x with value 0. Processor 1 adds 1 to x and processor
2 adds 2 to x. What is the final value of x?
You should have met this problem before when
studying locks and critical sections in the operating
systems module.

43
Shared Memory Conflicts 2
The following outcomes are possible
1. If P1 executes and completes x=x+1 before P2 reads the value of x
from memory then x is 3. Similarly, if P2 executes and completes
x=x+2 before P1 reads the value of x from memory then x is 3.
2. If P1 or P2 reads x from memory before the other has written back
its result, then the final value of x depends on which finishes last.
– if P1 finishes last the value of x is 1
– if P2 finishes last the value of x is 2
Shared memory
X = 0
Processor 1
X = X + 1
Processor 2
X = X + 2
This is an example
of non-determinism
or non-determinancy

44
Non-Determinancy
• Non-determinancy is caused by race conditions.
• A race condition occurs when two statements in
concurrent tasks access the same memory location,
at least one of which is a write, and there is no
guaranteed execution ordering between accesses.
• The problem of non-determinancy can be solved
by synchronising the use of shared data. That is if
x=x+1 and x=x+2 were mutually exclusive then
the final value of x would always be 3.
• Portions of a parallel program that require
synchronisation to avoid non-determinancy are
called critical sections.

45
Locks and Mutual Exclusion
In shared memory programs locks can be used
to give mutually exclusive access.
Processor 1:
LOCK (X)
X = X + 1
UNLOCK (X)
Processor 2:
LOCK (X)
X = X + 2
UNLOCK (X)

46
Classifying Shared Memory
Computers
Shared memory computers can be classified as follows
depending on whether two or more processors can gain access
to the same memory simultaneously.
1. Exclusive Read, Exclusive Write (EREW)
• Access to memory locations is exclusive, i.e., no 2 processors are
allowed to simultaneously read from or write into the same location.
2. Concurrent Read, Exclusive Write (CREW)
• Multiple processors are allowed to read from the same location, but
write is still exclusive, i.e., no 2 processors are allowed to write into
the same location simultaneously.
3. Exclusive Read, Concurrent Write (ERCW)
• Multiple processors are allowed to write into the same location, but
read access remains exclusive.
4. Concurrent Read, Concurrent Write (CRCW)
• Both multiple read and write privileges are allowed.

47
Notes on Shared Memory 1
• Allowing concurrent read access to the same
address should pose no problems in the sense
that such an operation is well-defined.
Conceptually each processor makes a copy of the
contents of the memory location and stores it in
its own register.
• Problems arise with concurrent write access,
because if several processors write
simultaneously to the same address, which
should “succeed?”

48
• There are several ways of deterministically
specifying the contents of a memory location after
a concurrent write
1. Assign priorities to processors and store value from
processor with highest priority.
2. All the processors are allowed to write, provided all the
values they are attempting to store are the same.
3. The max, min, sum, or average of the values is stored
(for numeric data).

49
• SIMD machines usually have 1000's of very
simple processors. Shared memory SIMD
machines are unrealistic because of the cost and
difficulty in arranging for efficient access to
shared memory for so many processors. There are
no commercial shared memory SIMD machines.
• MIMD machines use more powerful processors
and shared memory machines exist for small
numbers of processors (up to about 100).

50
Examples of Shared Memory
To show how the 4 subclasses of shared
memory machines behave, consider the
following example.
Problem:
We have N processors to search a list S = {L1,
L2, …,Lm} for the index of a given element x.
Assume x may appear several times, and any
index will do. 1< N≤m.

51
The Algorithm
procedure SM_search (S, x, k)
STEP 1: for i=1 to N do in parallel
read x
end for
Si = {L((i-1)m/N+1), ...,L(im/N)}
perform sequential search on sublist Si
(return Ki= -1 if not in list, otherwise index)
end for
if Ki > 0 then k=Ki end if
end for
end procedure

52
Time Complexity for EREW
If the sequential search step takes O(m/N) time, what
is the time complexity for each of the 4 subclasses of
shared memory computer?
•EREW
Step 1 takes O(N) (N reads, one at a time).
Step 2 takes O(m/N) time.
Step 3 takes O(N) time.
Total time is O(N)+O(m/N).

53
Time Complexity for ERCW
• ERCW
Step 1 takes O(N) time.
Step 3 takes constant time.

54
Time Complexity for CREW
• CREW
Step 3 takes O(N) time

55
Time Complexity for CRCW
• CRCW
Total time is O(m/N).

56
Limits on Shared Memory
• Shared memory computers are often implemented by
incorporating a fast bus to connect processors to memory.
• However, because the bus has a finite bandwidth, i.e., it can
carry only a certain maximum amount of data at any one
time, then as the number of processors increase the
contention for the bus becomes a problem. So it is feasible to
build shared memory machines with up to only about 100
processors.
Shared Memory
Bus
Processor Processor Processor

57
Quick Overview of OpenMP
• OpenMP can be used to represent task and data
parallelism.
• In case of data parallelism, OpenMP is used to
split loop iterations over multiple threads.
• Threads can execute different code but share the
same address space.
• OpenMP is most often used on machines with
support for a global address space.

