C H A P T E R 10 |
Parallelization |
This chapter presents an overview of multiprocessor parallelization and describes the capabilities of Fortran 95 on Solaris SPARC and x86 multiprocessor platforms.
See also Techniques for Optimizing Applications: High Performance Computing by Rajat Garg and Ilya Sharapov, a Sun Microsystems BluePrints publication (http://www.sun.com/blueprints/pubs.html)
Parallelizing (or multithreading) an application compiles the program to run on a multiprocessor system or in a multithreaded environment. Parallelization enables a single task, such as a DO loop, to run over multiple processors (or threads) with a potentially significant execution speedup.
Before an application program can be run efficiently on a multiprocessor system like the Ultra 60, Sun Enterprise Server 6500, or Sun Enterprise Server 10000, it needs to be multithreaded. That is, tasks that can be performed in parallel need to be identified and reprogrammed to distribute their computations across multiple processors or threads.
Multithreading an application can be done manually by making appropriate calls to the libthread primitives. However, a significant amount of analysis and reprogramming might be required. (See the Solaris Multithreaded Programming Guide for more information.)
Sun compilers can automatically generate multithreaded object code to run on multiprocessor systems. The Fortran compilers focus on DO loops as the primary language element supporting parallelism. Parallelization distributes the computational work of a loop over several processors without requiring modifications to the Fortran source program.
The choice of which loops to parallelize and how to distribute them can be left entirely up to the compiler (-autopar), specified explicitly by the programmer with source code directives (-explicitpar), or done in combination (-parallel).
Not all loops in a program can be profitably parallelized. Loops containing only a small amount of computational work (compared to the overhead spent starting and synchronizing parallel tasks) may actually run more slowly when parallelized. Also, some loops cannot be safely parallelized at all; they would compute different results when run in parallel due to dependencies between statements or iterations.
Implicit loops (IF loops and Fortran 95 array syntax, for example) as well as explicit DO loops are candidates for automatic parallelization by the Fortran compilers.
f95 can detect loops that might be safely and profitably parallelized automatically. However, in most cases, the analysis is necessarily conservative, due to the concern for possible hidden side effects. (A display of which loops were and were not parallelized can be produced by the -loopinfo option.) By inserting source code directives before loops, you can explicitly influence the analysis, controlling how a specific loop is (or is not) to be parallelized. However, it then becomes your responsibility to ensure that such explicit parallelization of a loop does not lead to incorrect results.
The Fortran 95 compiler provides explicit parallelization by implementing the OpenMP 2.0 Fortran API directives. For legacy programs, f95 also supports the older Sun and Cray style directives. OpenMP has become an informal standard for explicit parallelization in Fortran 95, C, and C++ and is recommended over the older directive styles.
For information on OpenMP, see the OpenMP API User's Guide, or the OpenMP web site at http://www.openmp.org/.
For a discussion of legacy parallelization directives, see Section 10.3.3, Sun-Style Parallelization Directives, and Section 10.3.4, Cray-Style Parallelization Directives.
If you parallelize a program so that it runs over four processors, can you expect it to take (roughly) one fourth the time that it did with a single processor (a fourfold speedup)?
Probably not. It can be shown (by Amdahl's law) that the overall speedup of a program is strictly limited by the fraction of the execution time spent in code running in parallel. This is true no matter how many processors are applied. In fact, if p is the percentage of the total program execution time that runs in parallel mode, the theoretical speedup limit is 100/(100-p); therefore, if only 60% of a program's execution runs in parallel, the maximum increase in speed is 2.5, independent of the number of processors. And with just four processors, the theoretical speedup for this program (assuming maximum efficiency) would be just 1.8 and not 4. With overhead, the actual speedup would be less.
As with any optimization, choice of loops is critical. Parallelizing loops that participate only minimally in the total program execution time has only minimal effect. To be effective, the loops that consume the major part of the runtime must be parallelized. The first step, therefore, is to determine which loops are significant and to start from there.
