2. Course Objectives
• Introduction and comprehensive
coverage of the OpenMP
standard
• After completion of this course
users should be equipped to
implement OpenMP constructs
in their applications and realize
performance improvements on
shared memory machines
3. Course Overview
• Introduction
– Objectives, History, Overview, Motivation
• Getting Our Feet Wet
– Memory Architectures, Models (programming,
execution, memory, …), compiling and running
• Diving In
– Control constructs, worksharing, data scoping,
synchronization, runtime control
4. Course Overview Cont.
• Swimming Like a Champion
– more on parallel clauses, all work sharing
constructs, all synchronization constructs,
orphaned directives, examples and tips
• On to the Olympics
– Performance! Coarse/Fine grain
parallelism, data decomposition and
SPMD, false sharing and cache
considerations, summary and suggestions
5. Motivation
• No portable standard for shared memory
parallelism existed
– Each SMP vendor has proprietary API
– X3H5 and PCF efforts failed
• Portability only through MPI
• Parallel application availability
– ISVs have big investment in existing code
– New parallel languages not widely adopted
6. In a Nutshell
• Portable, Shared Memory Multiprocessing API
– Multi-vendor support
– Multi-OS support (Unixes, Windows, Mac)
• Standardizes fine grained (loop) parallelism
• Also supports coarse grained algorithms
• The MP in OpenMP is for multi-processing
• Don’t confuse OpenMP with Open MPI! :)
7. Version History
• First
– Fortran 1.0 was released in October 1997
– C/C++ 1.0 was approved in November 1998
• Recent
– version 2.5 joint specification in May 2005
– this may be the version most available
• Current
– OpenMP 3.0 API released May 2008
– check compiler for level of support
8. Who’s Involved
OpenMP Architecture Review Board (ARB)
Permanent Members of ARB:
• AMD
• Cray
• Fujitsu
• HP
• IBM
• Intel
• NEC
• The Portland Group
• SGI
• Sun
• Microsoft
Auxiliary Members of ARB:
• ASC/LLNL
• cOMPunity
• EPCC
• NASA
• RWTH Aachen University
9. Why choose OpenMP ?
• Portable
– standardized for shared memory architectures
• Simple and Quick
– incremental parallelization
– supports both fine grained and coarse grained
parallelism
– scalable algorithms without message passing
• Compact API
– simple and limited set of directives
13. Distributed vs. Shared
Memory
• Shared - all processors share a global pool of
memory
– simpler to program
– bus contention leads to poor scalability
• Distributed - each processor physically has it’s
own (private) memory associated with it
– scales well
– memory management is more difficult
14. Models …
No Not These!
Nor These Either!
We want programming models, execution models,
communication models and memory models!
15. OpenMP - User Interface Model
• Shared Memory with thread based parallelism
• Not a new language
• Compiler directives, library calls and environment
variables extend the base language
– f77, f90, f95, C, C++
• Not automatic parallelization
– user explicitly specifies parallel execution
– compiler does not ignore user directives even if wrong
16. What is a thread?
• A thread is an independent instruction stream,
thus allowing concurrent operation
• threads tend to share state and memory
information and may have some (usually
small) private data
• Similar (but distinct) from processes. Threads
are usually lighter weight allowing faster
context switching
• in OpenMP one usually wants no more than
one thread per core
17. Execution Model
• OpenMP program starts single threaded
• To create additional threads, user starts a parallel
region
– additional threads are launched to create a team
– original (master) thread is part of the team
– threads “go away” at the end of the parallel region:
usually sleep or spin
• Repeat parallel regions as necessary
– Fork-join model
18. Communicating Between Threads
• Shared Memory Model
– threads read and write shared variables
• no need for explicit message passing
– use synchronization to protect against race
conditions
– change storage attributes for minimizing
synchronization and improving cache reuse
19. Storage Model – Data Scoping
• Shared memory programming model: variables are
shared by default
• Global variables are SHARED among threads
– Fortran: COMMON blocks, SAVE variables, MODULE variables
– C: file scope variables, static
• Private Variables:
– exist only within the new scope, i.e. they are uninitialized and
undefined outside the data scope
– loop index variables
– Stack variables in sub-programs called from parallel regions
20. Creating Parallel Regions
• Only one way to create threads in OpenMP API:
• Fortran:
!$OMP parallel
< code to be executed in parallel >
!$OMP end parallel
• C
#pragma omp parallel
{
code to be executed by each thread
}
21.
