Intel Parallel Studio timing inconsistencies -

-

i have code uses intel tbb , i'm running on 32 core machine. in code, use

parallel_for(blocked_range (2,left_image_width-2, left_image_width /32) ...

to spawn 32 threads concurrent work, there no race conditions , each thread given same amount of work. i'm using clock_t measure how long program takes. image, takes 19 seconds complete.

then ran code through intel parallel studio , ran code in 2 seconds. result expecting can't figure out why there's such large difference between two. time_t sum clock cycles on cores? doesn't make sense. below snippet in question.

clock_t begin=clock();  create_threads_and_do_work();  clock_t end=clock(); double diffticks=end-begin; double diffms=(diffticks*1000)/clocks_per_sec; cout<<"and time "<<diffms<<" ms"<<endl;

any advice appreciated.

it's isn't quite clear if difference in run time result of 2 different inputs (images) or 2 different run-time measuring methods (clock_t difference vs. intel software measurement). furthermore, aren't showing goes on in create_threads_and_do_work(), , didn't mention tool within intel parallel studio using, vtune?

your clock_t difference method sum processing time of thread called (the main thread in example), might not count processing time of threads spawned within create_threads_and_do_work(). whether or doesn't depends on whether within function wait threads complete , exit function or if spawn threads , exit (before complete processing). if in function parallel_for(), clock_t difference should yield right result , should no different other run-time measurements.

within intel parallel studio there profiling tool called vtune. powerful tool , when run program through can view (in graphically pleasing way) processing time (as times called) of each function in code. i'm pretty sure after doing you'll figure out.

one last idea - did program complete course when using intel software? i'm asking because vtune collect data time , stop without allowing program complete.


Comments

Post a Comment

Popular posts from this blog

php - What is the difference between $_SERVER['PATH_INFO'] and $_SERVER['ORIG_PATH_INFO']? -

-
what’s difference between $_server['path_info'] , $_server['orig_path_info'] ? how use them? when run print_r($_server) , path_info , orig_path_info not present in array. why not? how can enable them? i have read php manual on them, still don’t understand them. the path_info variable present if invoke php script this: http://www.example.com/phpinfo.php/hello_there it's /hello_there part after .php script. if don't invoke url that, there won't $_server["path_info"] environment variable. the porig_ prefix uncommon. path_info standard cgi-environment variable, , should never prefixed. did read that? (there issues around php3/php4 if invoked php interpreter via cgi-bin/ - hardly has such setups today.) for reference: http://www.ietf.org/rfc/rfc3875
Read more

fortran - Function return type mismatch -

-
i'm attempting recode old c++ program in fortran make use of lapack (i'm aware c++ have lapack++, i'm having lot of trouble installing it, gave up). i didn't have problems compilation, when had variables declared real . when started coding section of program required lapack, found parameters passed dsyev need double precision . tried change double precision (including changing hard coded numbers double precision counterparts, i.e. 0.0 -> 0.0d0) when try compile, following error of functions , subroutines i've written: error: return type mismatch of function <function> @ (1) (real(4)/real(8)) i'm not sure coming from, in program has been changed double precision. for example, i've declared following: double precision :: alpha(3),d(3),zeta1,zeta2 double precision :: a1(3),a2(3),d1(3),d2(3) double precision :: pi pi = 3.14159265359d0 alpha = (/0.109818d0, 0.405771d0, 2.22766d0/) d = (/0.444635d0, 0.535328d0, 0.154329d0 /) 10 i=1,3 ...
Read more

queue - mq_receive: message too long -

-
i implementing communication between 2 processes using queue. problem when call function mq_receive, error: message long. i have done following: struct mq_attr attr; long size = attr.mq_msgsize; .... // initializing queue "/gateway" int rc = mq_receive(gateway, buffer, size, &prio); if print size value, size=1, while when print same size program (got same mechanism), not long integer ( -1217186280 )... how can solve error?....so while size = 1, believe it's right "message long" why 1? p.s. have tried put : int rc = mq_receive(gateway, buffer, sizeof(buffer), &prio); but no result. it seems need read docs more carefully. when call mq_receive should pass size of destination buffer. this size must greater mq_msgsize attribute of queue . in addition, seems have error in queue attributes initialisation makes proper mq_receive call impossible. here standard message queue session: fill mq_attr struct ( doc ): struct mq_a...
Read more