int main (int argc, char ** argv)
{
char filename [256];
int rank, size, i;
int NX = 10;
double t[NX];
MPI_Comm comm = MPI_COMM_WORLD;
/* ADIOS variables declarations for matching gwrite_temperature.ch */
int adios_err;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
for (i = 0; i < NX; i++)
t[i] = rank*NX + i;
strcpy (filename, "adios_global.bp");
adios_init ("adios_global.xml", comm);
adios_open (&adios_handle, "temperature", filename, "w", comm);
#include "gwrite_temperature.ch"
adios_close (adios_handle);
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int write_file (char *fname)
{
char filename [256];
int rank, size, i, block;
int adios_err;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
strcpy (filename, "reuse_dim.bp");
adios_open (&m_adios_file, "restart", filename, "w", comm);
adios_groupsize = (2*sizeof(int) + 2*NX*sizeof(double));
adios_group_size (m_adios_file, adios_groupsize, &adios_totalsize);
adios_write(m_adios_file, "NX", (void *) &NX);
adios_write(m_adios_file, "t1", t);
adios_write(m_adios_file, "NX", (void *) &NX);
adios_write(m_adios_file, "t2", t);
adios_close (m_adios_file);
MPI_Barrier (comm);
return 0;
}
int write_file (int step)
{
int64_t fh;
uint64_t groupsize=0, totalsize;
log ("Write step %d to %s\n", step, FILENAME);
adios_open (&fh, "selections", FILENAME, (step ? "a" : "w"), comm);
groupsize = 9 * sizeof(int); // dimensions
groupsize += 3 * sizeof(int); // scalars
groupsize += 3 * ldim1 * sizeof(int); // 1D
groupsize += 3 * ldim1 * ldim2 * sizeof(int); // 2D
groupsize += 3 * ldim1 * ldim2 * ldim3 * sizeof(int); // 3D
adios_group_size (fh, groupsize, &totalsize);
adios_write (fh, "gdim1", &gdim1);
adios_write (fh, "gdim2", &gdim2);
adios_write (fh, "gdim3", &gdim3);
adios_write (fh, "ldim1", &ldim1);
adios_write (fh, "ldim2", &ldim2);
adios_write (fh, "ldim3", &ldim3);
adios_write (fh, "offs1", &offs1);
adios_write (fh, "offs2", &offs2);
adios_write (fh, "offs3", &offs3);
adios_write (fh, "a0", &a0);
adios_write (fh, "a1", a1);
adios_write (fh, "a2", a2);
adios_write (fh, "a3", a3);
adios_close (fh);
MPI_Barrier (comm);
return 0;
}
int main (int argc, char ** argv)
{
char filename [256];
int rank;
int NX = 10;
double t[NX];
char result[1024], s[32];
int i;
/* ADIOS variables declarations for matching gread_temperature.ch */
int adios_err;
uint64_t adios_groupsize, adios_totalsize, adios_buf_size;
int64_t adios_handle;
MPI_Comm comm = MPI_COMM_WORLD;
MPI_Init (&argc, &argv);
MPI_Comm_rank (MPI_COMM_WORLD, &rank);
sprintf (filename, "restart.bp");
adios_init ("config.xml", comm);
adios_open (&adios_handle, "temperature", filename, "r", comm);
#include "gread_temperature.ch"
adios_close (adios_handle);
adios_finalize (rank);
MPI_Finalize ();
sprintf(result, "rank=%d t=[%g", rank, t[0]);
for (i=1; i<NX; i++) {
sprintf (s, ",%g", t[i]);
strcat (result, s);
}
printf("%s]\n", result);
return 0;
}
int read_write(int step)
{
int retval = 0;
int i;
uint64_t total_size;
// open output file
adios_open (&fh, group_namelist[0], outfilename, (step==1 ? "w" : "a"), comm);
adios_group_size (fh, write_total, &total_size);
for (i=0; i<f->nvars; i++)
{
if (varinfo[i].writesize != 0) {
// read variable subset
print ("rank %d: Read variable %d: %s\n", rank, i, f->var_namelist[i]);
ADIOS_SELECTION *sel = adios_selection_boundingbox (varinfo[i].v->ndim,
varinfo[i].start,
varinfo[i].count);
adios_schedule_read_byid (f, sel, i, 1, 1, readbuf);
adios_perform_reads (f, 1);
// write (buffer) variable
print ("rank %d: Write variable %d: %s\n", rank, i, f->var_namelist[i]);
adios_write(fh, f->var_namelist[i], readbuf);
}
}
adios_release_step (f); // this step is no longer needed to be locked in staging area
adios_close (fh); // write out output buffer to file
return retval;
}
int main(int argc, char ** argv){
int rank=0, size=0;
int NX = NX_DIM; // size of 1D array we will write
double t[NX_DIM]; // this will contain the variables
MPI_Comm comm = MPI_COMM_WORLD; // required for ADIOS
int64_t adios_handle; // the ADIOS file handler
int retval;
struct adios_tsprt_opts adios_opts;
int err_count = 0;
GET_ENTRY_OPTIONS(adios_opts, "Runs writers. It is recommended to run as many writers as readers.");
// I assume that I have all required options set in adios_opts
// sanity check
assert(NX==NX_DIM);
// ADIOS initialization
MPI_Init(&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
SET_ERROR_IF_NOT_ZERO(adios_init(adios_opts.xml_adios_init_filename, comm), err_count);
RET_IF_ERROR(err_count, rank);
// init the array that I will transport
if (gen_1D_array(t, NX, rank) == DIAG_ERR){
printf("ERROR: Generating 1D array. Quitting ...\n");
return DIAG_ERR;
}
uint64_t adios_groupsize, adios_totalsize;
// open with the group name as specified in the xml file
adios_open( &adios_handle, "temperature", FILE_NAME, "w", comm);
adios_groupsize = 4 + 4 + 4 + 8 * (NX);
retval=adios_group_size (adios_handle, adios_groupsize, &adios_totalsize);
fprintf(stderr, "Rank=%d adios_group_size(): adios_groupsize=%" PRIu64 ", adios_totalsize=%" PRIu64 ", retval=%d\n",
rank, adios_groupsize, adios_totalsize, retval);
// write; don't check errors for simplicity reasons
adios_write(adios_handle, "NX", &NX);
adios_write(adios_handle, "size", &size);
adios_write(adios_handle, "rank", &rank);
adios_write(adios_handle, "var_1d_array", t);
fprintf(stderr, "Rank=%d committed write\n", rank);
adios_close(adios_handle);
// clean and finalize the system
adios_finalize(rank);
MPI_Finalize();
return DIAG_OK;
}
int main (int argc, char ** argv)
{
char filename [256];
int rank;
MPI_Comm comm = MPI_COMM_WORLD;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
int8_t v1 = -4;
int16_t v2 = -3;
int32_t v3 = -2;
int64_t v4 = -1;
uint8_t v5 = 1;
uint16_t v6 = 2;
uint32_t v7 = 3;
uint64_t v8 = 4;
float v9 = 5.0;
double v10 = 6.0;
char * v11 = "ADIOS example";
complex v12;
v12.r = 8.0;
v12.i = 9.0;
double_complex v13;
v13.r = 10.0;
v13.i = 11.0;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
strcpy (filename, "scalars.bp");
/* adios_open() opens a "group in a file", here the "scalars" group.
GWRITE is the convenient way to write all variables defined in the
xml file but of course one can write the individual adios_write()
statements here too
*/
adios_init ("scalars.xml", comm);
adios_open (&adios_handle, "scalars", filename, "w", comm);
#include "gwrite_scalars.ch"
adios_close (adios_handle);
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
static void write_test_file(double *arr) {
int64_t fd;
uint64_t total_size;
// Write the data to file
adios_open(&fd, BP_GROUP, BP_FILENAME, "w", &comm);
adios_group_size(fd, 2 * N * sizeof(double), &total_size);
adios_write(fd, RAW_VAR, arr);
adios_write(fd, XFORM_VAR, arr);
adios_close(fd);
}
int main (int argc, char ** argv)
{
MPI_Comm comm = 0; // dummy mpi
/* ADIOS variables declarations for matching gwrite_temperature.ch */
uint64_t adios_groupsize, adios_totalsize;
int64_t g;
int64_t f;
int64_t Tid, Pid, Vid; // variable IDs
char dimstr[32];
sprintf (dimstr, "%d,%d", NX, NY);
adios_init_noxml (comm);
adios_set_max_buffer_size (1);
adios_declare_group (&g, "vars", "", adios_flag_yes);
adios_select_method (g, "POSIX", "", "");
Tid = adios_define_var (g, "T" ,"", adios_double, dimstr, dimstr, "0,0");
adios_set_transform (Tid, "none");
Pid = adios_define_var (g, "P" ,"", adios_double, dimstr, dimstr, "0,0");
adios_set_transform (Pid, "none");
Vid = adios_define_var (g, "V" ,"", adios_double, dimstr, dimstr, "0,0");
adios_set_transform (Vid, "none");
adios_read_init_method(ADIOS_READ_METHOD_BP,0,"");
if (adios_query_is_method_available (ADIOS_QUERY_METHOD_ALACRITY)) {
adios_set_transform (Tid, "alacrity");
adios_set_transform (Pid, "alacrity");
adios_set_transform (Vid, "alacrity");
printf ("Turned on ALACRITY transformation for array variables\n");
}
adios_open (&f, "vars", "vars.bp", "w", comm);
adios_groupsize = 3*NX*NY*sizeof(double);
adios_group_size (f, adios_groupsize, &adios_totalsize);
adios_write (f, "T", T);
adios_write (f, "P", P);
adios_write (f, "V", V);
adios_close (f);
adios_finalize (0);
return 0;
}
int main (int argc, char ** argv)
{
char filename [256];
int rank, size;
int NX = 10;
int N = 3; /* number of files to write */
double t[NX];
int i;
/* ADIOS variables declarations for matching gwrite_temperature.ch */
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
int color, key;
MPI_Comm comm;
MPI_Init (&argc, &argv);
MPI_Comm_rank (MPI_COMM_WORLD, &rank);
MPI_Comm_size (MPI_COMM_WORLD, &size);
/* MPI_Comm_split partitions the world group into N disjoint subgroups,
* the processes are ranked in terms of the argument key.