22.
23. Comparison of Programming Models
Feature Open MP MPI
Portable yes yes
Scalable less so yes
Incremental
Parallelization
yes no
Fortran/C/C++
Bindings
yes yes
High Level yes mid level
24. Compiling
• Intel (icc, ifort, icpc)
– -openmp
• PGI (pgcc, pgf90, …)
– -mp
• GNU (gcc, gfortran, g++)
– -fopenmp
– need version 4.2 or later
– g95 was based on GCC but branched off
• I don’t think it has Open MP support
25. Compilers
• No specific Fortran 90 or C++ features are
required by the OpenMP specification
• Most compilers support OpenMP, see
compiler documentation for the appropriate
compiler flag to set for other compilers, e.g.
IBM, SGI, …
26. Compiler Directives
C Pragmas
• C pragmas are case sensitive
• Use curly braces, {}, to enclose parallel regions
• Long directive lines can be "continued" by
escaping the newline character with a backslash
("") at the end of a directive line.
Fortran
• !$OMP, c$OMP, *$OMP – fixed format
• !$OMP – free format
• Comments may not appear on the same line
• continue w/ &, e.g. !$OMP&
27. Specifying threads
• The simplest way to specify the number of threads
used on a parallel region is to set the environment
variable (in the shell where the program is executing)
– OMP_NUM_THREADS
• For example, in csh/tcsh
– setenv OMP_NUM_THREADS 4
• in bash
– export OMP_NUM_THREADS=4
• Later we will cover other ways to specify this
30. OpenMP Language Features
• Compiler Directives – 3 categories
– Control Constructs
• parallel regions, distribute work
– Data Scoping for Control Constructs
• control shared and private attributes
– Synchronization
• barriers, atomic, …
• Runtime Control
– Environment Variables
• OMP_NUM_THREADS
– Library Calls
• OMP_SET_NUM_THREADS(…)
33. Supported Clauses for the Parallel
Construct
Valid Clauses:
– if (expression)
– num_threads (integer expression)
– private (list)
– firstprivate (list)
– shared (list)
– default (none|shared|private *fortran only*)
– copyin (list)
– reduction (operator: list)
• More on these later …
34. Loop Worksharing directive
• Loop directives:
– !$OMP DO [clause[[,] clause]… ] Fortran do
– [!$OMP END DO [NOWAIT]] optional end
– #pragma omp for [clause[[,] clause]… ] C/C++ for
• Clauses:
– PRIVATE(list)
– FIRSTPRIVATE(list)
– LASTPRIVATE(list)
– REDUCTION({op|intrinsic}:list})
– ORDERED
– SCHEDULE(TYPE[, chunk_size])
– NOWAIT
35. All Worksharing Directives
• Divide work in enclosed region among threads
• Rules:
– must be enclosed in a parallel region
– does not launch new threads
– no implied barrier on entry
– implied barrier upon exit
– must be encountered by all threads on team or none
36. Loop Constructs
• Note that many of the clauses are the same as
the clauses for the parallel region. Others are not,
e.g. shared must clearly be specified before a
parallel region.
• Because the use of parallel followed by a loop
construct is so common, this shorthand notation
is often used (note: directive should be followed
immediately by the loop)
– !$OMP parallel do …
– !$OMP end parallel do
– #pragma parallel for …
37. Uninitialized!
Undefined!
Private Clause
• Private, uninitialized copy is created for each
thread
• Private copy is not storage associated with the
original
program wrong
I = 10
!$OMP parallel private(I)
I = I + 1
!$OMP end parallel
print *, I
38. Firstprivate Clause
• Firstprivate, initializes each private copy with
the original
program correct
I = 10
!$OMP parallel firstprivate(I)
I = I + 1
!$OMP end parallel
Initialized!