Problem size also plays an important role in determining the fraction of the program running in parallel and consequently the speedup. Increasing the problem size increases the amount of work done in loops. A triply nested loop could see a cubic increase in work. If the outer loop in the nest is parallelized, a small increase in problem size could contribute to a significant performance improvement (compared to the unparallelized performance).
Here is a very general outline of the steps needed to parallelize an application:
1. Optimize. Use the appropriate set of compiler options to get the best serial performance on a single processor.
2. Profile. Using typical test data, determine the performance profile of the program. Identify the most significant loops.
3. Benchmark. Determine that the serial test results are accurate. Use these results and the performance profile as the benchmark.
4. Parallelize. Use a combination of options and directives to compile and build a parallelized executable.
5. Verify. Run the parallelized program on a single processor and single thread and check results to find instabilities and programming errors that might have crept in. (Set $PARALLEL or $OMP_NUM_THREADS to 1; see Section 10.1.5, Number of Threads).
6. Test. Make various runs on several processors to check results.
7. Benchmark. Make performance measurements with various numbers of processors on a dedicated system. Measure performance changes with changes in problem size (scalability).
8. Repeat steps 4 to 7. Make improvements to your parallelization scheme based on performance.
Not all loops are parallelizable. Running a loop in parallel over a number of processors usually results in iterations executing out of order. Moreover, the multiple processors executing the loop in parallel may interfere with each other whenever there are data dependencies in the loop.
Situations where data dependence issues arise include recurrence, reduction, indirect addressing, and data dependent loop iterations.
You might be able to rewrite a loop to eliminate data dependencies, making it parallelizable. However, extensive restructuring could be needed.
These are general conditions for parallelization. The compilers' automatic parallelization analysis considers additional criteria when deciding whether to parallelize a loop. However, you can use directives to explicitly force loops to be parallelized, even loops that contain inhibitors and produce incorrect results.
Variables that are set in one iteration of a loop and used in a subsequent iteration introduce cross-iteration dependencies, or recurrences. Recurrence in a loop requires that the iterations to be executed in the proper order. For example:
requires the value computed for A(I) in the previous iteration to be used (as A(I-1)) in the current iteration. To produce correct results, iteration I must complete before iteration I+1 can execute.
Reduction operations reduce the elements of an array into a single value. For example, summing the elements of an array into a single variable involves updating that variable in each iteration:
If each processor running this loop in parallel takes some subset of the iterations, the processors will interfere with each other, overwriting the value in SUM. For this to work, each processor must execute the summation one at a time, although the order is not significant.
Certain common reduction operations are recognized and handled as special cases by the compiler.
Loop dependencies can result from stores into arrays that are indexed in the loop by subscripts whose values are not known. For example, indirect addressing could be order dependent if there are repeated values in the index array:
In the example, repeated values in ID cause elements in A to be overwritten. In the serial case, the last store is the final value. In the parallel case, the order is not determined. The values of A(L) that are used, old or updated, are order dependent.
The Sun Studio compilers support the OpenMP parallelization model natively as the primary parallelization model. For information on OpenMP parallelization, see the OpenMP API User's Guide. Sun and Cray-style parallelization refer to legacy applications.
The PARALLEL (or OMP_NUM_THREADS) environment variable controls the maximum number of threads available to the program. Setting the environment variable tells the runtime system the maximum number of threads the program can use. The default is 1. In general, set the PARALLEL or OMP_NUM_THREADS variable to the available number of processors on the target platform.
The following example shows how to set it:
In this example, setting PARALLEL to four enables the execution of a program using at most four threads. If the target machine has four processors available, the threads will map to independent processors. If there are fewer than four processors available, some threads could run on the same processor as others, possibly degrading performance.
The SunOS operating system command psrinfo(1M) displays a list of the processors available on a system:
demo% psrinfo 0 on-line since 03/18/99 15:51:03 1 on-line since 03/18/99 15:51:03 2 on-line since 03/18/99 15:51:03 3 on-line since 03/18/99 15:51:03 |
The executing program maintains a main memory stack for the initial thread executing the program, as well as distinct stacks for each helper thread. Stacks are temporary memory address spaces used to hold arguments and AUTOMATIC variables over subprogram invocations.
