slides8 SharedMemory.ppt

8a-1
Programming with Shared Memory
Part 1
Threads
Accessing shared data
Critical sections
ITCS4145/5145, Parallel Programming B. Wilkinson Fall 2010.

2
Shared memory multiprocessor
system
Any memory location can be accessible by any of the
processors.
A single address space exists, meaning that each
memory location is given a unique address within a
single range of addresses.
Programming a shared memory multiprocessor
system can take advantage of data stored in the
shared memory, which is accessible by all processors
without having to send the data to destination through
message passing

3
Shared Memory Programming
Generally, shared memory programming more
convenient although it does require access to shared
data by different processors to be carefully controlled
by the programmer.
Shared memory systems have been around for a long
time but with the advent of multi-core systems, it has
become very important to able to program for them

4
Methods for Programming Shared
Memory Multiprocessors
• Using heavyweight processes.
• Using threads. Example Pthreads, Java threads
• Using a completely new programming language for parallel
programming - not popular. Example Ada.
• Modifying the syntax of an existing sequential programming
language to create a parallel programming language.
Example UPC
• Using an existing sequential programming language
supplemented with compiler directives and libraries for
specifying parallelism. Example OpenMP
We will look mostly at threads and OpenMP

5
Using Heavyweight Processes
Operating systems often based upon notion of a process.
Processor time shares between processes, switching from
one process to another. Might occur at regular intervals or
when an active process becomes delayed.
Offers opportunity to de-schedule processes blocked from
proceeding for some reason, e.g. waiting for an I/O operation
to complete.
Concept could be used for parallel programming. Not much
used because of overhead but fork/join concepts used
elsewhere.

8
UNIX System Calls
SPMD model with different code for master process and
forked slave process.

9
Differences between a process and threads

10
Statement Execution Order
Single processor: Processes/threads typically executed until blocked.
Multiprocessor: Instructions of processes/threads interleaved in time.
Example
Process 1 Process 2
Instruction 1.1 Instruction 2.1
Many possible orderings, e.g.:
Instruction 1.1
Instruction 1.2
Instruction 2.1
Instruction 1.3
Instruction 2.2
Instruction 2.3
assuming instructions cannot be divided into smaller steps.

11
A sequence of instructions might perform a
specific task, for example print out a message.
If two processes were to print messages, using
interleaved instructions, the messages could
appear garbled - the individual characters of
each message could be interleaved if special
care is not taken in coding the print routines.

12
Thread-Safe Routines
Thread safe if they can be called from multiple threads
simultaneously and always produce correct results.
Standard I/O thread safe (prints messages without
interleaving the characters).
System routines that return time may not be thread safe.
Routines that access shared data may require special care to
be made thread safe.

8a-13
Re-ordering code also done intentionally by:
• Compilers to optimize performance
• Processors internally, again to optimize
performance
Compilers will re-order code prior to execution while
processors will re-order the code during execution.
In both cases the objective is to best utilize the
available computer resources and minimize any
waiting time.

14
Compiler/Processor Optimizations
Compiler and processor reorder instructions to improve
performance (execution speed).
Example
Suppose one had the code:
.
.
a = b + 5;
x = y * 4;
p = x + 9;
.
.
and the processor can perform, as is usual, multiple arithmetic
operations at the same time.

15
Compiler/Processor Optimizations
Can reorder to:
.
.
x = y * 4;
a = b + 5;
p = x + 9;
.
.
and still be logically correct. This gives the multiply operation
longer time to complete before the result (x) is needed in the last
statement.
Very common for processors to execute machines instructions out
of order for increased speed.

16
Accessing Shared Data
Accessing shared data needs careful control.
Consider two processes each of which is to add one to a
shared data item, x.
Location x is read, x + 1 computed, and the result written
back to the location:

17
Conflict in accessing shared variable

18
Critical Section
A mechanism for ensuring that only one process accesses a
particular resource at a time.
critical section – a section of code for accessing resource
Arrange that only one such critical section is executed at a
time.
This mechanism is known as mutual exclusion.
Concept also appears in an operating systems.

19
Locks
Simplest mechanism for ensuring mutual exclusion of critical
sections.
A lock - a 1-bit variable that is a 1 to indicate that a process
has entered the critical section and a 0 to indicate that no
process is in the critical section.
Operates much like that of a door lock:
A process coming to the “door” of a critical section and
finding it open may enter the critical section, locking the door
behind it to prevent other processes from entering. Once the
process has finished the critical section, it unlocks the door
and leaves.