59
OpenMP and Loops
#include <omp.h>
#define CHUNKSIZE 100
#define N 1000
main () {
int i, chunk;
float a[N], b[N], c[N];
/* Some initializations */
for (i=0; i < N; i++) a[i] = b[i] = i * 1.0;
chunk = CHUNKSIZE;
#pragma omp parallel shared(a,b,c,chunk) private(i)
{
#pragma omp for schedule(dynamic,chunk) nowait
for (i=0; i < N; i++) c[i] = a[i] + b[i];
}
/* end of parallel section */ }

60
Number of Threads
#include <omp.h>
#define CHUNKSIZE 100
#define N 1000
main () {
int i, chunk;
float a[N], b[N], c[N];
for (i=0; i < N; i++) a[i] = b[i] = i * 1.0;
chunk = CHUNKSIZE;
#pragma omp parallel shared(a,b,c,chunk) private(i) num_threads(4)
{
#pragma omp for schedule(dynamic,chunk) nowait
for (i=0; i < N; i++) c[i] = a[i] + b[i];
}
/* end of parallel section */ }

61
Reduction Operations
#include <omp.h>
main () {
int i, n, chunk;
float a[100], b[100], result;
n = 100;
chunk = 10;
result = 0.0;
for (i=0; i < n; i++) {
a[i] = i * 1.0;
b[i] = i * 2.0;
}
#pragma omp parallel for default(shared) private(i) schedule(static,chunk) reduction(+:result)
for (i=0; i < n; i++)
result = result + (a[i] * b[i]);
printf("Final result= %fn",result);
}

62
Interconnection Networks and
Message Passing
In this case each processor has its own private (local)
memory and there is no global, shared memory. The
processors need to be connected in some way to allow
them to communicate data.
Interconnectio
n
Network
Processor
+ memory
Processor
+ memory
Processor
+ memory
Processor
+ memory
Processor
+ memory

63
Message Passing
• If a processor requires data contained on a
different processor then it must be explicitly
passed by using communication
instructions, e.g., send and receive.
• The value x is explicitly passed from P2 to P1.
This is known as message passing.
P1 P2
receive (x, P2) send (x, P1)

64
Hybrid Computers
• In addition to the cases of
shared memory and
distributed memory there
are possibilities for hybrid
designs that incorporate
features of both.
• Clusters of processors are
connected via a high speed
bus for communication
within a cluster, and
communicate between
clusters via an
interconnection network.
Bus
Interconnection
network

65
Comparison of Shared and
Distributed Memory
Distributed memory Shared memory
Large number of processors
(100’s to 1000’s)
Moderate number of
processor (10’s to 100)
High peak performance Modest peak performance
Unlimited expansion Limited expansion
Difficult to fully utilise
Relatively easy to fully
utilise
Revolutionary parallel
computing
Evolutionary parallel
computing

66
Memory Hierarchy 1
• In general, certain memory locations have a greater affinity for
certain processes.
• Parallel programming language may make distinction between
“near” and “far” memory, but will not fully represent the
memory hierarchy.
P0 P1
Registers
Cache(s)
Main memory
Faster
access
More
memory
Data sharing between processes

67
Typical Quad-Core Chip
L1 cache
Instruction cache
CPU
L1 cache
Instruction cache
CPU
L2 cache
L1 cache
Instruction cache
CPU
L1 cache
Instruction cache
CPU
L2 cache
Shared Memory

68
Summing m Numbers
Example: summing m numbers
On a sequential computer we have,
sum = a[0];
for (i=1;i<m;i++) {
sum = sum + a[i];
}
Would expect the running time be roughly proportional
to m. We say that the running time is (m).

69
Summing m Numbers in Parallel
• What if we have N processors, with each
calculating the m/N numbers assigned to it?
• We must add these partial sums together to
get the total sum.

70
Summing Using Shared Memory
The m numbers, and the global sum, are held in global shared
memory.
global_sum = 0;
for (each processor){
local_sum = 0;
calculate local sum of m/N numbers
LOCK
global_sum = global_sum + local_sum;
UNLOCK
}

71
Notes on Shared Memory
Algorithm
• Since global_sum is a shared variable each
processor must have mutually exclusive access to
it – otherwise the final answer may be incorrect.
• The running time (or algorithm time complexity) is
 (m/N)+  (N)
• where
– m/N comes from finding the local sums in parallel
– N comes from adding N numbers in sequence

72
Summing Using Distributed
Memory
P1
1
P1
3
P1
2
P2
1
P2
3
P2
2
P3
1
P3
3
P3
2
j
i
Suppose we have a square mesh of N processors.
The algorithm is as follows:
1. Each processor finds the local
sum of its m/N numbers
2. Each processor passes its local
sum to another processor in a
coordinated way
3. The global sum is finally in
processor P11.

73
Distributed Memory Algorithm
The algorithm proceeds as follows:
1. Each processor finds its local sum.
2. Sum along rows:
a) If the processor is in the rightmost column it sends its local sum to the left.
b) If the processor is not in the rightmost or leftmost column it receives the
number form the processor on its right, adds it to its local, and send the
result to the processor to the left.
c) If the processor is in the leftmost column it receives the number from the
processor on its right and adds it to it local sum to give the row sum.
3. Leftmost column only – sum up the leftmost column:
a) If the processor is in the last row send the row sum to the processor above
b) If the processor is not in the last or first row receive the number from the
processor below, add it to the row sum, and send result to processor above
c) If the processor is in the first row receive the number from the processor
below. This is the global sum.