The default size of the main stack is about 8 megabytes. The Fortran compilers normally allocate local variables and arrays as STATIC (not on the stack). However, the -stackvar option forces the allocation of all local variables and arrays on the stack (as if they were AUTOMATIC variables). Use of -stackvar is recommended with parallelization because it improves the optimizer's ability to parallelize subprogram calls in loops. -stackvar is required with explicitly parallelized loops containing subprogram calls. (See the discussion of -stackvar in the Fortran User's Guide.)
Using the C shell (csh), the limit command displays the current main stack size as well as sets it:
With Bourne or Korn shells, the corresponding command is ulimit:
Each helper thread of a multithreaded program has its own thread stack. This stack mimics the initial thread stack but is unique to the thread. The thread's PRIVATE arrays and variables (local to the thread) are allocated on the thread stack. The default size is 8 megabytes on SPARC V9 (UltraSPARC) and 64-bit x86 platforms, 4 megabytes otherwise. The size is set with the STACKSIZE environment variable:
demo% setenv STACKSIZE 8192 <- Set thread stack size to 8 Mb C shell -or- demo$ STACKSIZE=8192 Bourne/Korn Shell demo$ export STACKSIZE |
Setting the thread stack size to a value larger than the default may be necessary for some parallelized Fortran codes. However, it may not be possible to know just how large it should be, except by trial and error, especially if private/local arrays are involved. If the stack size is too small for a thread to run, the program will abort with a segmentation fault.
With the -autopar and -parallel options, the f95 compiler automatically finds DO loops that can be parallelized effectively. These loops are then transformed to distribute their iterations evenly over the available processors. The compiler generates the thread calls needed to make this happen.
The compiler's dependency analysis transforms a DO loop into a parallelizable task. The compiler may restructure the loop to split out unparallelizable sections that will run serially. It then distributes the work evenly over the available processors. Each processor executes a different chunk of iterations.
For example, with four CPUs and a parallelized loop with 1000 iterations, each thread would execute a chunk of 250 iterations:
Only loops that do not depend on the order in which the computations are performed can be successfully parallelized. The compiler's dependence analysis rejects from parallelization those loops with inherent data dependencies. If it cannot fully determine the data flow in a loop, the compiler acts conservatively and does not parallelize. Also, it may choose not to parallelize a loop if it determines the performance gain does not justify the overhead.
Note that the compiler always chooses to parallelize loops using a static loop scheduling--simply dividing the work in the loop into equal blocks of iterations. Other scheduling schemes may be specified using explicit parallelization directives described later in this chapter.
A few definitions, from the point of view of automatic parallelization, are needed:
Both m and a are array variables; s is pure scalar. The variables u, x, z, and px are scalar variables, but not pure scalars.
DO loops that have no cross-iteration data dependencies are automatically parallelized by -autopar or -parallel. The general criteria for automatic parallelization are:
The compilers may automatically eliminate a reference that appears to create a data dependence in the loop. One of the many such transformations makes use of private versions of some of the arrays. Typically, the compiler does this if it can determine that such arrays are used in the original loops only as temporary storage.
Example: Using -autopar, with dependencies eliminated by private arrays:
parameter (n=1000) real a(n), b(n), c(n,n) do i = 1, 1000 <--Parallelized do k = 1, n a(k) = b(k) + 2.0 end do do j = 1, n-1 c(i,j) = a(j+1) + 2.3 end do end do end |
In the example, the outer loop is parallelized and run on independent processors. Although the inner loop references to array a appear to result in a data dependence, the compiler generates temporary private copies of the array to make the outer loop iterations independent.
Under automatic parallelization, the compilers do not parallelize a loop if:
In a multithreaded, multiprocessor environment, it is most effective to parallelize the outermost loop in a loop nest, rather than the innermost. Because parallel processing typically involves relatively large loop overhead, parallelizing the outermost loop minimizes the overhead and maximizes the work done for each thread. Under automatic parallelization, the compilers start their loop analysis from the outermost loop in a nest and work inward until a parallelizable loop is found. Once a loop within the nest is parallelized, loops contained within the parallel loop are passed over.