20
Control of critical sections through
busy waiting

21
Critical Sections Serializing Code
High performance programs should have as few as
possible critical sections as their use can serialize the
code.
Suppose, all processes happen to come to their critical
section together.
They will execute their critical sections one after the other.
In that situation, the execution time becomes almost that of
a single processor.

23
Deadlock
Can occur with two processes when one requires a resource
held by the other, and this process requires a resource held by
the first process.

24
Deadlock (deadly embrace)
Deadlock can also occur in a circular fashion with several
processes having a resource wanted by another.

25
Program example
To sum the elements of an array, a[1000]:
int sum, a[1000];
sum = 0;
for (i = 0; i < 1000; i++)
sum = sum + a[i];

26
UNIX Processes
Calculation will be divided into two parts, one doing even i and
one doing odd i; i.e.,
Process 1 Process 2
sum1 = 0; sum2 = 0;
for (i = 0; i < 1000; i = i + 2) for (i = 1; i < 1000; i = i + 2)
sum1 = sum1 + a[i]; sum2 = sum2 + a[i];
Each process will add its result (sum1 or sum2) to an
accumulating result, sum :
sum = sum + sum1; sum = sum + sum2;
Sum will need to be shared and protected by a lock. Shared
data structure is created:

27
ITCS4145/5145, Parallel Programming B. Wilkinson Fall 2010, Oct 27, 2010
Introduction to OpenMP

28
OpenMP
An accepted standard developed in the late 1990s by a
group of industry specialists.
Consists of a small set of compiler directives, augmented
with a small set of library routines and environment variables
using the base language Fortran and C/C++.
Several OpenMP compilers available.

29
Wikipedia OpenMP
http://en.wikipedia.org/wiki/File:OpenMP_language_extensions.svg

30
OpenMP
• Uses a thread-based shared memory programming
model
• OpenMP programs will create multiple threads
• All threads have access to global memory
• Data can be shared among all threads or private to one
thread
• Synchronization occurs but often implicit

31
OpenMP uses “fork-join” model but thread-based.
Initially, a single thread executed by a master thread.
parallel directive creates a team of threads with a specified
block of code executed by the multiple threads in parallel.
The exact number of threads in the team determined by one
of several ways.
Other directives used within a parallel construct to specify
parallel for loops and different blocks of code for threads.

32
parallel region
Multiple threads
parallel region
Master thread
Fork/join model
Synchronization

33
For C/C++, the OpenMP directives contained in #pragma
statements.
Format:
#pragma omp directive_name ...
where omp is an OpenMP keyword.
May be additional parameters (clauses) after directive name
for different options.
Some directives require code to specified in a structured
block that follows directive and then directive and structured
block form a “construct”.

34
Parallel Directive
#pragma omp parallel
structured_block
creates multiple threads, each one executing the specified
structured_block, (a single statement or a compound
statement created with { ...} with a single entry point and a
single exit point.)
Implicit barrier at end of construct.
Directive corresponds to forall construct.

35
Hello world example
#pragma omp parallel
{
printf("Hello World from thread %d of %dn",
omp_get_thread_num(),
omp_get_num_threads() );
}
Output from an 8-processor/core machine:
Hello World from thread 0 of 8
OpenMP
directive for a
parallel region
Opening
brace must
be on a new
line

36
Private and shared variables
Variables could be declared within each parallel region but
OpenMP provides private clause.
int myid;
…
#pragma omp parallel private(myid)
{
myid = omp_get_thread_num();
printf("Hello World from thread = %dn", myid);
}
Each thread
has a local
variable tid
Also a shared clause available.

37
Example
#pragma omp parallel private(myid, threads)
{
myid = omp_get_thread_num(); // myid is private
threads = omp_get_num_threads(); // threads is private
arr[myid] = 10 * threads; // arr is global
}

38
Number of threads in a team
Established by either:
1. num_threads clause after the parallel directive, or
2. omp_set_num_threads() library routine being previously
called, or
3. Environment variable OMP_NUM_THREADS is defined
in order given or is system dependent if none of above.
Number of threads available can also be altered dynamically to
achieve best use of system resources.

39
Work-Sharing
Three constructs in this classification:
sections
for
single
In all cases, there is an implicit barrier at end of construct
unless a nowait clause included, which overrides the barrier.
Note: These constructs do not start a new team of threads.
That done by an enclosing parallel construct.