74
Summing Example
There are N-1 additions and N-1 communications in
each direction, so the total time complexity is
(m/N) +  (N) + C
where C is the time spent communicating.
10 12 7
6 9 17
9 11 18
9910 19
6 26
9 29
29
70
29
32
38
Initially Shift left
and sum
Shift left
and sum
Shift up
and sum
Shift up
and sum

75
Interconnection Networks
• Parallel computers with many processors do not
use shared memory hardware.
• Instead each processor has its own local memory
and data communication takes place via message
passing over an interconnection network.
• The characteristics of the interconnection network
are important in determining the performance of a
multicomputer.
• If network is too slow for an application,
processors may have to wait for data to arrive.

76
Examples of Networks
Important networks include:
• fully connected or all-to-all
• mesh
• ring
• hypercube
• shuffle-exchange
• butterfly
• cube-connected cycles

77
Network Metrics
A number of metrics can be used to evaluate
and compare interconnection networks.
• Network connectivity is the minimum number of
nodes or links that must fail to partition the network
into two or more disjoint networks.
• Network connectivity measures the resiliency of a
network, and its ability to continue operation
despite disabled components. When components
fail we would like the network to continue
operation with reduced capacity.

78
Network Metrics 2
• Bisection width is the minimum number of links
that must be cut to partition the network into two
equal halves (to within one). The bisection
bandwidth is the bisection width multiplied by the
data transfer rate of each link.
• Network diameter is the maximum internode
distance, i.e., the maximum number of links that
must be traversed to send a message to any node
along the shortest path. The lower the network
diameter the shorter the time to send messages to
distant nodes.

79
Network Metrics 3
Network narrowness measures congestion in
a network.
• Partition the network into two
groups A and B, containing
NA and NB nodes,
respectively, with NB < NA.
• Let I be the number on
connections between nodes in
A and nodes in B. The
narrowness of the network is
the maximum value of NB/I for
all partitionings of the network.
Group A
Group B
I connections
• If the narrowness is high
(NB>I) then if the group B
nodes want to communicate
with the group A nodes
congestion in the network
will be high.

80
Network Metrics 4
• Network Expansion Increment is the minimum
number of nodes by which the network can be
expanded.
– A network should be expandable to create larger and
more powerful parallel systems by simply adding more
nodes to the network.
– For reasons of cost it is better to have the option of
small increments since this allows you to upgrade your
machine to the required size.
• Number of edges per node. If this is independent
of the size of the network then it is easier to
expand the system.

81
Fully Connected Network
• In the fully connected, or all-to-all, network each
node is connected directly to all other nodes.
• This is the most general and powerful
interconnection network, but it can be
implemented for only a small number of nodes.

82
Fully Connected Network 2
For n even:
• Network connectivity = n - 1
• Network diameter = 1
• Network narrowness = 2/n
• Bisection width = n2
/4
• Expansion Increment = 1
• Edges per node = n - 1

83
Mesh Networks
• In a mesh network nodes are arranged as a q-
dimensional lattice, and communication is allowed
only between neighboring nodes.
• In a periodic mesh, nodes on the edge of the mesh
have wrap-around connections to nodes on the other
side. This is sometimes called a toroidal mesh.

84
Mesh Metrics
For a q-dimensional non-periodic lattice with
kq
nodes:
• Network connectivity = q
• Network diameter = q(k-1)
• Network narrowness = k/2
• Bisection width = kq-1
• Expansion Increment = kq-1
• Edges per node = 2q

85
Ring Networks
A simple ring network is just a 1D periodic mesh.
• Network connectivity = 2
• Network diameter = n/2
• Network narrowness = n/4
• Bisection width = 2
• Expansion Increment = 1
• Edges per node = 2
The problem for a simple ring is its large diameter.

86
Chordal Ring Networks
• A chordal ring uses extra chordal links to reduce
the diameter.
• For a ring with extra diametric links we have (for
n even)
– Network connectivity = 3
– Network diameter = ceiling(n/4)
– Network narrowness = n/(n+4)
– Bisection width = 2+n/2
– Expansion Increment = 2
– Edges per node = 3

87
Examples of Ring Networks
• Here are a simple ring and a chordal ring
with diametric links, each of size 6 nodes.

88
Hypercube Networks
• A hypercube network consists of n=2k
nodes
arranged as a k-dimensional hypercube.
Sometimes called a binary n-cube.
• Nodes are numbered 0, 1,…,n-1, and two nodes
are connected if their node numbers differ in
exactly one bit.
– Network connectivity = k
– Network diameter = k
– Network narrowness = 1
– Bisection width = 2k-1
– Expansion increment = 2k
– Edges per node = k

89
Examples of Hypercubes
1D 2D 3D
4D

90
Mapping Grids to Hypercubes
• In the example in which we summed a set of numbers
over a square mesh of processors each processor needs
to know where it is in the mesh.
• We need to be able to map node numbers to locations
in the process mesh
– Given node number k what is its location (i,j) in the
processor mesh?
– Given a location (i,j) in the processor mesh what is the node
number, k, of the processor at that location?
– We want to choose a mapping such that neighbouring
processes in the mesh are also neighbours in the hypercube.
This ensures that when neighbouring processes in the mesh
communicate, this entails communication between
neighbouring processes in the hypercube.