39. LASTPRIVATE clause
• Useful when loop index is live out
– Recall that if you use PRIVATE the loop index
becomes undefined
do i=1,N-1
a(i)= b(i+1)
enddo
a(i) = b(0)
In Sequential
case
i=N
!$OMP PARALLEL
!$OMP DO LASTPRIVATE(i)
do i=1,N-1
a(i)= b(i+1)
enddo
a(i) = b(0)
!$OMP END PARALLEL
40. NOWAIT clause
• NOWAIT clause
!$OMP PARALLEL
!$OMP DO
do i=1,n
a(i)= cos(a(i))
enddo
!$OMP END DO
!$OMP DO
do i=1,n
b(i)=a(i)+b(i)
enddo
!$OMP END DO
!$OMP END PARALLEL
!$OMP PARALLEL
!$OMP DO
do i=1,n
a(i)= cos(a(i))
enddo
!$OMP END DO NOWAIT
!$OMP DO
do i=1,n
b(i)=a(i)+b(i)
enddo
!$OMP END DO
Implied
BARRIER
No
BARRIER
By default loop
index is PRIVATE
41. Reductions
• Assume no reduction clause
do i=1,N
X = X + a(i)
enddo
Sum Reduction
!$OMP PARALLEL DO SHARED(X)
do i=1,N
X = X + a(i)
enddo
!$OMP END PARALLEL DO
!$OMP PARALLEL DO SHARED(X)
do i=1,N
!$OMP CRITICAL
X = X + a(i)
!$OMP END CRITICAL
enddo
!$OMP END PARALLEL DO
Wrong!
What’s wrong?
42. REDUCTION clause
• Parallel reduction operators
– Most operators and intrinsics are supported
– +, *, -, .AND. , .OR., MAX, MIN
• Only scalar variables allowed
!$OMP PARALLEL DO REDUCTION(+:X)
do i=1,N
X = X + a(i)
enddo
!$OMP END PARALLEL DO
do i=1,N
X = X + a(i)
enddo
43. Ordered clause
• Executes in the same order as sequential code
• Parallelizes cases where ordering needed
do i=1,N
call find(i,norm)
print*, i,norm
enddo
!$OMP PARALLEL DO ORDERED PRIVATE(norm)
do i=1,N
call find(i,norm)
!$OMP ORDERED
print*, i,norm
!$OMP END ORDERED
enddo
!$OMP END PARALLEL DO
1 0.45
2 0.86
3 0.65
44. Schedule clause
• Controls how the iterations of the loop are assigned to
threads
– static: Each thread is given a “chunk” of iterations in a round
robin order
• Least overhead - determined statically
– dynamic: Each thread is given “chunk” iterations at a time; more
chunks distributed as threads finish
• Good for load balancing
– guided: Similar to dynamic, but chunk size is reduced
exponentially
– runtime: User chooses at runtime using environment variable
(note no space before chunk value)
• setenv OMP_SCHEDULE “dynamic,4”
45. Performance Impact of Schedule
• Static vs. Dynamic across multiple
do loops
– In static, iterations of the do loop
executed by the same thread in both
loops
– If data is small enough, may be still in
cache, good performance
• Effect of chunk size
– Chunk size of 1 may result in multiple
threads writing to the same cache line
– Cache thrashing, bad performance
a(1,1) a(1,2)
a(2,1) a(2,2)
a(3,1) a(3,2)
a(4,1) a(4,2)
!$OMP DO SCHEDULE(STATIC)
do i=1,4
!$OMP DO SCHEDULE(STATIC)
do i=1,4
47. Barrier Synchronizaton
• Syntax:
– !$OMP barrier
– #pragma omp barrier
• Threads wait until all
threads reach this point
• implicit barrier at the end
of each parallel region
48. Atomic Update
• Specifies a specific memory location can only be
updated atomically, i.e. 1 thread at a time
• Optimization of mutual exclusion for certain cases
(i.e. a single statement CRITICAL section)
– applies only to the statement immediately following the
directive
– enables fast implementation on some HW
• Directive:
– !$OMP atomic
– #pragma atomic
49.
50.