40
Sections
The construct
#pragma omp sections
{
#pragma omp section
structured_block
.
.
.
#pragma omp section
structured_block
}
cause structured blocks to be shared among threads in team.
The first section directive optional.
Blocks
executed by
available
threads

41
Example
#pragma omp parallel shared(a,b,c,d,nthreads) private(I,tid)
{
tid = omp_get_thread_num();
#pragma omp sections nowait
{
#pragma omp section
{
printf("Thread %d doing section 1n",tid);
for (i=0; i<N; i++) {
c[i] = a[i] + b[i];
printf("Thread %d: c[%d]= %fn",tid,i,c[i]);
}
}
#pragma omp section
{
printf("Thread %d doing section 2n",tid);
for (i=0; i<N; i++) {
d[i] = a[i] * b[i];
printf("Thread %d: d[%d]= %fn",tid,i,d[i]);
}
}
} /* end of sections */
} /* end of parallel section */
One
thread
does this
Another
thread
does this

42
For Loop
#pragma omp for
for ( i = 0; …. )
causes for loop to be divided into parts and parts shared
among threads in the team. for loop must be of a simple form.
Loops divided can be specified by additional “schedule”
clause.
For loop of a
simple form

43
Loop Scheduling and Partitioning
• Static
#pragma omp parallel for schedule (static,chunk_size)
Partitions loop iterations into equal sized chunks specified by
chunk_size. Chunks assigned to threads in round robin
fashion.
• Dynamic
#pragma omp parallel for schedule (dynamic,chunk_size)
Uses internal work queue. Chunk-sized block of loop assigned
to threads as they become available.

44
Example
#pragma omp parallel shared(a,b,c,nthreads,chunk) private(i,tid)
{
tid = omp_get_thread_num();
if (tid == 0) {
nthreads = omp_get_num_threads();
printf("Number of threads = %dn", nthreads);
}
printf("Thread %d starting...n“,tid);
#pragma omp for schedule(dynamic,chunk)
for (i=0; i<N; i++) {
c[i] = a[i] + b[i];
printf("Thread %d: c[%d]= %fn",tid,i,c[i]);
}
} /* end of parallel section */
For loop
Executed by
one thread

45
Single
The directive
#pragma omp single
structured block
cause the structured block to be executed by one thread only.

46
Combined Parallel Work-sharing
Constructs
If a parallel directive is followed by a single for directive, it
can be combined into:
#pragma omp parallel for
<for loop>
with similar effects.

47
If a parallel directive is followed by a single sections directive,
it can be combined into
#pragma omp parallel sections
{
#pragma omp section
structured_block
#pragma omp section
structured_block
.
.
.
}
with similar effect. (In both cases, the nowait clause is not
allowed.)

48
Master Directive
The master directive:
#pragma omp master
structured_block
causes the master thread to execute the structured block.
Different to those in the work sharing group in that there is
no implied barrier at the end of the construct (nor the
beginning). Other threads encountering this directive will
ignore it and the associated structured block, and will move
on.

49
• Guided
#pragma omp parallel for schedule (guided,chunk_size)
Similar to dynamic but chunk size starts large and gets smaller
to reduce time threads have to go to work queue.
chunk size = number of iterations remaining
2 * number of threads
• Runtime
#pragma omp parallel for schedule (runtime)
Uses OMP_SCHEDULE environment variable to specify which
of static, dynamic or guided should be used.

50
Reduction clause
Used combined the result of the iterations into a single
value c.f. with MPI _Reduce().
Can be used with parallel, for, and sections,
Example
sum = 0
#pragma omp parallel for reduction(+:sum)
for (k = 0; k < 100; k++ ) {
sum = sum + funct(k);
}
Private copy of sum created for each thread by complier.
Private copy will be added to sum at end.
Eliminates here the need for critical sections.
Operation
Variable

51
Private variables
private clause – creates private copies of variables for
each thread
firstprivate clause - as private clause but initializes each
copy to the values given immediately prior to parallel
construct.
lastprivate clause – as private but “the value of each
lastprivate variable from the sequentially last iteration of
the associated loop, or the lexically last section directive,
is assigned to the variable’s original object.”

52
Synchronization Constructs
Critical
critical directive will only allow one thread execute the
associated structured block. When one or more threads
reach the critical directive:
#pragma omp critical name
structured_block
they will wait until no other thread is executing the same
critical section (one with the same name), and then one
thread will proceed to execute the structured block.
name is optional. All critical sections with no name map to
one undefined name.