91
Binary Gray Codes
• Consider just one dimension – a periodic
processor mesh in this case is just a ring.
• Let G(i) be the node number of the processor at
position i in the ring, where 0  i < n. The mapping
G must satisfy the following,
– It must be unique, i.e., G(i) = G(j)  i = j.
– G(i) and G(i-1) must differ in exactly one bit for all i, 0
 i < n-1.
– G(n-1) and G(0) must differ in exactly one bit.

92
Binary Gray Codes 2
• A class of mappings known as binary Gray
codes satisfy these requirements. There are
several n-bit Gray codes. Binary Gray codes
can be defined recursively as follows:
Given a d-bit Gray code, a (d+1)-bit Gray code
can be constructed by listing the d-bit Gray
code with the prefix 0, followed by the d-bit
Gray code in reverse order with prefix 1.

93
Example of a Gray Code
• Start with the Gray code G(0)=0, G(1)=1.
• Then the 2-bit Gray code is given in Table 1,
and the 3-bit Gray code is given in Table 2.
i [G(i)]2 G(i)
0 00 0
1 01 1
2 11 3
3 10 2
Table 1: A 2-bit Gray code

94
Example of a Gray Code 2
i [G(i)]2 G(i)
0 000 0
1 001 1
2 011 3
3 010 2
4 110 6
5 111 7
6 101 5
7 100 4
Table 2: A 3-bit Gray code

95
Example of a Gray Code 3
• A ring can be embedded in a hypercube as
follows:
0
1
2
3
4
5
6
7

96
Multi-Dimensional Gray Codes
• To map a multidimensional mesh of
processors to a hypercube we require that
the number of processors in each direction
of the mesh be a power of 2. So
2dr-1 2dr-2  …  2d0
is an r-dimensional mesh and if d is the
hypercube dimension then:
d0+ d1+… + dr-1= d

97
Multi-Dimensional Gray Codes 2
• We partition the bits of node number and
assign them to each dimension of the mesh.
The first d0 go to dimension 0, the next d1
bits go to dimension 1, and so on. Then we
apply separate inverse Gray code mappings
to each group of bits.

98
Mapping a 24 Mesh to a
Hypercube
K [k1]2,[k0]2 [G-1
(k1)]2,[G-1
(k0)]2 (i,j)
0 0, 00 0, 00 (0,0)
1 0, 01 0, 01 (0,1)
2 0, 10 0, 11 (0,3)
3 0, 11 0, 10 (0,2)
4 1, 00 1, 00 (1,0)
5 1, 01 1, 01 (1,1)
6 1, 10 1, 11 (1,3)
7 1, 11 1, 10 (1,2)

99
Mapping a 24 Mesh to a
Hypercube 2
• A 2  4 mesh is embedded into a 3D
hypercube as follows:
0 1 3 2
4 5 7 6

100
Shuffle-Exchange Networks
• A shuffle-exchange network consists of
n=2k
nodes, and two kinds of connections.
– Exchange connections links nodes whose
numbers differ in their lowest bit.
– Perfect shuffle connections link node i with
node 2i mod(n-1), except for node n-1 which is
connected to itself.

101
8-Node Shuffle-Exchange
Network
• Below is an 8-node shuffle-exchange
network, in which shuffle links are shown
with solid lines, and exchange links with
dashed lines.
0 2 3 4 5 6 71

102
• What is the origin of the name “shuffle-
exchange”?
• Consider a deck of 8 cards numbered 0, 1, 2,…,7.
The deck is divided into two halves and shuffled
perfectly, giving the order:
0, 4, 1, 5, 2, 6, 3, 7
• The final position of a card i can be found by
following the shuffle link of node i in a shuffle-
exchange network.

103
• Let ak-1, ak-2,…, a1, a0 be the address of a node in a
shuffle-exchange network in binary.
• A datum at this node will be at node number
ak-2,…, a1, a0, ak-1
after a shuffle operation.
• This corresponds to a cyclic leftward shift in the
binary address.
• After k shuffle operations we get back to the node
we started with, and nodes through which we pass
are called a necklace.