* A new communicator comm is returned for this specific grid configuration
*/
color = rank % N;
key = rank / N;
MPI_Comm_split (MPI_COMM_WORLD, color, key, &comm);
for (i=0; i<NX; i++)
t[i] = rank*NX + i;
/* every P/N processes write into the same file
* there are N files generated.
*/
sprintf (filename, "restart_%5.5d.bp", color);
adios_init ("config.xml", MPI_COMM_WORLD);
adios_open (&adios_handle, "temperature", filename, "w", comm);
#include "gwrite_temperature.ch"
adios_close (adios_handle);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int main (int argc, char ** argv)
{
int size, i, block;
MPI_Comm comm = 0; // dummy mpi
/* ADIOS variables declarations for matching gwrite_temperature.ch */
uint64_t adios_groupsize, adios_totalsize;
int64_t g;
int64_t f;
char dimstr[32];
adios_init_noxml (comm);
adios_allocate_buffer (ADIOS_BUFFER_ALLOC_NOW, 1);
adios_declare_group (&g, "table", "", adios_flag_yes);
adios_select_method (g, "POSIX1", "", "");
sprintf (dimstr, "%d,%d", NX, NY);
adios_define_var (g, "A" ,"", adios_integer, dimstr, dimstr, "0,0");
sprintf (dimstr, "%d,%d", n_of_elements, Elements_length);
adios_define_var (g, "Elements" ,"", adios_byte, dimstr, dimstr, "0,0");
sprintf (dimstr, "%d,%d", NY, Columns_length);
adios_define_var (g, "Columns" ,"", adios_byte, dimstr, dimstr, "0,0");
adios_open (&f, "table", "table.bp", "w", comm);
adios_groupsize = NX*NY*sizeof(int32_t) /* size of A */
+ n_of_elements * Elements_length /* size of Elements */
+ NY * Columns_length; /* size of Columns */
adios_group_size (f, adios_groupsize, &adios_totalsize);
adios_write (f, "A", A);
adios_write (f, "Elements", Elements);
adios_write (f, "Columns", Columns);
adios_close (f);
adios_finalize (0);
return 0;
}
int write_file (int step)
{
int64_t fh;
uint64_t groupsize=0, totalsize;
log ("Write step %d to %s\n", step, FILENAME);
adios_open (&fh, "connect", FILENAME, (step ? "a" : "w"), &subcomm);
groupsize = 3 * sizeof(int); // dimensions
groupsize += ldim1 * sizeof(int); // 1D
adios_group_size (fh, groupsize, &totalsize);
adios_write (fh, "gdim1", &gdim1);
adios_write (fh, "ldim1", &ldim1);
adios_write (fh, "offs1", &offs1);
adios_write (fh, "a1", a1);
adios_close (fh);
MPI_Barrier (subcomm);
return 0;
}
int output_dump(char *filename, int step, void *data)
{
int64_t fh;
uint64_t tsize;
double t1, t2;
char fname[256], mode[2]="w";
if (streaming)
{
snprintf (fname, sizeof(fname), "data.bp",filename);
mode[0]='a'; mode[1] = 0;
}
else if (file_per_process)
{
snprintf (fname, sizeof(fname), "%s_%d.bp",filename, rank);
mode[0]='w'; mode[1] = 0;
}
else
{
snprintf (fname, sizeof(fname), "%s.bp",filename);
mode[0]='w'; mode[1] = 0;
}
t1 = MPI_Wtime();
adios_open (&fh, "writer", fname, mode, iocomm);
t2 = MPI_Wtime();
Tio_open[step] = t2-t1;
adios_group_size (fh, groupsize, &tsize);
t1 = MPI_Wtime();
Tio_group[step] = t1-t2;
adios_write (fh, "xy", data);
t2 = MPI_Wtime();
Tio_write[step] = t2-t1;
adios_close (fh);
t1 = MPI_Wtime();
Tio_close[step] = t1-t2;
return 0;
}
int main (int argc, char ** argv)
{
char filename [256];
int rank, size, i, j;
int NX = 10, NY = 100;
double t[NX][NY];
int p[NX];
MPI_Comm comm = MPI_COMM_WORLD;
int adios_err;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
for (i = 0; i < NX; i++)
for (j = 0; j< NY; j++)
t[i][j] = rank * NX + i + j*(1.0/NY);
for (i = 0; i < NX; i++)
p[i] = rank * NX + i;
strcpy (filename, "arrays.bp");
adios_init ("arrays.xml", comm);
adios_open (&adios_handle, "arrays", filename, "w", comm);
#include "gwrite_arrays.ch"
adios_close (adios_handle);
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int main (int argc, char ** argv) {
//For varriable definitions:
//gbounds = global bounds string, lbounds = local bounds string, offs = offset string, tstring = temp string to hold temperary stuff
char gbounds[1007], lbounds[1007], offs[1007],tstring[100];
//size = number of cores, gidx = adios group index
int rank, size, gidx, i, j, k, ii;
//data = pointer to read-in data
void * data = NULL;
uint64_t s[] = {0,0,0,0,0,0,0,0,0,0}; //starting offset
uint64_t c[] = {1,1,1,1,1,1,1,1,1,1}; //chunk block array
uint64_t bytes_read = 0;
int element_size;
int64_t new_adios_group, m_adios_file;
uint64_t var_size; //portion_bound,
uint64_t adios_groupsize, adios_totalsize;
int read_buffer; //possible maximum size you the user would like for each chunk in MB
int write_buffer = 1536; //actual buffer size you use in MB
int itime;
int WRITEME=1;
uint64_t chunk_size; //chunk size in # of elements
char *var_path, *var_name; // full path cut into dir path and name
MPI_Init(&argc,&argv);
MPI_Comm_rank(comm,&rank);
MPI_Comm_size(comm,&size);
// timing numbers
// we will time:
// 0: adios_open, adios_group_size
// 1: the total time to read in the data
// 2: times around each write (will only work if we do NOT buffer....
// 3: the time in the close
// 4: fopen, fclose
// 5: total time
// timers: the total I/O time
int timers = 6;
double start_time[timers], end_time[timers], total_time[timers];
if (TIMING==100) {
for (itime=0;itime<timers;itime++) {
start_time[itime] = 0;
end_time[itime] = 0;
total_time[itime]=0;
}
//MPI_Barrier(MPI_COMM_WORLD);
start_time[5] = MPI_Wtime();
}
if(rank==0)
printf("converting...\n");
if (argc < 5) {
if (rank==0) printf("Usage: %s <BP-file> <ADIOS-file> read_buffer(MB) write_buffer(MB) METHOD (LUSTRE_strip_count) (LUSTRE_strip_size) (LUSTRE_block_size)\n", argv[0]);
return 1;
}
if(TIMING==100)
start_time[4] = MPI_Wtime();
ADIOS_FILE * f = adios_fopen (argv[1], MPI_COMM_SELF);
if(TIMING==100){
end_time[4] = MPI_Wtime();
total_time[4] = end_time[4]-start_time[4];
}
adios_init_noxml(comm); // no xml will be used to write the new adios file
read_buffer = atoi(argv[3]);
write_buffer = atoi(argv[4]);
adios_allocate_buffer (ADIOS_BUFFER_ALLOC_NOW, write_buffer); // allocate MB buffer
if (f == NULL) {
printf("rank=%d, file cant be opened\n", rank);
if (DEBUG) printf ("%s\n", adios_errmsg());
return -1;
}
for (gidx = 0; gidx < f->groups_count; gidx++) { //group part
adios_groupsize = 0;
ADIOS_GROUP * g = adios_gopen (f, f->group_namelist[gidx]);
if (g == NULL) {
if (DEBUG) printf ("%s\n", adios_errmsg());
printf("rank %d: group cannot be opened.\n", rank);
return -1;
}
/* First create all of the groups */
// now I need to create this group in the file that will be written
adios_declare_group(&new_adios_group,f->group_namelist[gidx],"",adios_flag_yes);
if(strcmp(argv[5],"MPI_LUSTRE")!=0) //see whether or not the user uses MPI_LUSTRE method
adios_select_method (new_adios_group, argv[5], "", ""); //non-MPI_LUSTRE methods... like MPI, POSIX....
else{
char lustre_pars[1000];
strcpy(lustre_pars, "");
strcat(lustre_pars, "stripe_count=");
sprintf(tstring, "%d", atoi(argv[6]));
strcat(lustre_pars, tstring);
strcat(lustre_pars, ",stripe_size=");
sprintf(tstring, "%d", atoi(argv[7]));
strcat(lustre_pars, tstring);
strcat(lustre_pars, ",block_size=");
sprintf(tstring, "%d", atoi(argv[8]));
strcat(lustre_pars, tstring);
if(rank==0)
printf("lustre_pars=%s\n", lustre_pars);
adios_select_method (new_adios_group, argv[5], lustre_pars, ""); //Use MPI Lustre method
}
// variable definition part
for (i = 0; i < g->vars_count; i++) {
ADIOS_VARINFO * v = adios_inq_var_byid (g, i);
getbasename (g->var_namelist[i], &var_path, &var_name);
if (v->ndim == 0)
{
// scalars: every process does them the same.