53
Barrier
When a thread reaches the barrier
#pragma omp barrier
it waits until all threads have reached the barrier and then they
all proceed together.
There are restrictions on the placement of barrier directive in a
program. In particular, all threads must be able to reach the
barrier.

54
Atomic
The atomic directive
#pragma omp atomic
expression_statement
implements a critical section efficiently when the critical
section simply updates a variable (adds one, subtracts one,
or does some other simple arithmetic operation as defined
by expression_statement).

55
Flush
A synchronization point which causes thread to have a
“consistent” view of certain or all shared variables in memory.
All current read and write operations on variables allowed to
complete and values written back to memory but any memory
operations in code after flush are not started.
Format:
#pragma omp flush (variable_list)
Only applied to thread executing flush, not to all threads in team.
Flush occurs automatically at entry and exit of parallel and critical
directives, and at the exit of for, sections, and single (if a no-wait
clause is not present).

56
Ordered clause
Used in conjunction with for and parallel for directives to
cause an iteration to be executed in the order that it
would have occurred if written as a sequential loop.

57
More information
Full information on OpenMP at
http://openmp.org/wp/

slides 8d-58
Specifying parallelism
Performance issues
ITCS4145/5145, Parallel Programming B. Wilkinson Fall 2010.

59
We have seen OpenMP for specifying parallelism
Programmer decides on what parts of the code
should be parallelized and inserts compiler
directives (pragma’s)
Whatever programming environment we use where
the programmers explicitly says what should be
done in parallel, the issue for the programmer is
deciding what can be done in parallel.
Let us use generic language constructs for
parallelism.

60
par Construct
For specifying concurrent statements:
par {
S1;
S2;
.
.
Sn;
}
Says one can execute
all statement S1 to Sn
simultaneously if
resources available, or
execute them in any
order and still get the
correct result

61
forall Construct
To start multiple similar processes together:
forall (i = 0; i < n; i++) {
S1;
S2;
.
.
Sm;
}
Says each iteration of
body can be executed
simultaneously if
resources available, or in
any order and still get the
correct result.
The statements of each
instance of body executed
in order given.
Each instance of the body
uses a different value of i.

62
Example
forall (i = 0; i < 5; i++)
a[i] = 0;
clears a[0], a[1], a[2], a[3], and a[4] to zero concurrently.

63
Dependency Analysis
To identify which processes could be executed together.
Example
Can see immediately in the code
forall (i = 0; i < 5; i++)
a[i] = 0;
that every instance of the body is independent of other
instances and all instances can be executed simultaneously.
However, it may not be that obvious. Need algorithmic way
of recognizing dependencies, for a parallelizing compiler.

66
Can use Berstein’s conditions at:
• Machine instruction level inside processor – have logic
to detect if conditions satisfied (see computer
architecture course)
• At the process level to detect whether two processes
can be executed simultaneously (using the inputs and
outputs of processes).
• Can be extended to more than two processes but
number of conditions rises – need every input/output
combination checked.
For three statements, need how many conditions
checked?

67
Shared Memory Programming
Performance Issues

68
Performance issues with Threads
Program might actually go slower when parallelized.
Too many threads can significantly reduce the program
performance.

69
Reasons:
• Work split among too many threads gives each thread too
little work, so overhead of starting and terminating threads
swamps useful work.
• To many concurrent threads incurs overhead from having to
share fixed hardware resources
• OS typically schedules threads in round robin with a time-
slice
• Time-slicing incurs overhead
• Saving registers, effects on cache memory, virtual
memory management … .
• Waiting to acquire a lock.
• When a thread is suspended while holding a lock, all
threads waiting for lock will have to wait for thread to re-
start.
Source: Multi-core programming by S. Akhter and J. Roberts, Intel Press.

70
Some Strategies
• Limit number of runnable threads to number of hardware
threads. (See later we do not do this with GPUs)
• For a n-core machine (not hyper-threaded) have n runnable
threads.
• If hyper-threaded (with 2 virtual threads per core) double
this.
• Can have more threads in total but others may be blocked.
• Separate I/O threads from compute threads
• I/O threads wait for external events
• Never hard-code number of threads – leave as a tuning
parameter.

71
• Let OpenMP optimize number of threads
• Implement a thread pool
• Implement a work stealing approach in which threads
has a work queue. Threads with no work take work
from other threads

72
Critical Sections Serializing Code
High performance programs should have as few as
possible critical sections as their use can serialize the
code.
Suppose, all processes happen to come to their critical
section together.
They will execute their critical sections one after the other.
In that situation, the execution time becomes almost that of
a single processor.