104
Butterfly Network
• A butterfly network consists of (k+1)2k
nodes
divided into k+1 rows, or ranks.
• Let node (i,j) refer to the jth node in the ith rank.
Then for i > 0 node (i,j) is connected to 2 nodes in
rank i-1, node (i-1,j) and node (i-1,m), where m is
the integer found by inverting the ith most
significant bit of j.
• Note that if node (i,j) is connected to node (i-1,m),
then node (i,m) is connected to node (i-1,j). This
forms a butterfly pattern.
– Network diameter = 2k
– Bisection width = 2k

105
Example of a Butterfly Network
Here is a butterfly network for k = 3.
Rank 0
Rank 1
Rank 2
Rank 3
(0,2) (0,6)
(1,0) (1,2) (1,6)
(2,0) (2,2) (2,3)
(3,2) (3,3)
i = 1, j = 2 = (010)2, j
= (110)2 = 6
i = 2, j = 2 = (010)2, j
= (000)2 = 0
i = 3, j = 2 = (010)2, j
= (011)2 = 3

106
Cube-Connected Cycles Network
• A cube-connected cycles network is a k-
dimensional hypercube whose 2k
vertices are
actually cycles of k nodes.
• The advantage compared with the hypercube is
that the number of edges per node is a constant, 3.
• Disadvantages are that network diameter is twice
that of a hypercube, and the bisection width is
lower.
• For a cube-connected cycle network of size k2k
,
– Network diameter = 2k
– Bisection width = 2k-1
– Edges per node = 3

107
Example of a Cube-Connected
Cycles Network
A cube-connected cycles network with k = 3
looks like this:

108
Complete Binary Tree Network
• Tree-based networks use switches to
connect processors. An example is the
binary tree network.
•This has a bisection width of 1, and a
connectivity of 1. The low bisection width
can result in congestion in the upper levels of
the network.

109
Fat Tree Network
• The fat tree network seeks to reduce the
congestion in the upper levels of the network by
adding extra links.
•The connectivity is still 1, but if there are 2d
processing nodes the bisection width is 2d-1
.
•This type of network was used in the CM-5.

110
Classifying Parallel Algorithms
• Parallel algorithms for MIMD machines can
be divided into 3 categories
– Pipelined algorithms
– Data parallel, or partitioned, algorithms
– Asynchronous, or relaxed, algorithms

111
Pipelined Algorithms
• A pipelined algorithm involves an ordered set of
processes in which the output from one process is
the input for the next.
• The input for the first process is the input for the
algorithm.
• The output from the last process is the output of
the algorithm.
• Data flows through the pipeline, being operated on
by each process in turn.

112
Pipelines Algorithms 2
• Example: Suppose it takes 3 steps, A, B, and C, to
assemble a widget, and each step takes one unit of
time.
• In the sequential case it takes 3 time units to
assemble each widget.
• Thus it takes 3n time units to produce n widgets.
A B C W2 W1

113
Example of Pipelined Algorithm
• In the pipelined case the following happens
– Time step 1: A operates on W1
– Time step 2: A operates on W2, B operates on W1
– Time step 3: A operates on W3, B operates on W2, C
completes W1
– Time step 4: A operates on W4, B operates on W3, C
completes W2
• After 3 time units, a new widget is produced every
time step.

114
Pipelined Algorithm
• If the pipeline is n processes long, a new widget is
produced every time step from the nth time step
onwards. We then say the pipeline is full.
• The pipeline start-up time is n-1.
• This sort of parallelism is sometimes called
algorithmic parallelism.
A B C W5W4W3W2W1
A B C

115
Performance of Pipelining
If
• N is the number of steps to be performed
• T is the time for each step
• M is the number of items (widgets)
then
Sequential time = NTM
Pipelined time = (N+M-1)T

116
Pipeline Performance Example
If T = 1, N = 100, and M = 106
, then
• Sequential time = 108
• Pipelined time = 1000099
The speed-up Tseq/Tpipe  100.

117
Data Parallelism
• Often there is a natural way of decomposing the
data into smaller parts, which are then allocated to
different processors.
• This way of exploiting parallelism is called data
parallelism or geometric parallelism.
• In general the processors can do different things to
their data, but often they do the same thing.
• Processors combine the solutions to their sub-
problems to form the complete solution. This may
involve communication between processors.

118
Data Parallelism Example
Data parallelism can be exploited in the
widget example. For 3-way data parallelism
we have:
A B C W4 W1
A B C W5 W2
A B C W6 W3

119
Relaxed Parallelism
• Relaxed parallelism arises when there is no
explicit dependency between processes.
• Relaxed algorithms never wait for input –
they use the most recently available data.

120
Relaxed Parallelism Example
Suppose we have 2 processors, A and B.
• A produces a sequence of numbers, ai,
i=1,2,….
• B inputs ai and performs some calculation
on it to produce Fi.
• Say B runs much faster than A.

121
Synchronous Operation
• A produces a1, passes it to B, which
calculates F1
• A produces a2, passes it to B, which
calculates F2
• and so on…..

122
Asynchronous Operation
1. A produces a1, passes it to B, which calculates F1
2. A is in the process of computing a2, but B does
not wait – it uses a1 to calculate F2, i.e., F1=F2.
• Asynchronous algorithms keep processors busy.
Drawbacks of asynchronous algorithms are
– they are difficult to analyse
– an algorithm that is known to converge in
synchronous mode
– may not converge in asynchronous mode.

123
Example of Asynchronous
Algorithm
• The Newton-Raphson method is an iterative
algorithm for solving non-linear equations
f(x)=0.
xn+1= xn- f(xn) / f '(xn)
• generates a sequence of approximations to
the root, starting with some initial value x0.