51. Environment Variables
• These are set outside the program and control execution of
the parallel code
• There are only 4, all are uppercase but values are case
insensitive
• OMP_SCHEDULE – determines how iterations are scheduled
when a schedule clause is set to “runtime”
– “type[, chunk]”
• OMP_NUM_THREADS – set maximum number of threads
– integer value
• OMP_DYNAMIC – dynamic adjustment of threads for parallel
regions
– true or false
• OMP_NESTED – nested parallelism
– true or false
52. Run-time Library Routines
• There are 17 different library routines, we will
cover just 3 of them now
• omp_get_thread_num()
– Returns the thread number (w/i the team) of the calling
thread. Numbering starts w/ 0.
• integer function omp_get_thread_num()
• #include <omp.h>
int omp_get_thread_num()
53. Run-time Library: Timing
• There are 2 portable timing routines
• omp_get_wtime
– portable wall clock timer returns a double
precision value that is number of elapsed seconds
from some point in the past
– gives time per thread - possibly not globally
consistent
– difference 2 times to get elapsed time in code
• omp_get_wtick
– time between ticks in seconds
54. Run-time Library: Timing
• double precision function omp_get_wtime()
• include <omp.h>
double omp_get_wtime()
• double precision function omp_get_wtick()
• include <omp.h>
double omp_get_wtick()
56. Supported Clauses for the Parallel
Construct
Valid Clauses:
– if (expression)
– num_threads (integer expression)
– private (list)
– firstprivate (list)
– shared (list)
– default (none|shared|private *fortran only*)
– copyin (list)
– reduction (operator: list)
• More on these later …
57. Shared Clause
• Declares variables in the list to be shared
among all threads in the team.
• Variable exists in only 1 memory location and
all threads can read or write to that address.
• It is the user’s responsibility to ensure that
this is accessed correctly, e.g. avoid race
conditions
• Most variables are shared by default (a
notable exception, loop indices)
58. Changing the default
• List the variables in one of the following
clauses
– SHARED
– PRIVATE
– FIRSTPRIVATE, LASTPRIVATE
– DEFAULT
– THREADPRIVATE, COPYIN
59. Default Clause
• Note that the default storage attribute is
DEFAULT (SHARED)
• To change default: DEFAULT(PRIVATE)
– each variable in static extent of the parallel region
is made private as if specified by a private clause
– mostly saves typing
• DEFAULT(none): no default for variables in
static extent. Must list storage attribute for
each variable in static extent
USE THIS!
60. If Clause
• if (logical expression)
• executes (in parallel) normally if the
expression is true, otherwise it executes the
parallel region serially
• Used to test if there is sufficient work to justify
the overhead in creating and terminating a
parallel region
61. Num_threads clause
• num_threads(integer expression)
• executes the parallel region on the number of
threads specified
• applies only to the specified parallel region
62. How many threads?
• Order of precedence:
– if clause
– num_threads clause
– omp_set_num_threads function
call
– OMP_NUM_THREADS
environment variable
– implementation default (usually
the number of cores on a node)
63. Work Sharing Constructs -
The Complete List
• Loop directives (DO/for)
• sections
• single
• workshare
64. SECTIONS directive
• Enclosed sections are divided among threads in
the team
• Each section is executed by exactly one thread. A
thread may execute more than one section
!$OMP SECTIONS [clause … ]
!$OMP SECTION
!$OMP SECTION
…
!$OMP END SECTIONS [nowait]
#pragma omp sections [clause …]
{
#pragma omp section
structured_block
#pragma omp section
structured_block
…
}
65. SECTIONS directive
• If there are N sections then there can be at
most N threads executing in parallel
• Any thread can execute a particular section.