adios_define_var(new_adios_group,var_name,var_path,v->type,0,0,0);
getTypeInfo( v->type, &element_size); //element_size is size per element based on its type
if (v->type == adios_string) { //special case when the scalar is string.
adios_groupsize += strlen(v->value);
} else {
adios_groupsize += element_size;
}
}
else
{
// vector variables
getTypeInfo( v->type, &element_size);
var_size=1;
for (ii=0;ii<v->ndim;ii++) {
var_size*=v->dims[ii];
}
uint64_t total_size = var_size; //total_size tells you the total number of elements in the current vector variable
var_size*=element_size; //var_size tells you the size of the current vector variable in bytess
//re-initialize the s and c variables
for(j=0; j<v->ndim; j++){
s[j] = 0;
c[j] = 1;
}
//find the approximate chunk_size you would like to use.
chunk_size = calcChunkSize(total_size, read_buffer*1024*1024/element_size, size);
//set the chunk block array with the total size as close to chunk_size as possible
calcC(chunk_size, v, c);
strcpy(lbounds,"");
for(j=0; j<v->ndim; j++){
sprintf(tstring, "%" PRId64 ",", c[j]);
strcat(lbounds, tstring);
}
printf("rank=%d, name=%s, chunk_size1=%" PRId64 " c[]=%s\n",rank,g->var_namelist[i],chunk_size,lbounds);
chunk_size = 1;
for(ii=0; ii<v->ndim; ii++) //reset chunk_size based on the created c. Now the chunk_size is exact.
chunk_size *= c[ii];
//current step points to where the process is in processing the vector. First sets with respect to rank.
uint64_t current_step = rank*chunk_size;
//First advance the starting point s by current_step. Of course, you don't do it if the current_step exceeds total_size.
if(current_step<total_size)
rS(v, s, current_step, rank);
uint64_t elements_defined = 0; //First, the number of elements you have defined is 0.
//You (the process) process your part of the vector when your current_step is smaller than the total_size
while(current_step < total_size)
{
//ts, temporary s, is introduced for the sake of the inner do while loop below. Copy s to ts.
uint64_t ts[] = {0,0,0,0,0,0,0,0,0,0};
arrCopy(s, ts);
//for every outer while iteration, you always have the whole chunk_size remained to process.
uint64_t remain_chunk = chunk_size;
if(current_step+chunk_size>total_size) //except when you are nearing the end of the vector....
remain_chunk = total_size-current_step;
//tc, temporary c, is introduced for the sake of the inner do while loop below. Copy s to tc.
uint64_t tc[] = {1,1,1,1,1,1,1,1,1,1};
arrCopy(c, tc);
do{
//how much of the remain chunk you wanna process? initially you think you can do all of it....
uint64_t used_chunk = remain_chunk;
//you feel like you should process the vector with tc block size, but given ts, you might go over bound.
uint64_t uc[] = {1,1,1,1,1,1,1,1,1,1};
//so you verify it by setting a new legit chunck block uc, and getting a new remain_chunk.
remain_chunk = checkBound(v, ts, tc, uc, remain_chunk);
//you check whether or not ts+uc goes over the bound. This is just checking to make sure there's no error.
//Thereotically, there should be no problem at all.
checkOverflow(0, v, ts, uc);
//the below code fragment simply calculates gbounds, and sets place holders for lbounds and offs.
strcpy(gbounds,"");
strcpy(lbounds,"");
strcpy(offs,"");
for(j=0; j<v->ndim-1; j++){
sprintf(tstring, "%d,", (int)v->dims[j]);
strcat(gbounds, tstring);
//sprintf(tstring, "ldim%d_%s,", j, var_name);
sprintf(tstring, "ldim%d,", j);
strcat(lbounds, tstring);
//sprintf(tstring, "offs%d_%s,", j, var_name);
sprintf(tstring, "offs%d,", j);
strcat(offs, tstring);
}
sprintf(tstring, "%d", (int)v->dims[v->ndim-1]);
strcat(gbounds, tstring);
//sprintf(tstring, "ldim%d_%s", v->ndim-1, var_name);
sprintf(tstring, "ldim%d", v->ndim-1);
strcat(lbounds, tstring);
//sprintf(tstring, "offs%d_%s", v->ndim-1, var_name);
sprintf(tstring, "offs%d", v->ndim-1);
strcat(offs, tstring);
//sprintf(tstring, "%d", v->ndim);
for(j=0; j<v->ndim; j++){
//sprintf(tstring, "ldim%d_%s", j, var_name);
sprintf(tstring, "ldim%d", j);
adios_define_var(new_adios_group, tstring, "bp2bp", adios_unsigned_long, 0, 0, 0);
//sprintf(tstring, "offs%d_%s", j, var_name);
sprintf(tstring, "offs%d", j);
adios_define_var(new_adios_group, tstring, "bp2bp", adios_unsigned_long, 0, 0, 0);
}
adios_define_var(new_adios_group,var_name,var_path,v->type,lbounds,gbounds,offs);
if (DEBUG){
strcpy(lbounds,"");
strcpy(offs,"");
for(j=0; j<v->ndim; j++){
sprintf(tstring, "%" PRId64 ",", ts[j]);
strcat(offs, tstring);
sprintf(tstring, "%" PRId64 ",", uc[j]);
strcat(lbounds, tstring);
}
printf("rank=%d, name=%s, gbounds=%s: lbounds=%s: offs=%s \n",rank,g->var_namelist[i],gbounds, lbounds, offs);
}
used_chunk -= remain_chunk; //you get the actual used_chunk here.
elements_defined += used_chunk;
if(remain_chunk!=0){
rS(v, ts, used_chunk, rank); //advance ts by used_chunk.
for(k=0; k<10; k++)
tc[k] = 1;
calcC(remain_chunk, v, tc); //based on the remain_chunk, calculate new tc chunk block remained to process.
}
adios_groupsize+= used_chunk*element_size+2*v->ndim*8;
}while(remain_chunk!=0);
current_step += size*chunk_size; //once a whole chunk_size is processed, advance the current_step in roll-robin manner.
if(current_step<total_size){ //advance s in the same way.
rS(v, s, size*chunk_size, rank);
}
}
//beside checkOverflow above, here you check whether or not the total number of elements processed across processes matches
//the total number of elements in the original vector.
if(DEBUG){
uint64_t* sb = (uint64_t *) malloc(sizeof(uint64_t));
uint64_t* rb = (uint64_t *) malloc(sizeof(uint64_t));
sb[0] = elements_defined;
MPI_Reduce(sb,rb,1,MPI_UNSIGNED_LONG_LONG,MPI_SUM,0, comm);
if(rank==0 && rb[0]!=total_size)
printf("some array define mismatch. please use debug mode\n");
free(sb); free(rb);
}
}
free (var_name);
free (var_path);
} // finished declaring all of the variables
// Now we can define the attributes....
for (i = 0; i < g->attrs_count; i++) {
enum ADIOS_DATATYPES atype;
int asize;
void *adata;
adios_get_attr_byid (g, i, &atype, &asize, &adata);
// if (DEBUG) printf("attribute name=%s\n",g->attr_namelist[i]);
adios_define_attribute(new_adios_group,g->attr_namelist[i],"",atype,adata,0);
}
/*------------------------------ NOW WE WRITE -------------------------------------------- */
// Now we have everything declared... now we need to write them out!!!!!!
if (WRITEME==1) {
// open up the file for writing....
if (DEBUG) printf("rank=%d, opening file = %s, with group %s, size=%" PRId64 "\n",rank,argv[2],f->group_namelist[gidx],adios_groupsize);
if(TIMING==100)
start_time[0] = MPI_Wtime();
adios_open(&m_adios_file, f->group_namelist[gidx],argv[2],"w",comm);
adios_group_size( m_adios_file, adios_groupsize, &adios_totalsize);
//get both the total adios_totalsize and total adios_groupsize summed across processes.
uint64_t* sb = (uint64_t *) malloc(sizeof(uint64_t));;
uint64_t* rb = (uint64_t *) malloc(sizeof(uint64_t));
sb[0] = adios_groupsize;
MPI_Reduce(sb,rb,1,MPI_UNSIGNED_LONG_LONG,MPI_SUM,0, comm);
uint64_t* sb2 = (uint64_t *) malloc(sizeof(uint64_t));;
uint64_t* rb2 = (uint64_t *) malloc(sizeof(uint64_t));
sb2[0] = adios_totalsize;
MPI_Reduce(sb2,rb2,1,MPI_UNSIGNED_LONG_LONG,MPI_SUM,0, comm);
if(rank==0){
printf("total adios_totalsize = %" PRId64 "\n", *rb2);
printf("total adios_groupsize = %" PRId64 "\n", *rb);
}
free(sb); free(rb); free(sb2); free(rb2);
if (TIMING==100) {
end_time[0] = MPI_Wtime();
total_time[0]+=end_time[0] - start_time[0]; //variable definition time taken
}
// now we have to write out the variables.... since they are all declared now
// This will be the place we actually write out the data!!!!!!!!
for (i = 0; i < g->vars_count; i++) {
ADIOS_VARINFO * v = adios_inq_var_byid (g, i);
getbasename (g->var_namelist[i], &var_path, &var_name);
if (v->ndim == 0)
{
if (DEBUG) {
printf ("ADIOS WRITE SCALAR: rank=%d, name=%s value=",
rank,g->var_namelist[i]);
print_data (v->value, 0, v->type);
printf ("\n");
}
if (TIMING==100) {
start_time[2] = MPI_Wtime();
}
adios_write(m_adios_file,g->var_namelist[i],v->value);
if (TIMING==100) {
end_time[2] = MPI_Wtime();
total_time[2]+=end_time[2] - start_time[2]; //IO write time...