A computation that transforms an array into a scalar is called a reduction operation. Typical reduction operations are the sum or product of the elements of a vector. Reduction operations violate the criterion that calculations within a loop not change a scalar variable in a cumulative way across iterations.
Example: Reduction summation of the elements of a vector:
However, for some operations, if reduction is the only factor that prevents parallelization, it is still possible to parallelize the loop. Common reduction operations occur so frequently that the compilers are capable of recognizing and parallelizing them as special cases.
Recognition of reduction operations is not included in the automatic parallelization analysis unless the -reduction compiler option is specified along with -autopar or -parallel.
If a parallelizable loop contains one of the reduction operations listed in TABLE 10-2, the compiler will parallelize it if -reduction is specified.
The following table lists the reduction operations that are recognized by the compiler.
All forms of the MIN and MAX function are recognized.
Floating-point sum or product reduction operations may be inaccurate due to the following conditions:
In some situations, the error may not be acceptable.
Example: Roundoff, get the sum of 100,000 random numbers between -1 and +1:
Results vary with the number of processors. The following table shows the sum of 100,000 random numbers between -1 and +1.
In this situation, roundoff error on the order of 10-14 is acceptable for data that is random to begin with. For more information, see the Sun Numerical Computation Guide.
This section describes the source code directives recognized by f95 to explicitly indicate which loops to parallelize and what strategy to use.
The Fortran 95 compiler now supports the OpenMP Fortran API as the primary parallelization model. See the OpenMP API User's Guide for additional information..
f95 will also accept legacy Sun-style and Cray-style parallelization directives to facilitate porting explicitly parallelized programs from other platforms.
Explicit parallelization of a program requires prior analysis and deep understanding of the application code as well as the concepts of shared-memory parallelization.
DO loops are marked for parallelization by directives placed immediately before them. Compile with -openmp to enable recognition of OpenMP Fortran 95 directives and generation of parallelized DO loop code. (Compile with -parallel or -explicitpar for legacy Sun or Cray directives.) Parallelization directives are comment lines that tell the compiler to parallelize (or not to parallelize) the DO loop that follows the directive. Directives are also called pragmas.
Take care when choosing which loops to mark for parallelization. The compiler generates threaded, parallel code for all loops marked with parallelization directives, even if there are data dependencies that will cause the loop to compute incorrect results when run in parallel.
If you do your own multithreaded coding using the libthread primitives, do not use any of the compilers' parallelization options--the compilers cannot parallelize code that has already been parallelized with user calls to the threads library.
A loop is appropriate for explicit parallelization if:
A private variable or array is private to a single iteration of a loop. The value assigned to a private variable or array in one iteration is not propagated to any other iteration of the loop.
A shared variable or array is shared with all other iterations. The value assigned to a shared variable or array in an iteration is seen by other iterations of the loop.
If an explicitly parallelized loop contains shared references, then you must ensure that sharing does not cause correctness problems. The compiler does not synchronize on updates or accesses to shared variables.
If you specify a variable as private in one loop, and its only initialization is within some other loop, the value of that variable may be left undefined in the loop.
A subprogram call in a loop (or in any subprograms called from within the called routine) may introduce data dependencies that could go unnoticed without a deep analysis of the data and control flow through the chain of calls. While it is best to parallelize outermost loops that do a significant amount of the work, these tend to be the very loops that involve subprogram calls.
Because such an interprocedural analysis is difficult and could greatly increase compilation time, automatic parallelization modes do not attempt it. With explicit parallelization, the compiler generates parallelized code for a loop marked with a PARALLEL DO or DOALL directive even if it contains calls to subprograms. It is still the programmer's responsibility to insure that no data dependencies exist within the loop and all that the loop encloses, including called subprograms.
Multiple invocations of a routine by different threads can cause problems resulting from references to local static variables that interfere with each other. Making all the local variables in a routine automatic rather than static prevents this. Each invocation of a subprogram then has its own unique store of local variables maintained on the stack, and no two invocations will interfere with each other.