This may vary from one execution to the next
(non-deterministic)
• The shorthand for combining sections with
the parallel directive is:
– $!OMP parallel sections …
– #pragma parallel sections
66. SINGLE directive
• SINGLE specifies that the enclosed code is
executed by only 1 thread
• This may be useful for sections of code that
are not thread safe, e.g. some I/O
• It is illegal to branch into or out of a single
block
• !$OMP single [clause …]
• !$OMP end single [nowait]
• #pragma single [clause …]
67. WORKSHARE directive
• Fortran only. It exists to allow work sharing for
code using f90 array syntax, forall statement,
and where statements
• !$OMP workshare [clause …]
• !$OMP end workshare [nowait]
• can combine with parallel construct
– !$OMP parallel workshare …
70. Mutual Exclusion - Critical Sections
• Critical Section
– only 1 thread executes at a time, others block
– can be named (names are global entities and must not conflict
with subroutine or common block names)
– It is good practice to name them
– all unnamed sections are treated as the same region
• Directives:
– !$OMP CRITICAL [name]
– !$OMP END CRITICAL [name]
– #pragma omp critical [name]
71. Master Section
• Only the master (0) thread executes the section
• Rest of the team skips the section and continues execution
from the end of the master
• No barrier at the end (or start) of the master section
• The worksharing construct, OMP single is similar in
behavior but has an implied barrier at the end. Single is
performed by any one thread.
• Syntax:
– !$OMP master
– !$OMP end master
– #pragma omp master
72. Ordered Section
• Enclosed code is executed in the same order
as would occur in sequential execution of the
loop
• Directives:
– !$OMP ORDERED
– !$OMP END ORDERED
– #pragma omp ordered
73. Flush
• synchronization point at which the system must provide
a consistent view of memory
– all thread visible variables must be written back to memory (if
no list is provided), otherwise only those in the list are written
back
• Implicit flushes of all variables occur automatically at
– all explicit and implicit barriers
– entry and exit from critical regions
– entry and exit from lock routines
• Directives
– !$OMP flush [(list)]
– #pragma omp flush [(list)]
74. Sub-programs in parallel regions
• Sub-programs can be called from parallel
regions
• static extent is code contained lexically
• dynamic extent includes static extent + the
statements in the call tree
• the called sub-program can contain OpenMP
directives to control the parallel region
– directives in dynamic extent but not in static
extent are called Orphan directives
75.
76. Threadprivate
• Makes global data private to a thread
– Fortran: COMMON blocks
– C: file scope and static variables
• Different from making them PRIVATE
– with PRIVATE global scope is lost
– THREADPRIVATE preserves global scope for each
thread
• Threadprivate variables can be initialized using
COPYIN
77.
78.
79. Run-time Library Routines
• omp_set_num_threads(integer)
– Set the number of threads to use in next parallel
region.
– can only be called from serial portion of code
– if dynamic threads are enabled, this is the maximum
number allowed, if they are disabled then this is the
exact number used
• omp_get_num_threads
– returns number of threads currently in the team
– returns 1 for serial (or serialized nested) portion of
code
80. Run-time Library Routines Cont.
• omp_get_max_threads
– returns maximum value that can be returned by a
call to omp_get_num_threads
– generally reflects the value as set by
OMP_NUM_THREADS env var or the
omp_set_num_threads library routine
– can be called from serial or parall region
• omp_get_thread_number
– returns thread number. Master is 0. Thread
numbers are contiguous and unique.
81. Run-time Library Routines Cont.
• omp_get_num_procs
– returns number of processors available
• omp_in_parallel
– returns a logical (fortran) or int (C/C++) value
indicating if executing in a parallel region
82. Run-time Library Routines Cont.
• omp_set_dynamic (logical(fortran) or int(C))
– set dynamic adjustment of threads by the run time
system
– must be called from serial region
– takes precedence over the environment variable
– default setting is implementation dependent
• omp_get_dynamic
– used to determine of dynamic thread adjustment is
enabled
– returns logical (fortran) or int (C/C++)
83. Run-time Library Routines Cont.
• omp_set_nested (logical(fortran) or int(C))
– enable nested parallelism
– default is disabled
– overrides environment variable OMP_NESTED
• omp_get_nested
– determine if nested parallelism is enabled
• There are also 5 lock functions which will not
be covered here.
84. Weather Forecasting Example 1
!$OMP PARALLEL DO
!$OMP& default(shared)
!$OMP& private (i,k,l)
do 50 k=1,nztop
do 40 i=1,nx
cWRM remove dependency
cWRM l = l+1
l=(k-1)*nx+i
dcdx(l)=(ux(l)+um(k))
*dcdx(l)+q(l)
40 continue
50 continue
!$OMP end parallel do
• Many parallel loops simply
use parallel do
• autoparallelize when possible
(usually doesn’t work)
• simplify code by removing
unneeded dependencies
• Default (shared) simplifies
shared list but Default (none)
is recommended.