}
}
else
{
for(j=0; j<v->ndim; j++){
s[j] = 0;
c[j] = 1;
}
getTypeInfo( v->type, &element_size);
uint64_t total_size = 1;
for (ii=0;ii<v->ndim;ii++)
total_size*=v->dims[ii];
chunk_size = calcChunkSize(total_size, read_buffer*1024*1024/element_size, size);
calcC(chunk_size, v, c);
chunk_size = 1;
for(ii=0; ii<v->ndim; ii++)
chunk_size *= c[ii];
uint64_t current_step = rank*chunk_size;
if(current_step<total_size)
rS(v, s, current_step, rank);
uint64_t elements_written = 0;
while(current_step < total_size)
{
uint64_t ts[] = {0,0,0,0,0,0,0,0,0,0};
arrCopy(s, ts);
uint64_t remain_chunk = chunk_size;
if(current_step+chunk_size>total_size)
remain_chunk = total_size-current_step;
uint64_t tc[] = {1,1,1,1,1,1,1,1,1,1};
arrCopy(c, tc);
do{
uint64_t uc[] = {1,1,1,1,1,1,1,1,1,1};
uint64_t used_chunk = remain_chunk;
remain_chunk = checkBound(v, ts, tc, uc, remain_chunk);
checkOverflow(1, v, ts, uc);
used_chunk -= remain_chunk;
elements_written += used_chunk;
//allocated space for data read-in
data = (void *) malloc(used_chunk*element_size);
if (TIMING==100) {
start_time[1] = MPI_Wtime();
}
if(PERFORMANCE_CHECK) printf("rank=%d, read start\n",rank);
bytes_read = adios_read_var_byid(g,v->varid,ts,uc,data);
if(PERFORMANCE_CHECK) printf("rank=%d, read end\n",rank);
if (TIMING==100) {
end_time[1] = MPI_Wtime();
total_time[1]+=end_time[1] -start_time[1]; //IO read time
}
if (DEBUG)
printf ("ADIOS WRITE: rank=%d, name=%s datasize=%" PRId64 "\n",rank,g->var_namelist[i],bytes_read);
if (TIMING==100) {
start_time[2] = MPI_Wtime();
}
if (DEBUG){
printf("rank=%d, write ts=",rank);
int k;
for(k=0; k<v->ndim; k++)
printf("%" PRId64 ",", ts[k]);
printf(" uc=");
for(k=0; k<v->ndim; k++)
printf("%" PRId64 ",", uc[k]);
printf("\n");
}
//local bounds and offets placeholders are not written out with actual values.
if(PERFORMANCE_CHECK) printf("rank=%d, adios write start\n", rank);
for(k=0; k<v->ndim; k++){
//sprintf(tstring, "ldim%d_%s", k, var_name);
sprintf(tstring, "ldim%d", k);
if (DEBUG) {
printf ("ADIOS WRITE DIMENSION: rank=%d, name=%s value=",
rank,tstring);
print_data (&uc[k], 0, adios_unsigned_long);
printf ("\n");
}
adios_write(m_adios_file, tstring, &uc[k]);
//sprintf(tstring, "offs%d_%s", k, var_name);
sprintf(tstring, "offs%d", k);
if (DEBUG) {
printf ("ADIOS WRITE OFFSET: rank=%d, name=%s value=",
rank,tstring);
print_data (&ts[k], 0, adios_unsigned_long);
printf ("\n");
}
adios_write(m_adios_file, tstring, &ts[k]);
}
adios_write(m_adios_file,g->var_namelist[i],data);
if(PERFORMANCE_CHECK) printf("rank=%d, adios write end\n", rank);
if (TIMING==100) {
end_time[2] = MPI_Wtime();
total_time[2]+=end_time[2] - start_time[2]; //IO write time
}
free(data);
if(remain_chunk!=0){
rS(v, ts, used_chunk, rank);
for(k=0; k<10; k++)
tc[k] = 1;
calcC(remain_chunk, v, tc);
}
}while(remain_chunk!=0);
current_step += size*chunk_size;
if(current_step<total_size)
rS(v, s, size*chunk_size,rank);
}
if(DEBUG){
uint64_t* sb = (uint64_t *) malloc(sizeof(uint64_t));;
uint64_t* rb = (uint64_t *) malloc(sizeof(uint64_t));
sb[0] = elements_written;
MPI_Reduce(sb,rb,1,MPI_UNSIGNED_LONG_LONG,MPI_SUM,0, comm);
if(rank==0 && rb[0]!=total_size)
printf("some array read mismatch. please use debug mode\n");
free(sb); free(rb);
}
}
free (var_name);
free (var_path);
}// end of the writing of the variable..
if (TIMING==100) {
start_time[3] = MPI_Wtime();
}
if(PERFORMANCE_CHECK) printf("rank=%d, adios_close start\n", rank);
adios_close(m_adios_file);
if(PERFORMANCE_CHECK) printf("rank=%d, adios_close end\n", rank);
if (TIMING==100) {
end_time[3] = MPI_Wtime();
total_time[3]+=end_time[3] - start_time[3];
}
adios_gclose(g);
} //end of WRITEME
} // end of all of the groups
if(rank==0)
printf("conversion done!\n");
if(TIMING==100)
start_time[4] = MPI_Wtime();
adios_fclose(f);
if(TIMING==100){
end_time[4] = MPI_Wtime();
total_time[4] = total_time[4]+end_time[4]-start_time[4];
}
adios_finalize(rank);
// now, we write out the timing data, for each category, we give max, min, avg, std, all in seconds, across all processes.
if(TIMING==100){
// 0: adios_open, adios_group_size
// 1: the total time to read in the data
// 2: times around each write (will only work if we do NOT buffer....
// 3: the time in the close
// 4: fopen, fclose
// 5: total time
end_time[5] = MPI_Wtime();
total_time[5] = end_time[5] - start_time[5];
double sb[7];
sb[0] = total_time[1]; sb[1] = total_time[4]; //read_var, fopen+fclose
sb[2] = sb[0]+sb[1];
sb[3] = total_time[0]; sb[4] = total_time[2]+total_time[3]; //adios_open+adios_group_size, write+close
sb[5] = sb[3]+sb[4];
sb[6] = total_time[5]; //total
double * rb = NULL;
if(rank==0)
rb = (double *)malloc(size*7*sizeof(double));
//MPI_Barrier(comm);
MPI_Gather(sb, 7, MPI_DOUBLE, rb, 7, MPI_DOUBLE, 0, comm);
if(rank==0){
double read_avg1 = 0;
double read_avg2 = 0;
double tread_avg = 0;
double write_avg1 = 0;
double write_avg2 = 0;
double twrite_avg = 0;
double total_avg = 0;
for(j=0; j<size; j++){
read_avg1 += rb[7*j];
read_avg2 += rb[7*j+1];
tread_avg += rb[7*j+2];
write_avg1 += rb[7*j+3];
write_avg2 += rb[7*j+4];
twrite_avg += rb[7*j+5];
total_avg += rb[7*j+6];
}
read_avg1 /= size;
read_avg2 /= size;
tread_avg /= size;
write_avg1 /= size;
write_avg2 /= size;
twrite_avg /= size;
total_avg /= size;
double read1_max = rb[0];
double read1_min = rb[0];
double read1_std = rb[0]-read_avg1; read1_std *= read1_std;
double read2_max = rb[1];
double read2_min = rb[1];
double read2_std = rb[1]-read_avg2; read2_std *= read2_std;
double tread_max = rb[2];
double tread_min = rb[2];
double tread_std = rb[2]-tread_avg; tread_std *= tread_std;
double write1_max = rb[3];
double write1_min = rb[3];
double write1_std = rb[3]-write_avg1; write1_std *= write1_std;
double write2_max = rb[4];
double write2_min = rb[4];
double write2_std = rb[4]-write_avg2; write2_std *= write2_std;
double twrite_max = rb[5];
double twrite_min = rb[5];
double twrite_std = rb[5]-twrite_avg; twrite_std *= twrite_std;
double total_max = rb[6];
double total_min = rb[6];
double total_std = rb[6]-total_avg; total_std *= total_std;
for(j=1; j<size; j++){
if(rb[7*j]>read1_max)
read1_max = rb[7*j];
else if(rb[7*j]<read1_min)
read1_min = rb[7*j];
double std = rb[7*j]-read_avg1; std *= std;
read1_std += std;
if(rb[7*j+1]>read2_max)
read2_max = rb[7*j+1];
else if(rb[7*j+1]<read2_min)
