parallel processing - OpenMP not work with an array access indirect -

-

i'm trying optimize loop in parallel way openmp.

  double precision, intent(inout) :: x(13541)   integer         , intent(in)    :: nelem,nelmax !nelmax = 25996    integer         , intent(in)    :: ikle(nelmax)   double precision, intent(in)    :: w(nelmax) !$omp parallel private(ielem) reduction(+:x)   ielem = 1 , nelem     x(ikle(ielem)) = x(ikle(ielem)) + w(ielem)   enddo !$omp end parallel

for , j different, it's possible ikle(i)=ikle(j)

increasing number of threads, found takes longer run loop. use omp_get_wtime() time job.

1 t  0.21943306922912598      2 t  0.30610203742980957      3 t  0.43688893318176270      4 t  0.53783702850341797      5 t  0.61055016517639160      6 t  0.96715998649597168      7 t  0.89582014083862305      8 t   1.3073339462280273

i think problem caused array access indirect don't know how deal in openmp

if it's possible ikle(i)=ikle(j) have worse problem irregular access trashing hopes of efficient use of cached data, have data race -- there no protection in code against multiple threads writing same location 'simultaneously'. depending on frequency of occurrence of ikle(i)=ikle(j) may lucky , never experience race in practice. or may unlucky. either way, written, program wrong 'un.

fwiw agree suspicion irregular pattern of access elements of x root of timing peculiarity have reported, requires lot more movement of data through caches.

also, while i'm writing, line

  integer, intent(in) :: nelem,nelmax = 25996

is wrong too, it's syntax error try initialise routine argument.

edit, in response op's comment:

the dependence surely problem -- not way of thinking, dependence (by take mean call data race) makes program wrong, broken, erroneous. poor scaling inconvenience. problem arises in openmp version of program because two, or more, threads may try update same element of x @ same time. update not atomic, various operations go on behind scenes (read data register, add values in 2 registers together, write data memory location, sort of thing) may interleaved in way, , of interleavings not leave value in x have been in sequential execution of program.

would give hints movement of data said? not @ time, i've written enough.


Comments

Post a Comment

Popular posts from this blog

yii2 - Yii 2 Running a Cron in the basic template -

-
i have seen instructions on running cron in advanced template, cant figure out how on basic template. have following controller in basic controllers folder. <?php namespace app\controllers; use yii\console\controller; /** * cron controller */ class croncontroller extends controller { public function actionindex() { echo "cron service runnning"; } public function actionmail($to) { echo "sending mail " . $to; } } i have navigated root of application , tried these comands yii cron php yii cron im getting unknown comand "cron" i know old question failed find actual answer this. if @ code yii exec file you'll notice requires /common/config/aliases.php file console/config/main.php. can naturally change these or copy on these advanced template. hope solves problem wanting same thing.
Read more

asp.net - 'System.Web.HttpContext' does not contain a definition for 'GetOwinContext' Mystery -

-
i realize question might seem trivial some, it's these types of things find myself fighting quite bit , want make sense of despite seeming losing battle in .net (for me anyway). so, if following: using system.web; ... applicationuser user = system.web.httpcontext.getowincontext().getusermanager<applicationusermanager>().findbyid(system.web.httpcontext.user.identity.getuserid()); that produces error in title , red getowincontext() , error cannot resolve symbol 'getowincontext()' however, if following (remove system.web in front of httpcontext ), works expected (or @ least no errors): using system.web; ... applicationuser user = httpcontext.getowincontext().getusermanager<applicationusermanager>().findbyid(system.web.httpcontext.user.identity.getuserid()); however, if (same line that's working using system.web commented out): //using system.web; ... applicationuser user = httpcontext.getowincontext().getusermana...
Read more

mercurial graft feature, can it copy? -

-
there new feature in mercurial 3, called 'graft' (graft local). job of moving change-set different branch. is there way "copy" change-set branch? i interested in in moving stuff qa branch production branch, still need code reside in qa. want copy. found , have tried few things, maybe it'll become clear more tinkering , reading... does have better way of 'cherry picking' change-sets move cross branch (or trunk)??? graft copies changesets, doesn't move them. hg graft (emphasis mine): this command uses mercurial's merge logic copy individual changes other branches without merging branches in history graph. known 'backporting' or 'cherry-picking' . default, graft copy user, date, , description source changesets.
Read more