arrays - Trying to pass MPI derived types between processors (and failing) -


i trying parallelize customer's fortran code mpi. f array of 4-byte reals dimensioned f(dimx,dimy,dimz,dimf). need various processes work on different parts of array's first dimension. (i have rather started last, wasn't me.) define derived type mpi_x_inteface so

call mpi_type_vector(dimy*dimz*dimf, 1, dimx, mpi_real,  &                      mpi_x_interface, mpi_err) call mpi_type_commit(mpi_x_interface, mpi_err)

my intent single mpi_x_interface contain of data in 'f' @ given first index "i". is, given i, should contain f(i,:,:,:). (note @ stage of game, procs have complete copy of f. intend split f between procs, except want proc 0 have full copy purpose of gathering.)

ptsinproc array containing number of "i" indices handled each proc. x_slab_displs displacement beginning of array each proc. 2 procs, testing on, ptsinproc=(/61,60/), x_slab_displs=(/0,61/). myminpt simple integer giving minimum index handled in each proc.

so want gather of f proc 0 , run

    if (myrank == 0)       call mpi_gatherv(mpi_in_place, ptsinproc(myrank),  +                     mpi_x_interface, f(1,1,1,1), ptsinproc,  +                     x_slab_displs, mpi_x_interface, 0,  +                     mpi_comm_world, mpi_err)     else       call mpi_gatherv(f(myminpt,1,1,1), ptsinproc(myrank),  +                     mpi_x_interface, f(1,1,1,1), ptsinproc,  +                     x_slab_displs, mpi_x_interface, 0,  +                     mpi_comm_world, mpi_err)     endif

i can send @ 1 "slab" this. if try send entire 60 "slabs" proc 1 proc 0 seg fault due "invalid memory reference". btw, when send single slab, data winds in wrong places.

i've checked obvious stuff maiking sure myrank , ptsinproc , x_slab_dislps should on procs. i've looked difference between "size" , "extent" , on, no avail. i'm @ wit's end. don't see doing wrong. , might remember asked similar (but different!) question few months back. admit i'm not getting it. patience appreciated.

first off, want reason you're running many problems because trying split first (fastest) axis. not recommended @ because as-is packing mpi_x_interface requires lot of non-contiguous memory accesses. we're talking huge loss in performance.

splitting slowest axis across mpi processes better strategy. highly recommend transposing 4d matrix x axis last if can.

now actual problem(s)...

derived datatypes

as have deduced, 1 problem size , extent of derived datatype might incorrect. let's simplify problem bit can draw picture. dimy*dimz*dimf=3, , dimx=4. as-is, datatype mpi_x_interface describes following data in memory:

| x |   |   |   | x |   |   |   | x |   |   |   |

that is, every 4th mpi_real, , 3 of them total. seeing want, far good: size of variable correct. however, if try , send "the next" mpi_x_interface, see implementation of mpi start @ next point in memory (which in case has not been allocated), , throw "invalid memory access" @ you:

                                             tries access , bombs                                                  vvv | x |   |   |   | x |   |   |   | x |   |   |   | y |   |   |   | y | ...

what need tell mpi part of datatype "the next" mpi_x_interface starts 1 real array. accomplished redefining "extent" of derived datatype calling mpi_type_create_resized(). in case, need write

integer :: mpi_x_interface, mpi_x_interface_resized integer, parameter :: sizeof_real = 4 ! or whatever f  call mpi_type_vector(dimy*dimz*dimf, 1, dimx, mpi_real,  &                  mpi_x_interface, mpi_err) call mpi_type_create_resized(mpi_x_interface, 0, 1*sizeof_real, &                              mpi_x_interface_resized, mpi_err) call mpi_type_commit(mpi_x_interface_resized, mpi_err)

then, calling "the next" 3 mpi_x_interface_resized result in:

| x | y | z | | x | y | z | | x | y | z | |

as expected.

mpi_gatherv

note have correctly defined extent of datatype, calling mpi_gatherv offset in terms of datatype should work expected.

personally, wouldn't think there need try fancy logic mpi_in_place collective operation. can set myminpt=1 on myrank==0. can call on every rank:

   call mpi_gatherv(f(myminpt,1,1,1), ptsinproc(myrank), +                     mpi_x_interface_resized, f, ptsinproc, +                     x_slab_displs, mpi_x_interface_resized, 0, +                     mpi_comm_world, mpi_err)

Comments

Post a Comment

Popular posts from this blog

java - pagination of xlsx file to XSSFworkbook using apache POI -

right in code, reading xlsx file, xssfworkbook, , writing database. but, when size of xlsx file increases, causes outofmemory error. can not increase server size, or divide xlsx file pieces. tried loading workbook using file (instead of inputstream), didn't either. i looking way read 10k rows @ time (instead of entire file @ once) , iteratively write workbook , database. is there way apache poi? poi contains called "eventmodel" designed purpose. it's mentioned in faq : the ss eventmodel package api reading excel files without loading whole spreadsheet memory. require more knowledge on part of user, reduces memory consumption more tenfold. based on awt event model in combination sax. if need read-only access, best way it. however, may want double check first if issue somewhere else. check out this item : i think poi using memory! can do? 1 comes quite lot, reason isn't might think. so, first thing check - what's source of prob...
Read more

Unlimited choices in BASH case statement -

consider simple bash snippet: case $option in 1) image=${options[0]%.tar} ;; 2) image=${options[1]%.tar} ;; 3) image=${options[2]%.tar} ;; 4) image=${options[3]%.tar} ;; *) echo "invalid option" exit 1 esac in real script numbers goes 30. makes pretty long. can somehow specify cases variable? something this: case $option in $i) image=${options[$(($i-1))]%.tar} any pointer appreciated. you can match against multiple patterns in single clause: case $option in 1|2|3: echo "$option one, 2 or three" ;; esac if still typing you, can use simple pattern matching: case $option in # following matches # - single-digit numbers 0-9 # - two-digit numbers starting either 1 or 2 # - number 30 [0-9]|[12][0-9]|30) image=${options[$(($option-1))]%.tar} ;; *) echo "invalid option" 1>&2 exit 1 esac
Read more

apache - How do I stop my index.php being run twice for every user -

we've got webapp initial html served using apache , php. i noticed index.php being run twice every request. seems caused / effected rewrite rules in our .htaccess - rewritecond %{http_host} ^(www\.|hotels\.)ourdomain\.com$ [nc,or] rewritecond %{server_name} ^(www\.|hotels\.)ourdomain\.com$ [nc] rewriterule (.*) http://ourdomain.com/$1 [r=301] rewriterule ^hotels/([^/]+)/?\??(.*)$ ?d=$1&$2 [qsa] the final rule moving parameter url path query string. i.e. http://ourdomain.com/hotels/vegas?someparam=1 becomes http://ourdomain.com?d=vegas&someparam=1 if go directly query string version index.php run once. if use url redirected, index.php run twice (i'm checking adding error_log('end of index.php') file). so example going http://ourdomain.com/hotels/paris hits file twice http://ourdomain.com?d=paris hits once. i've seen this question , had @ blog mentioned , , can't find empty string url's (i've tried using yslow process). ...
Read more