85. Weather - Example 2a
cmass = 0.0
!$OMP parallel default
(shared)
!$OMP&
private(i,j,k,vd,help,..
)
!$OMP& reduction(+:cmass)
do 40 j=1,ny
!$OMP do
do 50 i=1,nx
vd = vdep(i,j)
do 10 k=1,nz
help(k) = c(i,j,k)
10 continue
• Parallel region makes
nested do more efficient
– avoid entering and exiting
parallel mode
• Reduction clause
generates parallel
summing
86. Weather - Example 2a Continued
…
do 30 k=1,nz
c(I,j,k)=help(k)
cmass=cmass+help(k)
30 continue
50 continue
!$OMP end do
40 continue
!$omp end parallel
• Reduction means
– each thread gets private
cmass
– private cmass added at
end of parallel region
– serial code unchanged
87. Weather Example - 3
!$OMP parallel
do 40 j=1,ny
!$OMP do schedule(dynamic)
do30 i=1,nx
if(ish.eq.1)then
call upade(…)
else
call ucrank(…)
endif
30 continue
40 continue
!$OMP end parallel
• Schedule(dynamic) for
load balancing
88. Weather Example - 4
!$OMP parallel !don’t it
slows down
!$OMP& default(shared)
!$OMP& private(i)
do 30 I=1,loop
y2=f2(I)
f2(i)=f0(i) +
2.0*delta*f1(i)
f0(i)=y2
30 continue
!$OMP end parallel do
• Don’t over parallelize
small loops
• Use if(<condition>) clause
when loop is sometimes
big, other times small
89. Weather Example - 5
!$OMP parallel do
schedule(dynamic)
!$OMP& shared(…)
!$OMP& private(help,…)
!$OMP& firstprivate
(savex,savey)
do 30 i=1,nztop
…
30 continue
!$OMP end parallel do
• First private (…) initializes
private variables
90. On to the Olympics - Advanced Topics
The Water Cube
Beijing National Aquatics Center
91. !$OMP PARALLEL DO
do i=1,n
…………
enddo
alpha = xnorm/sum
!$OMP PARALLEL DO
do i=1,n
…………
enddo
!$OMP PARALLEL DO
do i=1,n
…………
enddo
Loop Level Paradigm
• Execute each loop in
parallel
• Easy to parallelize code
• Similar to automatic
parallelization
– Automatic Parallelization
Option (API) may be good
start
• Incremental
– One loop at a time
– Doesn’t break code
92. !$OMP PARALLEL DO
do i=1,n
…………
enddo
alpha = xnorm/sum
!$OMP PARALLEL DO
do i=1,n
…………
enddo
!$OMP PARALLEL DO
do i=1,n
…………
enddo
Performance
• Fine Grain Overhead
– Start a parallel region
each time
– Frequent
synchronization
• Fraction of non
parallel work will
dominate
– Amdahl’s law
• Limited scalability
93. Reducing Overhead
• Convert to coarser grain model:
– More work per parallel region
– Reduce synchronization across threads
• Combine multiple DO directives into single
parallel region
– Continue to use Work-sharing directives
• Compiler does the work of distributing iterations
• Less work for user
– Doesn’t break code
94. !$OMP PARALLEL DO
do i=1,n
…………
enddo
!$OMP PARALLEL DO
do i=1,n
…………
enddo
!$OMP PARALLEL
!$OMP DO
do i=1,n
…………
enddo
!$OMP DO
do i=1,n
…………
enddo
!$OMP END PARALLEL
Coarser Grain
95. !$OMP PARALLEL DO
do i=1,n
…………
enddo
call MatMul(y)
subroutine MatMul(y)
!$OMP PARALLEL DO
do i=1,n
…………
enddo
!$OMP PARALLEL
!$OMP DO
do i=1,n
…………
enddo
call MatMul(y)
!$OMP END PARALLEL
subroutine MatMul(y)
!$OMP DO
do i=1, N
……
enddo
Orphaned
Directive
Using Orphaned Directives
96. !$OMP PARALLEL DO
!$OMP& REDUCTION(+: sum)
do i=1,n
sum = sum + a[i]
enddo
alpha = sum/scale
!$OMP PARALLEL DO
do i=1,n
a[i] = alpha * a[i]
enddo
!$OMP PARALLEL
!$OMP DO REDUCTION(+: sum)
do i=1,n
sum = sum + a[i]
enddo
!$OMP SINGLE
alpha = sum/scale
!$OMP END SINGLE
!$OMP DO
do i=1,n
a[i] = alpha * a[i]
enddo
!$OMP END PARALLEL
Load
Imbalance
Statements Between Loops
97. Replicated
execution
Cannot use
NOWAIT
!$OMP PARALLEL
!$OMP DO REDUCTION(+: sum)
do i=1,n
sum = sum + a[i]
enddo
!$OMP SINGLE
alpha = sum/scale
!$OMP END SINGLE
!$OMP DO
do i=1,n
a[i] = alpha * a[i]
enddo
!$OMP END PARALLEL
!$OMP PARALLEL PRIVATE(alpha)
!$OMP DO REDUCTION(+: sum)
do i=1,n
sum = sum + a[i]
enddo
alpha = sum/scale