Local subprogram variables can be made automatic variables that reside on the stack either by listing them on an AUTOMATIC statement or by compiling the subprogram with the -stackvar option. However, local variables initialized in DATA statements must be rewritten to be initialized in actual assignments.
Note - Allocating local variables to the stack can cause stack overflow. See Section 10.1.6, Stacks, Stack Sizes, and Parallelization about increasing the size of the stack. |
In general, the compiler parallelizes a loop if you explicitly direct it to. There are exceptions--some loops the compiler will not parallelize.
The following are the primary detectable inhibitors that might prevent explicitly parallelizing a DO loop:
This exception holds for indirect nesting, too. If you explicitly parallelize a loop that includes a call to a subroutine, then even if you request the compiler to parallelize loops in that subroutine, those loops are not run in parallel at runtime.
By compiling with -vpara and -loopinfo, you will get diagnostic messages if the compiler detects a problem while explicitly parallelizing a loop.
The following table lists typical parallelization problems detected by the compiler:
... !$OMP PARALLEL DO do 900 i = 1, 1000 ! Parallelized (outer loop) do 200 j = 1, 1000 ! Not parallelized, no warning ... 200 continue 900 continue ... |
Example: A parallelized loop in a subroutine:
In the example, the loop within the subroutine is not parallelized because the subroutine itself is run in parallel.
Example: Jumping out of a loop:
!$omp parallel do do i = 1, 1000 ! Not parallelized, error issued ... if (a(i) .gt. min_threshold ) go to 20 ... end do 20 continue ... |
The compiler issues an error diagnostic if there is a jump outside a loop marked for parallelization.
Example: A variable in a loop has a loop-carried dependency:
Here the loop is parallelized but the possible loop carried dependency is diagnosed in a warning. However, be aware that not all loop dependencies can be diagnosed by the compiler.
You can do I/O in a loop that executes in parallel, provided that:
Example: I/O statement in loop
However, I/O that is recursive, where an I/O statement contains a call to a function that itself does I/O, will cause a runtime error.
OpenMP is a parallel programming model for multi-processor platforms that is becoming standard programming practice for Fortran 95, C, and C++ applications. It is the preferred parallel programming model for Sun Studio compilers.
To enable OpenMP directives, compile with the -openmp option flag. Fortran 95 OpenMP directives are identified with the comment-like sentinel !$OMP followed by the directive name and subordinate clauses.
The !$OMP PARALLEL directive identifies the parallel regions in a program. The !$OMP DO directive identifies DO loops within a parallel region that are to be parallelized. These directives can be combined into a single !$OMP PARALLEL DO directive that must be placed immediately before the DO loop.
The OpenMP specification includes a number of directives for sharing and synchronizing work in a parallel region of a program, and subordinate clauses for data scoping and control.
One major difference between OpenMP and legacy Sun-style directives is that OpenMP requires explicit data scoping as either private or shared.
For more information, including guidelines for converting legacy programs using Sun and Cray parallelization directives, see the OpenMP API User's Guide.
Legacy Sun-style directives are enabled by default (or with the -mp=sun option) when compiling with the -explicitpar or -parallel options.
A parallel directive consists of one or more directive lines. A Sun-style directive line is defined as follows:
C$PAR Directive [ Qualifiers ] <- Initial directive line C$PAR& [More_Qualifiers] <- Optional continuation lines |
The Sun-style parallel directives are:
Examples of Sun-style parallel directives:
The TASKCOMMON directive declares variables in a global COMMON block as thread-private: Every variable declared in a common block becomes a private variable to the thread, but remains global within the thread. Only named COMMON blocks can be declared TASKCOMMON.
The syntax of the directive is:
C$PAR TASKCOMMON common_block_name
The directive must appear immediately after every COMMON declaration for that named block.
This directive is effective only when compiled with -explicitpar or -parallel. Otherwise, the directive is ignored and the block is treated as a regular COMMON block.
Variables declared in TASKCOMMON blocks are treated as thread-private variables in all the DOALL loops and routines called from within the DOALL loops. Each thread gets its own copy of the COMMON block, so data written by one thread is not directly visible to other threads. During serial portions of the program, accesses are to the initial thread's copy of the COMMON block.