124
Example of Asynchronous
Algorithm 2
Suppose we have 3 processors
• P1: given x, P1 calculates f(x) in t1, and sends it to P3.
• P2: given y, P2 calculates f '(x) in t2, and sends it to P3.
• P3: given a, b, and c, P3 calculates d = a - b/c.
If |d - a| >  then d is sent to P1 and P2; otherwise it is
output.
xn
f(xn)
f 
(xn)
f 
(xn)
f(xn)
xn -
P1
P2
P3

125
Serial Mode Time Complexity
Serial mode
• P1 computes f(xn), then P2 computes f 
(xn),
then P3 computes xn+1.
• Serial time is t1+ t2+t3 per iteration.
• If k iterations are needed, total time is
k(t1+ t2+t3)

126
Synchronous Parallel Mode
• P1 and P2 compute f(xn) and f 
(xn)
simultaneously, and when both have
finished the values of f(xn) and f 
(xn) are
used by P3 to find xn+1.
• Time per iteration is max(t1,t2) + t3.
• k iterations are necessary so the total time
is, k(max(t1,t2) + t3).

127
Asynchronous Parallel Mode
• P1 and P2 begin computing as soon as they
receive a new input value from P3.
• P3 computes a new value as soon as it
receives a new input value from either P1
or P2.

128
Asynchronous Parallel Mode
Example
• For example, if
t1=2, t2=3 and
t3=1.
• Ci indicates
processor is
using xi in its
calculation.
• Cannot predict
number of
iterations.
Time P1 P2 P3
1 C0 C0 –
2 f(x0) C0 –
3 – f
(x0)
4 – – x1= x0- f(x0)/ f
(x0)
5 C1 C1 –
6 f(x1) C1 –
7 – f
(x1) x2= x1- f(x1)/ f
(x0)
8 C2 C2 x3= x2- f(x1)/ f
(x1)
9 f(x2) C2 –
10 C3 f
(x2) x4= x3- f(x2)/ f
(x1)
11 f(x3) C4 x5= x4- f(x2)/ f
(x2)
12 C5 C4 x6= x5- f(x3)/ f
(x2)

129
Speed-up and Efficiency
• We now define some metrics which measure how
effectively an algorithm exploits parallelism.
• Speed-up is the ratio of the time taken to run the
best sequential algorithm on one processor of the
parallel machine divided by the time to run on N
processors of the parallel machine.
S(N) = Tseq/Tpar(N)
• Efficiency is the speed-up per processor.
 (N) = S(N)/N=(1/N)(Tseq/Tpar(N))
• Overhead is defined as
f(N) = 1/ (N) - 1

130
Example
• Suppose the best known sequential
algorithm takes 8 seconds, and a parallel
algorithm takes 2 seconds on 5 processors.
Then
Speed-up = 8/2 = 4
Efficiency = 4/5 = 0.8
Overhead = 1/0.8 – 1 = 0.25

131
Self Speed-up and Linear Speed-up
• Self speed-up is defined using the parallel
algorithm running on one processor.
• If the speed-up using N processors is N then
the algorithm is said to exhibit linear speed-
up.

132
Factors That Limit Speed-up
1. Software Overhead
Even when the sequential and parallel
algorithms perform the same computations,
software overhead may be present in the
parallel algorithm. This includes additional
index calculations necessitated by how the
data were decomposed and assigned to
processors, and other sorts of “bookkeeping”
required by the parallel algorithm but not the
sequential algorithm.

133
2. Load Imbalance
Each processor should be assigned the same
amount of work to do between
synchronisation points. Otherwise some
processors may be idle while waiting for
others to catch up. This is known as load
imbalance. The speedup is limited by the
slowest processor.

134
3. Communication Overhead
Assuming that communication and calculation
cannot be overlapped, then any time spent
communicating data between processors
reduces the speed-up.

135
Grain Size
The grain size or granularity is the amount of
work done between communication phases of
an algorithm. We want the grain size to be
large so the relative impact of communication
is less.

136
Definition of Load Imbalance
• Suppose the work done by processor i between two
successive synchronisation points is Wi
• If the number of processors is N, then the average
workload is:
∑
−
=






=
1
0
1 N
i
iW
N
W
• The amount of load imbalance is then given by:





 −
=
W
WW
L i
max
where the maximum is taken over all processors.

137
Analysis of Summing Example
P1
1
P1
3
P1
2
P2
1
P2
3
P2
2
P3
1
P3
3
P3
2
j
i
Recall the example of summing m numbers on a square
mesh of N processors.
The algorithm proceeds as follows
1. Each processor finds the local
sum of its m/N numbers
2. Each processor passes its local
sum to another processor in a
coordinated way
3. The global sum is finally in
processor P11.