read2_min = rb[7*j+1];
std = rb[7*j+1]-read_avg2; std *= std;
read2_std += std;
if(rb[7*j+2]>tread_max)
tread_max = rb[7*j+2];
else if(rb[7*j+2]<tread_min)
tread_min = rb[7*j+2];
std = rb[7*j+2]-tread_avg; std *= std;
tread_std += std;
if(rb[7*j+3]>write1_max)
write1_max = rb[7*j+3];
else if(rb[7*j+3]<write1_min)
write1_min = rb[7*j+3];
std = rb[7*j+3]-write_avg1; std *= std;
write1_std += std;
if(rb[7*j+4]>write2_max)
write2_max = rb[7*j+4];
else if(rb[7*j+4]<write2_min)
write2_min = rb[7*j+4];
std = rb[7*j+4]-write_avg2; std *= std;
write2_std += std;
if(rb[7*j+5]>twrite_max)
twrite_max = rb[7*j+5];
else if(rb[7*j+5]<twrite_min)
twrite_min = rb[7*j+5];
std = rb[7*j+5]-twrite_avg; std *= std;
twrite_std += std;
if(rb[7*j+6]>total_max)
total_max = rb[7*j+6];
else if(rb[7*j+6]<total_min)
total_min = rb[7*j+6];
std = rb[7*j+6]-total_avg; std *= std;
total_std += std;
}
read1_std /= size; read1_std = sqrt(read1_std);
read2_std /= size; read2_std = sqrt(read2_std);
tread_std /= size; tread_std = sqrt(tread_std);
write1_std /= size; write1_std = sqrt(write1_std);
write2_std /= size; write2_std = sqrt(write2_std);
twrite_std /= size; twrite_std = sqrt(twrite_std);
total_std /= size; total_std = sqrt(total_std);
printf("---type--- max\tmin\tavg\tstd\n");
printf("---read_var--- %lf\t%lf\t%lf\t%lf\n", read1_max, read1_min, read_avg1, read1_std);
printf("---fopen+fclose--- %lf\t%lf\t%lf\t%lf\n", read2_max, read2_min, read_avg2, read2_std);
printf("---total_read--- %lf\t%lf\t%lf\t%lf\n", tread_max, tread_min, tread_avg, tread_std);
printf("---adios_open+adios_groupsize--- %lf\t%lf\t%lf\t%lf\n", write1_max, write1_min, write_avg1, write1_std);
printf("---write+close--- %lf\t%lf\t%lf\t%lf\n", write2_max, write2_min, write_avg2, write2_std);
printf("---total_write--- %lf\t%lf\t%lf\t%lf\n", twrite_max, twrite_min, twrite_avg, twrite_std);
printf("---total--- %lf\t%lf\t%lf\t%lf\n", total_max, total_min, total_avg, total_std);
free(rb);
}
}
// if (TIMING==100 && rank==0) {
// printf("------------------------------------------------------------------\n");
// printf("Define variables = %lf\n",total_time[0]);
// printf("Read variables = %lf\n",total_time[1]);
// printf("Write variables = %lf\n",total_time[2]);
// printf("Close File for write = %lf\n",total_time[3]);
// printf("Total write time = %lf\n",total_time[2] + total_time[3]);
// for (itime=0;itime<timers-1;itime++)
// total_time[timers-1]+=total_time[itime];
// printf("Total I/O time = %lf\n",total_time[timers-1]);
// }
MPI_Finalize();
return(0);
}
int main (int argc, char ** argv )
{
MPI_Comm comm = MPI_COMM_WORLD;
int rank;
int ndx, ndy; // size of array per processor
double *data;
double *X; //X coordinate
double *Y; //Y coordinate
// Offsets and sizes
int offs_x, offs_y; //offset in x and y direction
int nx_local, ny_local; //local address
int nx_global, ny_global; //global address
int posx, posy; // position index in the array
int i,j;
/* ADIOS variables declarations for matching gwrite_temperature.ch */
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &nproc);
if (processArgs(argc, argv)) {
return 1;
}
//will work with each core writing ndx = 65, ndy = 129, (65*3,129*4) global
ndx = 65;
ndy = 129;
//2D array with block,block decomposition
posx = rank%npx; // 1st dim
posy = rank/npx; // 2nd dim
offs_x = posx * ndx;
offs_y = posy * ndy;
nx_local = ndx;
ny_local = ndy;
nx_global = npx * ndx;
ny_global = npy * ndy;
data = malloc (ndx * ndy * sizeof(double));
for( i = 0; i < ndx; i++ )
for( j = 0; j < ndy; j++)
data[i*ndy + j] = 1.0*rank;
X = malloc (ndx * sizeof(double));
for( i = 0; i < ndx; i++ )
//X[i] = 0.1*i*i+ndx*posx;
X[i] = 0.1*(i+offs_x)*(i+offs_x);
Y = malloc (ndy * sizeof(double));
for( i = 0; i < ndy; i++ )
//Y[i] = 0.1*i*i+ndx*posy;
Y[i] = 0.1*(i+offs_y)*(i+offs_y);
adios_init ("rectilinear2d.xml", comm);
adios_open (&adios_handle, "rectilinear2d", "rectilinear2d.bp", "w", comm);
adios_groupsize = 7*sizeof(int) \
+ sizeof(double) * (nx_local*ny_local) \
+ sizeof(double) * (nx_local) \
+ sizeof(double) * (ny_local);
adios_group_size (adios_handle, adios_groupsize, &adios_totalsize);
adios_write (adios_handle, "nproc", &nproc);
adios_write (adios_handle, "nx_global", &nx_global);
adios_write (adios_handle, "ny_global", &ny_global);
adios_write (adios_handle, "offs_x", &offs_x);
adios_write (adios_handle, "offs_y", &offs_y);
adios_write (adios_handle, "nx_local", &nx_local);
adios_write (adios_handle, "ny_local", &ny_local);
if( rank < npx ) {
adios_write (adios_handle, "X", X);
}
//printf ("rank %d: check if to print Y, rank%%npx=%d offs_y=%d\n", rank, rank%npx, offs_y);
if( rank % npx == 0 )
{
adios_write (adios_handle, "Y", Y);
}
adios_write (adios_handle, "data", data);
adios_close (adios_handle);
MPI_Barrier (comm);
free (data);
free (X);
free (Y);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int flexpath_writer (MPI_Comm adios_comm, int sink_index, bool append, bool shutdown_flag)
{
int i, j, offset, size_y;
int NX = 5;
int NY = 2;
int NZ = 2;
double t[NX*NY*NZ];
// this is our array of handles that we are writing to.
static int64_t * adios_handles = 0;
// this is our "current" handle, for convenience.
int64_t * adios_handle = 0;
// how much total data is there to transfer in the array?
int total = NX * NY * NZ * comm_size;
// this flag tells the workflow to shutdown.
int shutdown = shutdown_flag ? 1 : 0;
// offsets into the array for each MPI rank.
offset = my_rank*NY;
size_y = comm_size*NY;
// Each MPI rank only writes part of the array.
int myslice = NX * NY * NZ;
for (j=0; j<NY*NX*NZ; j++) {
t[j] = my_rank*myslice + (j);
}
// if we haven't allocated space for handles, do it.
if (adios_handles == 0) {
adios_handles = (int64_t*)(calloc(num_sinks, sizeof(int64_t)));
}
// if this file isn't open, open it.
adios_handle = &(adios_handles[sink_index]);
//if((*adios_handle) == 0) {
char file_name[256] = {0};
char group_name[256] = {0};
sprintf(file_name, "adios_%s_%s", my_name, sinks[sink_index]);
sprintf(group_name, "%s_to_%s", my_name, sinks[sink_index]);
//printf("Opening %s for write\n", file_name); fflush(stdout);
if(append) {
adios_open (adios_handle, group_name, file_name, "a", adios_comm);
} else {
adios_open (adios_handle, group_name, file_name, "w", adios_comm);
}
//}
/*
* Write our variables.