Variables in TASKCOMMON blocks should not appear on any DOALL qualifiers, such as PRIVATE, SHARED, READONLY, and so on.
It is an error to declare a common block as task common in some but not all compilation units where the block is defined. A check at runtime for task common consistency can be enabled by compiling the program with the -xcommonchk=yes flag. Enable the runtime check only during program development, as it can degrade performance.
The DOALL directive requests the compiler to generate parallel code for the one DO loop immediately following it (if compiled with the -parallel or -explicitpar options).
Note - Analysis and transformation of reduction operations is not performed within explicitly parallelized loops. |
Example: Explicit parallelization of a loop:
demo% cat t4.f ... C$PAR DOALL do i = 1, n a(i) = b(i) * c(i) end do do k = 1, m x(k) = x(k) * z(k,k) end do ... demo% f95 -explicitpar t4.f |
All qualifiers on the Sun-style DOALL directive are optional. The following table summarizes them:
The PRIVATE(varlist)qualifier specifies that all scalars and arrays in the list varlist are private for the DOALL loop. Both arrays and scalars can be specified as private. In the case of an array, each thread of the DOALL loop gets a copy of the entire array. All other scalars and arrays referenced in the DOALL loop, but not contained in the private list, conform to their appropriate default scoping rules. (See Section 10.3.1.1, Scoping Rules: Private and Shared).
Example: Specify array a private in loop i:
C$PAR DOALL PRIVATE(a) do i = 1, n a(1) = b(i) do j = 2, n a(j) = a(j-1) + b(j) * c(j) end do x(i) = f(a) end do |
The SHARED(varlist) qualifier specifies that all scalars and arrays in the list varlist are shared for the DOALL loop. Both arrays and scalars can be specified as shared. Shared scalars and arrays can be accessed in all the iterations of a DOALL loop. All other scalars and arrays referenced in the DOALL loop, but not contained in the shared list, conform to their appropriate default scoping rules.
Example: Specify a shared variable:
In the example, the variable y has been specified as a variable whose value should be shared among the iterations of the i loop.
The READONLY(varlist) qualifier specifies that all scalars and arrays in the list varlist are read-only for the DOALL loop. Read-only scalars and arrays are a special class of shared scalars and arrays that are not modified in any iteration of the DOALL loop. Specifying scalars and arrays as READONLY indicates to the compiler that it does not need to use a separate copy of that scalar variable or array for each thread of the DOALL loop.
Example: Specify a read-only variable:
In the preceding example, x is a shared variable, but the compiler can rely on the fact that its value will not be modified in any iteration of the i loop because of its READONLY specification.
A STOREBACK scalar variable or array is one whose value is computed in a DOALL loop. The computed value can be used after the termination of the loop. In other words, the last loop iteration values of storeback scalars or arrays are visible after the DOALL loop.
Example: Specify the loop index variable as storeback:
In the preceding example, both the variables x and i are storeback variables, even though both variables are private to the i loop. The value of i after the loop is n+1, while the value of x is whatever value it had at the end of the last iteration.
There are some potential problems for STOREBACK to be aware of.
The STOREBACK operation occurs at the last iteration of the explicitly parallelized loop, even if this is not the same iteration that last updates the value of the STOREBACK variable or array.
Example: STOREBACK variable potentially different from the serial version:
In the preceding example, the value of the STOREBACK variable x that is printed out might not be the same as that printed out by a serial version of the i loop. In the explicitly parallelized case, the processor that processes the last iteration of the i loop (when i = n) and performs the STOREBACK operation for x, might not be the same processor that currently contains the last updated value of x. The compiler issues a warning message about these potential problems.
The SAVELAST qualifier specifies that all private scalars and arrays are STOREBACK variables for the DOALL loop.
In the example, variables x, y, and i are STOREBACK variables.
The REDUCTION(varlist) qualifier specifies that all variables in the list varlist are reduction variables for the DOALL loop. A reduction variable (or array) is one whose partial values can be individually computed on various processors, and whose final value can be computed from all its partial values.