138
Analysis of Summing Example 2
• Time for best sequential algorithm is
Tseq = (m-1)tcalc
where tcalc is time to perform one floating-point
operation.
• Time for each phase of parallel algorithm
– Form local sums T1 = (m/N-1) tcalc
– Sum along processor rows T2 = (N - 1)(tcalc + tcomm)
– where tcomm is time to communicate one floating-point number
between neighbouring processors.
– Sum up first column of processors T3 = (N - 1)(tcalc + tcomm)

139
• Total time for the parallel algorithm is:
Tpar= (m/N + 2 N - 3)tcalc+ 2(N - 1) tcomm
• So the speed-up for the summing example is:
(m/N + 2 N - 3)tcalc + 2(N - 1) tcomm
(m-1)tcalc
S(N) =
1 + (N/m)(2 N - 3)+ 2(N/m)(N - 1) 
N(1-1/m)
=
where  = tcomm/ tcalc

140
• In this algorithm a good measure of the grain size,
g, is the number of elements per processor, m/N.
We can write S as:
1 + (N/m)(2 N - 3)+ 2(N/m)(N - 1) 
N(1-1/m)
S(g,N) =
• As g   with N constant, S  N.
• As N   with g constant, S  gN/(2(1+)).
• As N   with m constant, S  0.

141
• If m  1 and N  1,
1 + 2 N(1+)/g
N
S(g,N) =
1 + 2 N(1+)/g
1
(g,N) =
2 N(1+)/gf(g,N) =

142
Scalable Algorithms
• Scalability is a measure of how effectively an algorithm
makes use of additional processors.
• An algorithm is said to be scalable if it is possible to keep
the efficiency constant by increasing the problem size as
the number of processors increases.
• An algorithm is said to be perfectly scalable if the
efficiency remains constant when the problem size and the
number of processors increase by the same factor.
• An algorithm is said to be highly scalable if the efficiency
depends only weakly on the number of processors when
the problem size and the number of processors increase by
the same factor.

143
Scalability of the Summing
Example
• The summing algorithm is scalable since we
can take g  N.
• The summing algorithm is not perfectly
scalable, but it is highly scalable.
• “Problem size” may be either:
– the work performed, or
– the size of the data.

144
Amdahl’s Law
• Amdahl's Law states that the maximum speedup of
an algorithm is limited by the relative number of
operations that must be performed sequentially,
i.e., by its serial fraction.
• If  is the serial fraction, n is the number of
operations in the sequential algorithm, and N the
number of processors, then the time for the parallel
algorithm is:
Tpar(N) = (n + (1-)n/N)t + C(n,N)
where C(n,N) is the time for overhead due to
communication, load balancing, etc., and t is the
time for one operation.

145
Derivation of Amdahl’s Law
• The speed-up satisfies:
S(N) = Tseq/Tpar(N) = nt/[(n + (1-)n/N)t + C(n,N)]
= 1/[( + (1-)/N) + C(n,N)/(nt)]
< 1/( + (1-)/N)
• Note that as N, then S(N)1/, so the
speed-up is always limited to a maximum of
1/ no matter how many processors are
used.

146
Examples of Amdahl’s Law

0
S(N)
10
0 0.01 0.03 0.1
S(N) < 1/( + (1-)/N)
0.06
5.26
6.49
9.17
7.87
Consider the effect of Amdahl's Law on speed-up as a
function of serial fraction, , for N=10 processors.

147
Examples of Amdahl’s Law 2
Consider the effect of Amdahl's Law on speed-up as a
function of serial fraction, , for N=1000 processors.

0
S(N)
1000
0 0.01 0.03 0.1
S(N) < 1/( + (1-)/N)
0.06
9.9116.4190.99 32.29
If 1% of a parallel
program involves serial
code, the maximum
speed-up is 9 on a 10-
processor machine, but
only 91 on a 1000-
processor machine.

148
Implications of Amdahl’s Law
• Amdahl's Law says that the serial fraction puts a
severe constraint on the speed-up that can be
achieved as the number of processors increases.
• Amdahl's Law suggests that it is not cost effective
to build systems with large numbers of processors
because sufficient speed-up will not be achieved.
• It turns out that most important applications that
need to be parallelised contain very small serial
fractions, so large machines are justified.

149
Speed-Up for Large Problems
• Speed-up is the ratio between how long the best
sequential algorithm takes on a single processor
and how long it takes to run on multiple
processors.
• To measure the speed-up the problem must be
small enough to fit into the memory of one
processor.
• This limits us to measuring the speed-up of only
small problems.

150
Speed-Up for Large Problems 2
• In finding the speedup we can estimate the time to
run on one processor, so much larger problems
can be considered.
• In general overhead costs increase with problem
size, but at a slower rate than the amount of
computational work (measured by the grain size).
Thus, speed-up is an increasing function of
problem size, and so this approach to speed-up
allows us to measure larger speed-ups.

151
Speed-Up and Problem Size
0
0
Ideal speed-up
Slope = 1
Increasing M
S(N)
Number of processors, N
For a given number of processors, speed-up usually increases
with problem size, M.

152
Semantics of Message Sends
• Suppose one node sends a message to
another node:
send (data, count, datatype, destination)
• There are two possible behaviours:
– Blocking send
– Non-blocking send

153
Semantics of Blocking Send
• The send does not return until the data to be
sent has “left” the application.
• This usually means that the message has
been copied by the message passing system,
or it has been delivered to the destination
process.
• On return from the send() routine the data
buffer can be reused without corrupting the
message.

