24.05.2014 Views

XL Fortran Enterprise Edition for AIX : User's Guide - IBM

XL Fortran Enterprise Edition for AIX : User's Guide - IBM

XL Fortran Enterprise Edition for AIX : User's Guide - IBM

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Typically, parthreshold is set to be equal to the<br />

parallelization overhead. If the computation in<br />

a parallelized loop is very small and the time<br />

taken to execute these loops is spent primarily<br />

in the setting up of parallelization, these loops<br />

should be executed sequentially <strong>for</strong> better<br />

per<strong>for</strong>mance.<br />

seqthreshold=num<br />

profilefreq=num<br />

Specifies the time, in milliseconds, beyond<br />

which a loop that was previously serialized by<br />

the dynamic profiler should revert to being a<br />

parallel loop. The default setting is 5<br />

milliseconds, meaning that if a loop requires<br />

more than 5 milliseconds to execute serially, it<br />

should be parallelized.<br />

seqthreshold acts as the reverse of<br />

parthreshold.<br />

Specifies the frequency with which a loop<br />

should be revisited by the dynamic profiler to<br />

determine its appropriateness <strong>for</strong> parallel or<br />

serial execution. Loops in a program can be<br />

data dependent. The loop that was chosen to<br />

execute serially with a pass of dynamic<br />

profiling may benefit from parallelization in<br />

subsequent executions of the loop, due to<br />

different data input. There<strong>for</strong>e, you need to<br />

examine these loops periodically to reevaluate<br />

the decision to serialize a parallel loop at run<br />

time.<br />

The allowed values <strong>for</strong> this option are the<br />

numbers from 0 to 32. If you set profilefreq to<br />

one of these values, the following results will<br />

occur.<br />

v<br />

v<br />

v<br />

If profilefreq is 0, all profiling is turned off,<br />

regardless of other settings. The overheads<br />

that occur because of profiling will not be<br />

present.<br />

If profilefreq is 1, loops parallelized<br />

automatically by the compiler will be<br />

monitored every time they are executed.<br />

If profilefreq is 2, loops parallelized<br />

automatically by the compiler will be<br />

monitored every other time they are<br />

executed.<br />

v If profilefreq is greater than or equal to 2<br />

but less than or equal to 32, each loop will<br />

be monitored once every nth time it is<br />

executed.<br />

v<br />

If profilefreq is greater than 32, then 32 is<br />

assumed.<br />

It is important to note that dynamic profiling is<br />

not applicable to user-specified parallel loops<br />

Editing, Compiling, Linking, and Running <strong>XL</strong> <strong>Fortran</strong> Programs 63

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!