The presence of a list of reduction variables requests the compiler to handle a DOALL loop as reduction loop by generating parallel reduction code for it.
Example: Specify a reduction variable:
In the preceding example, the variable x is a (sum) reduction variable; the i loop is a (sum) reduction loop.
SCHEDTYPE(t) specifies the scheduling type t be used to schedule the DOALL loop.
Qualifiers can appear multiple times with cumulative effect. In the case of conflicting qualifiers, the compiler issues a warning message, and the qualifier appearing last prevails.
Example: A three-line Sun-style directive (note conflicting MAXCPUS, SHARED, and PRIVATE qualifiers):
C$PAR DOALL MAXCPUS(4), READONLY(S), PRIVATE(A,B,X), MAXCPUS(2) C$PAR DOALL SHARED(B,X,Y), PRIVATE(Y,Z) C$PAR DOALL READONLY(T) |
Example: A one-line equivalent of the preceding three lines:
The DOSERIAL directive disables parallelization of the specified loop. This directive applies to the one loop immediately following it.
Example: Exclude one loop from parallelization:
In the example, when compiling with -parallel, the j loop will not be parallelized by the compiler, but the i or k loop may be.
The DOSERIAL* directive disables parallelization of the specified nest of loops. This directive applies to the whole nest of loops immediately following it.
Example: Exclude a whole nest of loops from parallelization:
In the example, when compiling with -parallel, the j and k loops will not be parallelized by the compiler, but the i loop may be.
If both DOSERIAL* and DOALL are specified for the same loop, the last one prevails.
Example: Specifying both DOSERIAL* and DOALL:
In the example, the i loop is not parallelized, but the j loop is.
Also, the scope of the DOSERIAL* directive does not extend beyond the textual loop nest immediately following it. The directive is limited to the same function or subroutine that it appears in.
Example: DOSERIAL* does not extend to a loop in a called subroutine:
program caller common /block/ a(10,10) C$PAR DOSERIAL* do i = 1, 10 call callee(i) end do end subroutine callee(k) common /block/a(10,10) do j = 1, 10 a(j,k) = j + k end do return end |
In the preceding example, DOSERIAL* applies only to the i loop and not to the j loop, regardless of whether the call to the subroutine callee is inlined.
For Sun-style (C$PAR) explicit directives, the compiler uses default rules to determine whether a scalar or array is shared or private. You can override the default rules to specify the attributes of scalars or arrays referenced inside a loop. (With Cray-style !MIC$ directives, all variables that appear in the loop must be explicitly declared either shared or private on the DOALL directive.)
The compiler applies these default rules:
If inter-iteration dependencies exist in a loop, then the execution may result in erroneous results. You must ensure that these cases do not arise. The compiler may sometimes be able to detect such a situation at compile time and issue a warning, but it does not disable parallelization of such loops.
Example: Potential problem through equivalence:
In the example, since the scalar variable y has been equivalenced to a(1), we have a conflict with y as private and a(:) as shared by default, leading to possibly erroneous results when the parallelized i loop is executed. No diagnostic is issued in these situations.
You can fix the example by using C$PAR DOALL PRIVATE(y).
To use legacy Cray-style parallelization directives, you must compile with -mp=cray.
Mixing program units compiled with both Sun and Cray directives can produce incorrect results.
A major difference between Sun and Cray directives is that Cray style requires explicit scoping of every scalar and array in the loop as either SHARED or PRIVATE, unless AUTOSCOPE is specified.
The following table shows Cray-style directive syntax.
A parallel directive consists of one or more directive lines. A directive line is defined with the same syntax as Sun-style (see Section 10.3.3, Sun-Style Parallelization Directives), except:
The Cray directives are similar to Sun-style:
For Cray-style DOALL, the PRIVATE qualifier is required. Each variable within the DO loop must be qualified as private or shared, and the DO loop index must always be private. The following table summarizes available Cray-style qualifiers.
Specifying AUTOSCOPE directs the compiler to use the following rules to determine the scoping of a variable or array not explicitly scoped as PRIVATE or SHARED.