!$OMP DO
do i=1,n
a[i] = alpha * a[i]
enddo
!$OMP END PARALLEL
Loops (cont.)
98. Coarse Grain Work-sharing
• Reduced overhead by increasing work per
parallel region
• But…work-sharing constructs still need to
compute loop bounds at each construct
• Work between loops not always parallelizable
• Synchronization at the end of directive not
always avoidable
99. Domain Decomposition
• We could have computed loop bounds once
and used that for all loops
– i.e., compute loop decomposition a priori
• Enter Domain Decomposition
– More general approach to loop decomposition
– Decompose work into the number of threads
– Each threads gets to work on a sub-domain
100. Domain Decomposition (cont.)
• Transfers onus of decomposition from
compiler (work-sharing) to user
• Results in Coarse Grain program
– Typically one parallel region for whole program
– Reduced overhead good scalability
• Domain Decomposition results in a model of
programming called SPMD
101. Program
1 Domain n threads
n sub-domains
Program
SPMD Programming
• SPMD: Single Program Multiple Data
102. program work
!$OMP PARALLEL DEFAULT(PRIVATE)
!$OMP& SHARED(N,global)
nthreads = omp_get_num_threads()
iam = omp_get_thread_num()
ichunk = N/nthreads
istart = iam*ichunk
iend = (iam+1)*ichunk -1
call my_sum(istart, iend, local)
!$OMP ATOMIC
global = global + local
!$OMP END PARALLEL
print*, global
end
Program
Each thread
works on its
piece.
Decomposition
done manually
Implementing SPMD
103. Implementing SPMD (contd.)
• Decomposition is done manually
– Implement to run on any number of threads
• Query for number of threads
• Find thread number
• Each thread calculates its portion (sub-domain) of work
• Program is replicated on each thread, but with
different extents for the sub-domain
– all sub-domain specific data are PRIVATE
104. Handling Global Variables
• Global variables: spans whole domain
– Field variables are usually shared
• Synchronization required to update shared
variables
– ATOMIC
– CRITICAL
!$OMP ATOMIC
p(i) = p(i) + plocal
usually better than
!$OMP CRITICAL
p(i) = p(i) + plocal
!$OMP END CRITICAL
105. parameter (N=1000)
real A(N,N)
!$OMP THREADPRIVATE(/buf/)
common/buf/lft(N),rht(N)
!$OMP PARALLEL
call init
call scale
call order
!$OMP END PARALLEL
buf is common
within each thread
Global private to thread
• Use threadprivate for all other sub-domain
data that need file scope or common blocks
106. subroutine scale
parameter (N=1000)
!$OMP THREADPRIVATE(/buf/)
common/buf/lft(N),rht(N)
do i=1, N
lft(i)= const* A(i,iam)
end do
return
end
subroutine order
parameter (N=1000)
!$OMP THREADPRIVATE(/buf/)
common/buf/lft(N),rht(N)
do i=1, N
A(i,iam) = lft(index)
end do
return
end
Threadprivate
107. Cache Considerations
• Typical shared memory model: field variables shared
among threads
– Each thread updates field variables for its own subdomain
– Field variable is a shared variable with a size same as the
whole domain
• For fixed problem size scaling subdomain size
decreases
– Fortran Example
– P(20,20) Field variable
1 cache line
128 bytes on Origin
(16 double words)
P(20,20)
s1 s3
s2 s4
108. Two threads need to write the same cache
line
Cacheline pingpongs between the two
processors
•A cacheline is the smallest unit of transport
Poor performance: must avoid at all costs
1 cache line
Thread A writes
Thread B writes
Cache line must move from Processor A to Processor B
False sharing
109. Fixing False sharing
• Pay attention to granularity
• Diagnose by using performance tools if
available
• Pad false shared arrays or make private
110. Stacksize
• OpenMP standard does not specify how much stack
space a thread should have. Default values vary with
operating system and implementation.