*/
uint64_t adios_totalsize;
uint64_t adios_groupsize = 4 + 4 + 4 + 4 + 4 + 4 + 4 + 4 + 8 * (NZ) * (NY) * (NX);
adios_group_size ((*adios_handle), adios_groupsize, &adios_totalsize);
adios_write ((*adios_handle), "/scalar/dim/NX", &NX);
adios_write ((*adios_handle), "/scalar/dim/NY", &NY);
adios_write ((*adios_handle), "/scalar/dim/NZ", &NZ);
adios_write ((*adios_handle), "shutdown", &shutdown);
adios_write ((*adios_handle), "size", &comm_size);
adios_write ((*adios_handle), "rank", &my_rank);
adios_write ((*adios_handle), "offset", &offset);
adios_write ((*adios_handle), "size_y", &size_y);
adios_write ((*adios_handle), "var_2d_array", t);
//if (shutdown_flag) {
//printf("Closing %s for write\n", file_name); fflush(stdout);
adios_close (*adios_handle);
adios_handle = 0;
//}
return 0;
}
int main (int argc, char ** argv)
{
int i = 0;
if(argc < 4)
{
printf("wrong args\n");
usage();
return -1;
}
DIM_GLOBAL = atoi (argv[1]);
DIM_LOCAL = atoi (argv[2]);
char* option = argv[3];
char bp_file_name[NAME_LEN] = {0};
char xml_file_name[NAME_LEN] = {0};
snprintf(bp_file_name, NAME_LEN-1, "output/%s.bp", option);
snprintf(xml_file_name, NAME_LEN-1, "conf/%s.xml", option);
// MPI related intialization
int rank, nproc;
MPI_Comm comm = MPI_COMM_WORLD;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &nproc);
double t1 = 0.0;
double t2 = 0.0;
double t3 = 0.0;
double t4 = 0.0;
// variable dimensions
int gndx = DIM_GLOBAL;
int gndy = DIM_GLOBAL;
int gndz = DIM_GLOBAL;
int ndx = DIM_LOCAL;
int ndy = DIM_LOCAL;
int ndz = DIM_LOCAL;
int npx = gndx / ndx;
int npy = gndy / ndy;
int npz = gndz / ndz;
if(nproc != npx * npy * npz)
{
printf("process num error! nproc != npx * npy * npz\n");
MPI_Finalize();
return -1;
}
int posx = rank / (npx * npy);
int posy = rank % (npx * npy) / npy;
int posz = rank % (npx * npy) % npy;
// posx = mod(rank, npx) // 1st dim easy: 0, npx, 2npx... are in the same X position
// posy = mod(rank/npx, npy) // 2nd dim: (0, npx-1) have the same dim (so divide with npx first)
// posz = rank/(npx*npy) // 3rd dim: npx*npy processes belong into one dim
int offx = posx * ndx;
int offy = posy * ndy;
int offz = posz * ndz;
int timesteps = 0;
srand(0); // all procs generate the same random datasets
double* double_xyz = (double*) malloc (sizeof(double) * ndx * ndy * ndz);
for(i = 0; i < ndx * ndy * ndz; i++)
{
double_xyz[i] = (double) rand () / RAND_MAX;
}
int adios_err;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
if(rank == 0)
t3 = dclock();
MPI_Barrier(comm);
t1 = dclock();
adios_init (xml_file_name, comm);
adios_open (&adios_handle, GROUP_NAME, bp_file_name, "w", comm);
//////////////////////////////////////////////////////////////////////////////////////
adios_groupsize = 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 4
+ 8 * (ndx) * (ndy) * (ndz)
+ 8 * (ndx) * (ndy) * (ndz);
adios_group_size (adios_handle, adios_groupsize, &adios_totalsize);
adios_write (adios_handle, "gndx", &gndx);
adios_write (adios_handle, "gndy", &gndy);
adios_write (adios_handle, "gndz", &gndz);
adios_write (adios_handle, "nproc", &nproc);
adios_write (adios_handle, "npx", &npx);
adios_write (adios_handle, "npy", &npy);
adios_write (adios_handle, "npz", &npz);
adios_write (adios_handle, "offx", &offx);
adios_write (adios_handle, "offy", &offy);
adios_write (adios_handle, "offz", &offz);
adios_write (adios_handle, "ndx", &ndx);
adios_write (adios_handle, "ndy", &ndy);
adios_write (adios_handle, "ndz", &ndz);
adios_write (adios_handle, "temperature", double_xyz);
adios_write (adios_handle, "preasure", double_xyz);
//////////////////////////////////////////////////////////////////////////////////////
adios_close (adios_handle);
/*
t2 = dclock();
double tt = t2 - t1;
MPI_Barrier (comm);
if(rank == 0)
{
t4 = dclock();
}
*/
adios_finalize (rank);
/*
double* all_tt = (double*) malloc (sizeof(double) * nproc);
// calling MPI_Gather
int rtn = MPI_Gather (&tt, 1, MPI_DOUBLE, all_tt, 1, MPI_DOUBLE, 0, comm);
MPI_Barrier (comm);
if(rank == 0)
{
int k = 0;
double sum = 0.0;
for(k = 0; k < nproc; k++)
{
// printf("proc %d time %f\n", k, all_tt[k]);
sum += all_tt[k];
}
printf("%s average_write_time %f\n", xml_file_name, sum / nproc);
printf("%s total_write_time %f\n", xml_file_name, t4 - t3);
}
if(all_tt)
{
free(all_tt);
}
*/
MPI_Finalize ();
if(double_xyz)
{
free(double_xyz);
}
return 0;
}
int write_blocks ()
{
int NX, G, O;
double *t;
/* ADIOS variables declarations for matching gwrite_temperature.ch */
int it, i, r;
uint64_t adios_groupsize, adios_totalsize;
if (!rank) printf ("------- Write blocks -------\n");
// We will have "3 steps * 2 blocks per process * number of processes" blocks
nsteps = 3;
nblocks_per_step = 2;
block_offset = (uint64_t*) malloc (sizeof(uint64_t) * nsteps * nblocks_per_step * size);
block_count = (uint64_t*) malloc (sizeof(uint64_t) * nsteps * nblocks_per_step * size);
gdims = (uint64_t*) malloc (sizeof(uint64_t) * nsteps);
adios_init_noxml (comm);
adios_set_max_buffer_size (10);
int64_t m_adios_group;
int64_t m_adios_file;
adios_declare_group (&m_adios_group, "restart", "", adios_flag_yes);
adios_select_method (m_adios_group, "MPI", "", "");
adios_define_var (m_adios_group, "NX"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "G"
,"", adios_integer
,0, 0, 0);
/* have to define O and temperature as many times as we
write them within one step (twice) */
for (it=0; it < nblocks_per_step; it++) {
adios_define_var (m_adios_group, "O"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "t"
,"", adios_double
,"NX", "G", "O");
}
for (it =0; it < nsteps; it++) {
if (!rank) printf ("Step %d:\n", it);
NX = 10+it;
G = nblocks_per_step * NX * size;
t = (double *) malloc (NX*sizeof(double));
for (i = 0; i < NX; i++)
t[i] = rank + it*0.1 + 0.01;
MPI_Barrier (comm);
if (it==0)
adios_open (&m_adios_file, "restart", fname, "w", comm);
else
adios_open (&m_adios_file, "restart", fname, "a", comm);
adios_groupsize = 4 + 4 + 4 + NX * 8
+ 4 + 4 + 4 + NX * 8;
adios_group_size (m_adios_file, adios_groupsize, &adios_totalsize);
adios_write(m_adios_file, "NX", (void *) &NX);
adios_write(m_adios_file, "G", (void *) &G);
O = rank * nblocks_per_step * NX;
adios_write(m_adios_file, "O", (void *) &O);
adios_write(m_adios_file, "t", t);
printf ("rank %d: block 1: size=%d, offset=%d\n", rank, NX, O);
for (r = 0; r < size; r++) {
block_count [it*nblocks_per_step*size + nblocks_per_step*r] = NX;
block_offset [it*nblocks_per_step*size + nblocks_per_step*r] = r * nblocks_per_step * NX;
}
for (i = 0; i < NX; i++)
t[i] += 0.01;
O = rank * nblocks_per_step * NX + NX;
adios_write(m_adios_file, "O", (void *) &O);
adios_write(m_adios_file, "t", t);
printf ("rank %d: block 2: size=%d, offset=%d\n", rank, NX, O);
for (r = 0; r < size; r++) {
block_count [it*nblocks_per_step*size + nblocks_per_step*r + 1] = NX;
block_offset [it*nblocks_per_step*size + nblocks_per_step*r + 1] = r * nblocks_per_step * NX + NX;
}
gdims [it] = G;
adios_close (m_adios_file);
MPI_Barrier (comm);
free(t);
}
adios_finalize (rank);
return 0;
}
static void build_dataset_from_specs(
const char *filename_prefix,
const char *transform_name,
const dataset_xml_spec_t *xml_spec,
const dataset_global_spec_t *global_spec,
int num_ts, int num_pgs_per_ts,
dataset_pg_spec_t pg_specs[num_ts][num_pgs_per_ts]) // Not const because C has an corner case here (http://c-faq.com/ansi/constmismatch.html)
{
int var;
char xml_filename[strlen(filename_prefix) + strlen(".xml") + 1];
char bp_filename[strlen(filename_prefix) + strlen(".bp") + 1];
int timestep, pg_in_timestep;
char dimvar[32];
// Construct the XML and BP filenames
sprintf(xml_filename, "%s.xml", filename_prefix);
sprintf(bp_filename, "%s.bp", filename_prefix);
// Write out the XML file
FILE *xml_out = fopen(xml_filename, "w");
assert(xml_out);
produce_xml(xml_out, xml_spec, transform_name);
fclose(xml_out);
// Write out the BP file
adios_init(xml_filename, MPI_COMM_WORLD);
// Compute the groupsize contribution of the dimension scalars
const uint64_t base_groupsize = xml_spec->ndim * 3 * 4; // *3 for 3 scalars (N, D, O) *4 for sizeof(adios_integer) (not sure how what function in the User API to call to get this programatically
// For each timestep, for each PG in that timestep, write out all variables using the provided vardata buffers
int64_t adios_file;
for (timestep = 0; timestep < global_spec->num_ts; ++timestep) {
for (pg_in_timestep = 0; pg_in_timestep < global_spec->num_pgs_per_ts; ++pg_in_timestep) {
// (Re-)open the file in write or append mode, depending on whether or not this is the first PG written
const int is_first_pg = (timestep == 0 && pg_in_timestep == 0);
adios_open(&adios_file, xml_spec->group_name, bp_filename, is_first_pg ? "w" : "a", MPI_COMM_WORLD);
// Pin the timestep to allow multiple adios_open/adios_close cycles to write
// to the same timestep (this simulates a parallel file write with fewer core)
adios_pin_timestep(timestep + 1); // +1 because we want the timesteps to be 1-based
const dataset_pg_spec_t *pg_spec = &pg_specs[timestep][pg_in_timestep];
// Compute the group size
uint64_t groupsize = compute_groupsize(base_groupsize, xml_spec, pg_spec);
uint64_t out_groupsize;
adios_group_size(adios_file, groupsize, &out_groupsize);