For a variable or array to be SHARED, any of the following must be true:
For a variable or array to be PRIVATE, the following must be true:
Still, AUTOSCOPE cannot always determine the scope of variables or arrays at compile time. Conditional paths through the loop, among other things, can alter the scoping in ways that cannot be determined by the compiler. It is much safer to scope variables explicitly with PRIVATE and SHARED qualifiers.
For Cray-style directives, the DOALL directive allows a single scheduling qualifier, for example, !MIC$& CHUNKSIZE(100). TABLE 10-7 shows the Cray-style DOALL directive scheduling qualifiers:
The default scheduling type (when no scheduling type is specified on a Cray-style DOALL directive) is the Sun-style STATIC, for which there is no Cray-style equivalent.
There are a number of environment variables used with parallelization:
Sets the number of threads to use during execution of a parallel region. This number can be overriden by a NUM_THREADS clause, or a call to OMP_SET_NUM_THREADS(). If not set, a default of 1 is used. value is a positive integer. For compatibility with legacy programs, setting the PARALLEL environment variable has the same effect as setting OMP_NUM_THREADS. However, if they are both set to different values, the runtime library will issue an error message. |
|
Controls warning messages issued by the OpenMP runtime library. If set to TRUE the runtime library issues warning messages to stderr; FALSE disables warning messages. The default is FALSE.. |
|
Controls the end-of-task status of each helper thread executing the parallel part of a program. You can set the value to SPIN, SLEEP ns, or SLEEP nms. The default is SLEEP -- the thread sleeps after completing a parallel task until a new parallel task arrives. SLEEP with no argument puts the thread to sleep immediately after completing a parallel task. SLEEP, SLEEP (0), SLEEP (0s), and SLEEP (0ms) are all equivalent. |
|
The SUNW_MP_PROCBIND environment variable can be used to bind LWPs (lightweight processes) of an OpenMP program to processors. Performance can be enhanced with processor binding, but performance degradation will occur if multiple LWPs are bound to the same processor. |
|
Sets the stack size for each thread. The value is in kilobytes. The default thread stack sizes are 4 Mb on 32-bit SPARC V8 and x86 platforms, and 8 Mb on 64-bit SPARC V9 and x86 platforms. setenv STACKSIZE 8192
See also Section 10.1.6, Stacks, Stack Sizes, and Parallelization. |
There is more information about environment variables in the OpenMP API User's Guide.
Debugging parallelized programs requires some extra effort. The following schemes suggest ways to approach this task.
There are some steps you can try immediately to determine the cause of errors.
You can do one of the following:
If the problem disappears, then you can assume it was due to using multiple threads.
If you are using the -reduction option, summation reduction may be occurring and yielding slightly different answers. Try running without this option.
If you have many subroutines in your program, use fsplit(1) to break them into separate files. Then compile some files with and without -parallel, and use f95 to link the .o files. You must specify -parallel on this link step.
Execute the binary and verify results.
Repeat this process until the problem is narrowed down to one subroutine.
Check which loops are being parallelized and which loops are not.
Create a dummy subroutine or function that does nothing. Put calls to this subroutine in a few of the loops that are being parallelized. Recompile and execute. Use -loopinfo to see which loops are being parallelized.
Continue this process until you start getting the correct results.
Add the C$PAR DOALL directive to a couple of the loops that are being parallelized. Compile with -explicitpar, then execute and verify the results. Use -loopinfo to see which loops are being parallelized. This method permits the addition of I/O statements to the parallelized loop.
Repeat this process until you find the loop that causes the wrong results.
Note: if you need -explicitpar only (without -autopar), do not compile with -explicitpar and -depend. This method is the same as compiling with -parallel, which, of course, includes -autopar.
Replace DO I=1,N with DO I=N,1,-1. Different results point to data dependencies.
To use dbx on a parallel loop, temporarily rewrite the program as follows:
Example: Manually transform a loop to allow using dbx in parallel:
The following provide more information:
Copyright © 2004, Sun Microsystems, Inc. All Rights Reserved.