• Typically stack size is easy to exhaust.
• Threads that exceed their stack allocation may or may
not seg fault. An application may continue to run while
data is being corrupted.
• To modify:
– limit stacksize unlimited (for csh/tcsh shells)
– ulimit –s unlimited (for bash/ksh shells)
– (Note: here unlimited means increase up to the system
maximum)
111. Reducing Barriers
• Reduce BARRIERs to the minimum
• This synchronization is very often done using
barrier
– But OMP_BARRIER synchronizes all threads in the
region
– BARRIERS have high overhead at large processor
counts
– Synchronizing all threads makes load imbalance
worse
112. Summary
• Loop level paradigm is easiest to implement but
least scalable
• Work-sharing in coarse grain parallel regions
provides intermediate performance
• SPMD model provides good scalability
– OpenMP provides ease of implementation compared
to message passing
• Shared memory model forgiving of use error
– Avoid common pitfalls to obtain good performance
113. Suggestions from a tutorial at
WOMPAT 2002 at ARSC
• Scheduling - use static scheduling unless load balancing is a
real issue. If necessary use static with chunk specified;
dynamic or guided schedule types require too much
overhead.
• Synchronization - Don't use them if at all possible, esp.
'flush'
• Common Problems with OpenMP programs
– Too fine grain parallelism
– Overly synchronized
– Load Imbalance
– True Sharing (ping-ponging in cache)
– False Sharing (again a cache problem)
114. Suggestions from a tutorial at
WOMPAT 2002 at ARSC
• Tuning Keys
– Know your application
– Only use the parallel do/for statements
– Everything else in the language only slows you down,
but they are necessary evils.
• Use a profiler on your code whenever possible. This
will point out bottle necks and other problems.
• Common correctness issues are race conditions and
deadlocks.
115. Suggestions from a tutorial at
WOMPAT 2002 at ARSC
• Remember that reduction clauses need a barrier - so
make sure not to use the 'nowait' clause when a
reduction is being computed.
• Don't use locks if at all possible! If you have to use them,
make sure that you unset them!
• If you think you have a race condition, run the loops
backwards and see if you get the same result.
• Thanks to Tom Logan at ARSC for this report and Tim
Mattson & Sanjiv Shah for the tutorial
116. What’s new in OpenMP 3.0? -
some highlights
• task parallelism is the big news!
– Allows to parallelize irregular problems
• unbounded loops
• recursive algorithms
• producer/consumer
• Loop Parallelism Improvements
– STATIC schedule guarantees
– Loop collapsing
– New AUTO schedule
• OMP_STACK_SIZE environment variable
117. Tasks
• Tasks are work units which execution may be
deferred
– they can also be executed immediately
• Tasks are composed of:
– code to execute
– data environment
– internal control variables (ICV)
118. OpenMP 3.0 Compiler Support
• Intel 11.0: Linux (x86), Windows (x86) and
MacOS (x86)
• Sun Studio Express 11/08: Linux (x86) and
Solaris (SPARC + x86)
• PGI 8.0: Linux (x86) and Windows (x86)
• IBM 10.1: Linux (POWER) and AIX (POWER)
• GCC 4.4 will have support for OpenMP 3.0