write_adios_dimension_scalars(adios_file, "N", xml_spec->ndim, global_spec->global_dims);
write_adios_dimension_scalars(adios_file, "D", xml_spec->ndim, pg_spec->pg_dim);
write_adios_dimension_scalars(adios_file, "O", xml_spec->ndim, pg_spec->pg_offset);
// Write each variable
for (var = 0; var < xml_spec->nvar; ++var) {
adios_write(adios_file, xml_spec->varnames[var], (void*)pg_spec->vardata[var]); // (void*) to get rid of compiler complaining about constness
}
// Close the file to commit it
adios_close(adios_file);
}
}
}
/* --------------------------------- Main --------------------------------- */
int main( int argc, char ** argv)
{
char filename [256];
MPI_Comm comm = MPI_COMM_WORLD;
int rank, size;
/* ADIOS variables declarations for matching gwrite_schema.ch */
int adios_err;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
float tmax = 10.0;
float dt = 0.5; // run from 0.0 increasing with 'dt' up to 'tmax'
int i;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
adios_init ("local_array_time.xml", comm);
strcpy(filename, "local_array_time.bp");
// Declare and Initialize essential variables
int num_points = 37;
float angles[num_points];
float cos_of_angles[num_points];
float sin_of_angles[num_points];
float pi;
// Obtain pi once for all
pi = 4.0*atan(1.0);
// Initialize angles in degrees
float angle_degree = 0;
for (i=0; i<num_points; i++) {
angles[i] = pi * angle_degree/180.0;
angle_degree = angle_degree + 10.0;
}
// Scan over time
float timestep = 0.0;
for (timestep = 0.0; timestep <= tmax; timestep = timestep + dt) {
if (timestep == 0.0) {
printf("\n\n\nopen file\n\n\n");
adios_open (&adios_handle, "schema", filename, "w", comm);
} else {
adios_open (&adios_handle, "schema", filename, "a", comm);
}
for (i=0; i<num_points; i++) {
cos_of_angles[i] = cos(angles[i]*timestep);
sin_of_angles[i] = sin(angles[i]*timestep);
}
adios_groupsize = 4 + 4 \
+ 4*num_points \
+ 4*num_points;
if (timestep == 0 && rank == 0) {
adios_groupsize += 4 + 4 + 4*num_points;
}
adios_group_size (adios_handle, adios_groupsize, &adios_totalsize);
adios_write (adios_handle, "num_points", &num_points);
adios_write (adios_handle, "t", ×tep);
if (timestep == 0 && rank == 0) {
adios_write (adios_handle, "tmax", &tmax);
adios_write (adios_handle, "dt", &dt);
adios_write (adios_handle, "angles", angles);
}
adios_write (adios_handle, "cos", cos_of_angles);
adios_write (adios_handle, "sin", sin_of_angles);
adios_close (adios_handle);
// Write out raw data
print_data_1D(timestep, num_points, angles, sin_of_angles, 0);
}
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int main (int argc, char ** argv)
{
char filename [256];
char color_str[256];
int rank, size, i, color;
int NX = 100, Global_bounds, Offsets;
double t[NX];
MPI_Comm comm = MPI_COMM_WORLD;
/* ADIOS variables declarations for matching gwrite_temperature.ch */
uint64_t adios_groupsize, adios_totalsize;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
Global_bounds = NX * size;
strcpy (filename, "adios_global_aggregate_by_color.bp");
adios_init_noxml (comm);
adios_allocate_buffer (ADIOS_BUFFER_ALLOC_NOW, 10);
int64_t m_adios_group;
int64_t m_adios_file;
adios_declare_group (&m_adios_group, "restart", "iter", adios_flag_yes);
// split into 2 groups
color = (rank % 2 == 0 ? 0 : 1);
sprintf (color_str, "color=%d", color);
adios_select_method (m_adios_group, "MPI", color_str, "");
adios_define_var (m_adios_group, "NX"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "Global_bounds"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "Offsets"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "temperature"
,"", adios_double
,"NX", "Global_bounds", "Offsets");
adios_open (&m_adios_file, "restart", filename, "w", comm);
adios_groupsize = 4 + 4 + 4 + NX * 8;
adios_group_size (m_adios_file, adios_groupsize, &adios_totalsize);
adios_write(m_adios_file, "NX", (void *) &NX);
adios_write(m_adios_file, "Global_bounds", (void *) &Global_bounds);
Offsets = rank * NX;
adios_write(m_adios_file, "Offsets", (void *) &Offsets);
for (i = 0; i < NX; i++)
t[i] = Offsets + i;
adios_write(m_adios_file, "temperature", t);
adios_close (m_adios_file);
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int worker(int argc, char* argv[]) {
TAU_PROFILE_TIMER(timer, __func__, __FILE__, TAU_USER);
TAU_PROFILE_START(timer);
my_printf("%d of %d In worker B\n", myrank, commsize);
static bool announced = false;
/* validate input */
validate_input(argc, argv);
my_printf("Worker B will execute until it sees n iterations.\n", iterations);
/* ADIOS: These declarations are required to match the generated
* gread_/gwrite_ functions. (And those functions are
* generated by calling 'gpp.py adios_config.xml') ...
* EXCEPT THAT THE generation of Reader code is broken.
* So, we will write the reader code manually.
*/
uint64_t adios_groupsize;
uint64_t adios_totalsize;
uint64_t adios_handle;
void * data = NULL;
uint64_t start[2], count[2];
int i, j, steps = 0;
int NX = 10;
int NY = 1;
double t[NX];
double p[NX];
/* ADIOS: Can duplicate, split the world, whatever.
* This allows you to have P writers to N files.
* With no splits, everyone shares 1 file, but
* can write lock-free by using different areas.
*/
MPI_Comm adios_comm, adios_comm_b_to_c;
adios_comm = MPI_COMM_WORLD;
//MPI_Comm_dup(MPI_COMM_WORLD, &adios_comm);
adios_comm_b_to_c = MPI_COMM_WORLD;
//MPI_Comm_dup(MPI_COMM_WORLD, &adios_comm_b_to_c);
enum ADIOS_READ_METHOD method = ADIOS_READ_METHOD_FLEXPATH;
adios_read_init_method(method, adios_comm, "verbose=3");
if (adios_errno != err_no_error) {
fprintf (stderr, "rank %d: Error %d at init: %s\n", myrank, adios_errno, adios_errmsg());
exit(4);
}
if (send_to_c) {
adios_init("adios_config.xml", adios_comm);
}
/* ADIOS: Set up the adios communications and buffers, open the file.
*/
ADIOS_FILE *fp; // file handler
ADIOS_VARINFO *vi; // information about one variable
ADIOS_SELECTION * sel;
char adios_filename_a_to_b[256];
char adios_filename_b_to_c[256];
enum ADIOS_LOCKMODE lock_mode = ADIOS_LOCKMODE_NONE;
double timeout_sec = 1.0;
sprintf(adios_filename_a_to_b, "adios_a_to_b.bp");
sprintf(adios_filename_b_to_c, "adios_b_to_c.bp");
my_printf ("rank %d: Worker B opening file: %s\n", myrank, adios_filename_a_to_b);
fp = adios_read_open(adios_filename_a_to_b, method, adios_comm, lock_mode, timeout_sec);
if (adios_errno == err_file_not_found) {
fprintf (stderr, "rank %d: Stream not found after waiting %d seconds: %s\n",
myrank, timeout_sec, adios_errmsg());
exit(1);
} else if (adios_errno == err_end_of_stream) {
// stream has been gone before we tried to open
fprintf (stderr, "rank %d: Stream terminated before open. %s\n", myrank, adios_errmsg());
exit(2);
} else if (fp == NULL) {
// some other error happened
fprintf (stderr, "rank %d: Error %d at opening: %s\n", myrank, adios_errno, adios_errmsg());
exit(3);
} else {
my_printf("Found file %s\n", adios_filename_a_to_b);
my_printf ("File info:\n");
my_printf (" current step: %d\n", fp->current_step);
my_printf (" last step: %d\n", fp->last_step);
my_printf (" # of variables: %d:\n", fp->nvars);
vi = adios_inq_var(fp, "temperature");
adios_inq_var_blockinfo(fp, vi);
printf ("ndim = %d\n", vi->ndim);
printf ("nsteps = %d\n", vi->nsteps);
printf ("dims[%llu][%llu]\n", vi->dims[0], vi->dims[1]);
uint64_t slice_size = vi->dims[0]/commsize;
if (myrank == commsize-1) {
slice_size = slice_size + vi->dims[0]%commsize;
}
start[0] = myrank * slice_size;
count[0] = slice_size;
start[1] = 0;
count[1] = vi->dims[1];
data = malloc (slice_size * vi->dims[1] * 8);
/* Processing loop over the steps (we are already in the first one) */
while (adios_errno != err_end_of_stream && steps < iterations) {
steps++; // steps start counting from 1
TAU_PROFILE_TIMER(adios_recv_timer, "ADIOS recv", __FILE__, TAU_USER);
TAU_PROFILE_START(adios_recv_timer);
sel = adios_selection_boundingbox (vi->ndim, start, count);
adios_schedule_read (fp, sel, "temperature", 0, 1, data);
adios_perform_reads (fp, 1);
if (myrank == 0)
printf ("--------- B Step: %d --------------------------------\n",
fp->current_step);
#if 0
printf("B rank=%d: [0:%lld,0:%lld] = [", myrank, vi->dims[0], vi->dims[1]);
for (i = 0; i < slice_size; i++) {
printf (" [");
for (j = 0; j < vi->dims[1]; j++) {
printf ("%g ", *((double *)data + i * vi->dims[1] + j));
}
printf ("]");
}
printf (" ]\n\n");
#endif
// advance to 1) next available step with 2) blocking wait
adios_advance_step (fp, 0, timeout_sec);
if (adios_errno == err_step_notready)
{
printf ("B rank %d: No new step arrived within the timeout. Quit. %s\n",
myrank, adios_errmsg());
break; // quit while loop
}
TAU_PROFILE_STOP(adios_recv_timer);
/* Do some exchanges with neighbors */
//do_neighbor_exchange();
/* "Compute" */
compute(steps);
for (i = 0; i < NX; i++) {
t[i] = steps*100.0 + myrank*NX + i;
}
for (i = 0; i < NY; i++) {
p[i] = steps*1000.0 + myrank*NY + i;
}
if (send_to_c) {
TAU_PROFILE_TIMER(adios_send_timer, "ADIOS send", __FILE__, TAU_USER);
TAU_PROFILE_START(adios_send_timer);
/* ADIOS: write to the next application in the workflow */
if (steps == 0) {
adios_open(&adios_handle, "b_to_c", adios_filename_b_to_c, "w", adios_comm_b_to_c);
} else {
adios_open(&adios_handle, "b_to_c", adios_filename_b_to_c, "a", adios_comm_b_to_c);
}
/* ADIOS: Actually write the data out.