154
Semantics of Non-Blocking Send
• Upon return from the send() routine the data
buffer is volatile.
• This means that the data to be sent is not
guaranteed to have left the application, and if the
data buffer is changed the message may be
corrupted. The idea here is for the send() routine
to return as quickly as possible so the sending
process can get on with other useful work.
• A subsequent call is used to check for completion
of the send.

155
Semantics of Message Receives
• Suppose one node receives a message from
another node:
receive (data, count, datatype, source)
• There are two possible behaviours:
– Blocking receive
– Non-blocking receive

156
Semantics of Blocking Receive
• The receive does not return until the data to
be received has “entered” the application.
• This means that the message has been
copied into the data buffer and can be used
by the application on the receiving
processor.

157
Semantics of Non-Blocking
Receive
• Upon return from the receive() routine the status of
the data buffer is undetermined.
• This means that it is not guaranteed that the
message has yet been received into the data buffer.
• We say that a receive has been posted for the
message.
• The idea here is for the receive() routine to return
as quickly as possible so the receiving process can
get on with other useful work. A subsequent call is
used to check for completion of the receive.

158
Message Passing Protocols
• Suppose one node sends a message and another
receives it:
SOURCE: send (data, count, datatype, destination)
DEST: receive (data, count, datatype, source)
• Two important message passing protocols are
– Synchronous send protocol
– Asynchronous send protocol

159
Message Passing Protocols 2
• Synchronous: The send and receive routines
overlap in time. The send does not return
until the receive has started. This is also
known as a rendezvous protocol.
• Asynchronous: The send and receive
routines do not necessarily overlap in time.
The send can return regardless of whether
the receive has been initiated.

160
MPI Point-to-Point Communication
• MPI is a widely-used standard for message
passing on distributed memory concurrent
computers.
• Communication between pairs of processes is
called point-to-point communication.
• There are several Java versions of MPI, but we
shall use mpiJava.
• In mpiJava point-to-point communication is
provided through the methods of the Comm class.

161
mpiJava API
• The class MPI only has static members.
• It acts as a module containing global services, such as
initialisation, and many global constants including the
default communicator COMM_WORLD.
• The most important class in the package is the
communicator class Comm.
• All communication functions in mpiJava are
members of Comm or its subclasses.
• Another very important class is the Datatype class.
• The Datatype class describes the type of elements
in the message buffers to send and receive.

162
Class hierarchy
MPI
Group
Comm
Datatype
Status
Request
Package mpi
Intracomm
Intercomm
Prequest
Cartcomm
Graphcomm

163
Basic Datatypes
MPI Datatype Java Datatype
MPI.BYTE byte
MPI.CHAR char
MPI.SHORT short
MPI.BOOLEAN boolean
MPI.INT int
MPI.LONG long
MPI.FLOAT float
MPI.DOUBLE double
MPI.OBJECT object

164
mpiJava send()/recv()
• Send and receive members of Comm:
void Send(Object buf, int offset, int
count, Datatype type, int dst, int
tag);
Status Recv(Object buf, int offset,
int count, Datatype type, int src, int
tag);
• buf must be an array.
• offset is the element where message starts.
• Datatype class describes type of elements.

165
Communicators
• A communicator defines which processes may be
involved in the communication. In most
elementary applications the MPI-supplied
communicator MPI.COMM_WORLD is used.
• Two processes can communicate only if they use
the same communicator.
• User-defined datatypes can be used, but mostly the
standard MPI-supplied datatypes are used, such as
MPI.INT and MPI.FLOAT.

166
Process ranks
• When an MPI program is started the number of
processes ,N, is supplied to the program from the
invoking environment. The number of processes in
use can be determined from within the MPI program
with the Size() method.
int Comm.Size()
• Each of the N processes is identified by a unique
integer in the range 0 to N-1. This is called the
process rank. A processcan determine its rank with
the Rank() method.
int Comm.Rank()

167
Message Tags
• The message tag can be used to distinguish
between different types of message. The tag
specified by the receiver must match that of the
sender.
• In a Recv() routine the message source and tag
arguments can have the values
MPI.ANY_SOURCE and MPI.ANY_TAG. These
are called wildcards and indicate that the
requirement for an exact match does not apply.

168
Return Status Objects
• If the message source and/or tag are/is
wildcarded, then the actual source and tag can
be found from the publicly accessible
source and tag fields of the status object
returned by recv().
• The number of items received can be found
using:
int Status.Get_count(Datatype datatype)

169
Communication Completion
• A communication operation is locally complete on
a process if the process has completed its part in
the operation.
• A communication operation is globally complete if
all processes involved have completed their part in
the operation.
• A communication operation is globally complete
if and only if it is locally complete for all
processes.

170
Summary of Point-to-Point
Communication
• Message selectivity on the receiver is by rank and
message tag.
• Rank and tag are interpreted relative to the scope
of the communication.
• The scope is specified by the communicator.
• Source rank and tag may be wildcarded on the
receiver.
• Communicators must match on sender and
receiver.

PMSCS 657_Parallel and Distributed processing

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to PMSCS 657_Parallel and Distributed processing (20)

More from Md. Mashiur Rahman (20)

Recently uploaded (20)

PMSCS 657_Parallel and Distributed processing