74
Shared Data in Systems with Caches
All modern computer systems have cache memory, high-
speed memory closely attached to each processor for holding
recently referenced data and code.
Processors
Cache memory
Main memory

75
Cache coherence protocols
Update policy - copies of data in all caches are updated at
the time one copy is altered,
or
Invalidate policy - when one copy of data is altered, the same
data in any other cache is invalidated (by resetting a valid bit
in the cache). These copies are only updated when the
associated processor makes reference for it.

76
False Sharing
Different parts of block
required by different
processors but not
same bytes.
If one processor writes
to one part of the
block, copies of the
complete block in other
caches must be
updated or invalidated
although the actual
data is not shared.

77
Solution for False Sharing
Compiler to alter the layout of the data stored in the
main memory, separating data only altered by one
processor into different blocks.

78
Sequential Consistency
Formally defined by Lamport (1979):
A multiprocessor is sequentially consistent if the result
of any execution is the same as if the operations of all
the processors were executed in some sequential
order, and the operations of each individual
processors occur in this sequence in the order
specified by its program.
i.e., the overall effect of a parallel program is not
changed by any arbitrary interleaving of instruction
execution in time.

80
Writing a parallel program for a system which is
known to be sequentially consistent enables us to
reason about the result of the program.

81
Example
Process P1 Process 2
. .
data = new; .
flag = TRUE; .
. .
. while (flag != TRUE) { };
. data_copy = data;
. .
Expect data_copy to be set to new because we expect data =
new to be executed before flag = TRUE and while (flag !=
TRUE) { } to be executed before data_copy = data.
Ensures that process 2 reads new data from process 1.
Process 2 will simple wait for the new data to be produced.

82
Program Order
Sequential consistency refers to “operations of each
individual processor .. occur in the order specified in its
program” or program order.
In previous figure, this order is that of the stored machine
instructions to be executed.

83
Compiler Optimizations
The order of execution is not necessarily the same as the
order of the corresponding high level statements in the
source program as a compiler may reorder statements for
improved performance.
In this case, term program order will depend upon context,
either the order in the source program or the order in the
compiled machine instructions.

84
High Performance Processors
Modern processors also usually reorder machine instructions
internally during execution for increased performance.
Does not alter a multiprocessor being sequential consistency, if
processor only produces final results in program order (that is,
retrives values to registers in program order, which most
processors do).
All multiprocessors will have the option of operating under the
sequential consistency model. However, it can severely limit
compiler optimizations and processor performance.

85
Example of Processor Re-ordering
. .
new = a * b; .
data = new; .
flag = TRUE; .
. .
. data_copy = data;
. .
Multiply machine instruction corresponding to new = a * b is issued
for execution.
Next instruction corresponding to data = new cannot be issued until
the multiply has produced its result.
However following statement, flag = TRUE, completely independent
and a clever processor could start this operation before multiply has
completed leading to the sequence:

86
. .
new = a * b; .
flag = TRUE; .
data = new; .
. .
. data_copy = data;
. .
Now while statement might occur before new assigned to
data, and code would fail.
All multiprocessors have option of operating under sequential
consistency model, i.e. not reorder instructions and forcing
multiply instruction above to complete before starting
subsequent instruction that depend upon its result.

87
Relaxing Read/Write Orders
Processors may be able to relax the consistency in
terms of the order of reads and writes of one processor
with respect to those of another processor to obtain
higher performance, and instructions to enforce
consistency when needed.

88
Examples
Alpha processors
Memory barrier (MB) instruction - waits for all previously issued
memory accesses instructions to complete before issuing any
new memory operations.
Write memory barrier (WMB) instruction - as MB but only on
memory write operations, i.e. waits for all previously issued
memory write accesses instructions to complete before issuing
any new memory write operations - which means memory
reads could be issued after a memory write operation but
overtake it and complete before the write operation.

89
SUN Sparc V9 processors
memory barrier (MEMBAR) instruction with four bits for
variations Write-to-read bit prevent any reads that follow it
being issued before all writes that precede it have completed.
Other: Write-to-write, read-to-read, read-to-write.
IBM PowerPC processor
SYNC instruction - similar to Alpha MB instruction (check
differences)

slides8 SharedMemory.ppt

More Related Content

Similar to slides8 SharedMemory.ppt (20)

More from aminnezarat (16)

Recently uploaded (20)

slides8 SharedMemory.ppt