* Yes, this is the recommended method, and this way, changes in
* configuration with the .XML file will, even in the worst-case
* scenario, merely require running 'gpp.py adios_config.xml'
* and typing 'make'.
*/
#include "gwrite_b_to_c.ch"
/* ADIOS: Close out the file completely and finalize.
* If MPI is being used, this must happen before MPI_Finalize().
*/
adios_close(adios_handle);
TAU_PROFILE_STOP(adios_send_timer);
#if 1
if (!announced) {
SOS_val foo;
foo.i_val = NX;
SOS_pack(example_pub, "NX", SOS_VAL_TYPE_INT, foo);
SOS_announce(example_pub);
SOS_publish(example_pub);
announced = true;
}
#endif
}
MPI_Barrier(adios_comm_b_to_c);
}
MPI_Barrier(MPI_COMM_WORLD);
adios_read_close(fp);
/* ADIOS: Close out the file completely and finalize.
* If MPI is being used, this must happen before MPI_Finalize().
*/
adios_read_finalize_method(method);
}
if (send_to_c) {
adios_finalize(myrank);
}
free(data);
//MPI_Comm_free(&adios_comm);
//MPI_Comm_free(&adios_comm_b_to_c);
TAU_PROFILE_STOP(timer);
/* exit */
return 0;
}
int main (int argc, char ** argv)
{
char filename [256];
int rank, size, i, j, step, block;
int Offset;
int NX = 2; // number of records written per step per process
int Width=20;
int sub_blocks = 2; // number of record-blocks written per process in one step
int steps = 3;
char t[NX][Width];
MPI_Comm comm = MPI_COMM_WORLD;
/* ADIOS variables declarations for matching gwrite_temperature.ch */
int adios_err;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
//Global_bounds = sub_blocks * NX * size;
strcpy (filename, "steps.bp");
adios_init_noxml (comm);
adios_set_max_buffer_size (1);
int64_t m_adios_group;
int64_t m_adios_file;
adios_declare_group (&m_adios_group, "steps", "", adios_flag_yes);
adios_select_method (m_adios_group, "MPI", "", "");
adios_define_var (m_adios_group, "NX" ,"", adios_integer ,0, 0, 0);
adios_define_var (m_adios_group, "Width" ,"", adios_integer ,0, 0, 0);
adios_define_var (m_adios_group, "nproc" ,"", adios_integer ,0, 0, 0);
for (i=0;i<sub_blocks;i++) {
adios_define_var (m_adios_group, "record" ,"", adios_byte ,"NX,Width", "", "");
}
for (step=0; step<steps; step++) {
adios_open (&m_adios_file, "steps", filename, "a", comm);
adios_groupsize = sub_blocks * (4 + 4 + 4 + (uint64_t) NX * (uint64_t)Width);
adios_group_size (m_adios_file, adios_groupsize, &adios_totalsize);
adios_write(m_adios_file, "nproc", (void *) &size);
adios_write(m_adios_file, "NX", (void *) &NX);
adios_write(m_adios_file, "Width", (void *) &Width);
/* now we will write the data for each sub block */
for (block=0;block<sub_blocks;block++) {
for (i = 0; i < NX; i++)
//print 19 chars here + '\0'
sprintf (t[i], "r%2d b%2d s%2d i%2d ", rank, block, step, i);
adios_write(m_adios_file, "record", t);
}
adios_close (m_adios_file);
}
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int main (int argc, char ** argv)
{
MPI_Comm comm = MPI_COMM_WORLD;
int rank;
int ndx, ndy; // size of array per processor
double * data;
double *X; //X coordinate
double *Y; //Y coordinate
// Offsets and sizes
int offs_x, offs_y; //offset in x and y direction
int nx_local, ny_local; //local address
int nx_global, ny_global; //global address
int posx, posy; // position index in the array
int i,j;
int64_t m_adios_group;
uint64_t adios_groupsize, adios_totalsize;
int64_t adios_handle;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &nproc);
if (processArgs(argc, argv)) {
return 1;
}
//will work with each core writing ndx = 65, ndy = 129, (65*4,129*3) global
ndx = 65;
ndy = 129;
//2D array with block,block decomposition
posx = rank%npx; // 1st dim
posy = rank/npx; // 2nd dim
offs_x = posx * ndx;
offs_y = posy * ndy;
nx_local = ndx;
ny_local = ndy;
nx_global = npx * ndx;
ny_global = npy * ndy;
data = malloc (ndx * ndy * sizeof(double));
for( i = 0; i < ndx; i++ )
for( j = 0; j < ndy; j++)
data[i*ndy + j] = 1.0*rank;
X = malloc (ndx * ndy * sizeof(double));
for( i = 0; i < ndx; i++ )
for( j = 0; j < ndy; j++)
X[i*ndy + j] = offs_x + posy*ndx + i*ndx/ndx + (double)ndx*j/ndy;
Y = malloc (ndx * ndy * sizeof(double));
Y[0] = offs_y;
for( i = 0; i < ndx; i++ )
for( j = 0; j < ndy; j++)
Y[i*ndy + j] = offs_y + ndy*j/ndy;
char * schema_version = "1.1";
char * dimemsions = "nx_global,ny_global";
adios_init_noxml (comm);
adios_set_max_buffer_size (50);
adios_declare_group (&m_adios_group, "structured2d", "", adios_flag_yes);
adios_select_method (m_adios_group, "MPI", "", "");
adios_define_var (m_adios_group, "nx_global"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "ny_global"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "nproc"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "offs_x"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "offs_y"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "nx_local"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "ny_local"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "X"
,"", adios_double
,"nx_local,ny_local", "nx_global,ny_global", "offs_x,offs_y");
adios_define_var (m_adios_group, "Y"
,"", adios_double
,"nx_local,ny_local", "nx_global,ny_global", "offs_x,offs_y");
adios_define_var (m_adios_group, "data"
,"", adios_double
,"nx_local,ny_local", "nx_global,ny_global", "offs_x,offs_y");
adios_define_schema_version (m_adios_group, schema_version);
adios_define_mesh_structured (dimemsions, "X,Y", "2", m_adios_group, "structuredmesh");
adios_define_mesh_timevarying ("no", m_adios_group, "structuredmesh");
adios_define_var_mesh (m_adios_group, "data", "structuredmesh");
adios_define_var_centering (m_adios_group, "data", "point");
adios_open (&adios_handle, "structured2d", "structured2d_noxml.bp", "w", comm);
adios_groupsize = 7*sizeof(int) \
+ 3*sizeof(double) * (nx_local*ny_local);
adios_group_size (adios_handle, adios_groupsize, &adios_totalsize);
adios_write (adios_handle, "nproc", &nproc);
adios_write (adios_handle, "nx_global", &nx_global);
adios_write (adios_handle, "ny_global", &ny_global);
adios_write (adios_handle, "offs_x", &offs_x);
adios_write (adios_handle, "offs_y", &offs_y);
adios_write (adios_handle, "nx_local", &nx_local);
adios_write (adios_handle, "ny_local", &ny_local);
adios_write (adios_handle, "X", X);
adios_write (adios_handle, "Y", Y);
adios_write (adios_handle, "data", data);
adios_close (adios_handle);
MPI_Barrier (comm);
free (data);
free (X);
free (Y);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int main (int argc, char ** argv)
{
int rank, size;
int NX, NY;
int len, off;
double *t = NULL;
MPI_Comm comm = MPI_COMM_WORLD;
uint64_t start[2], count[2];
ADIOS_SELECTION *sel;
int steps = 0;
#ifdef _USE_GNUPLOT
int i, j;
double *tmp;
FILE *pipe;
#else
// Variables for ADIOS write
int64_t adios_handle;
uint64_t adios_groupsize, adios_totalsize;
char outfn[256];
#endif
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
adios_read_init_method(ADIOS_READ_METHOD_FLEXPATH, comm, "");
ADIOS_FILE* fp = adios_read_open("stream.bp",
ADIOS_READ_METHOD_FLEXPATH,
comm, ADIOS_LOCKMODE_NONE, 0.0);
assert(fp != NULL);
ADIOS_VARINFO* nx_info = adios_inq_var( fp, "NX");
ADIOS_VARINFO* ny_info = adios_inq_var( fp, "NY");
NX = *((int *)nx_info->value);
NY= *((int*)ny_info->value);
len = NX / size;
off = len * rank;
if (rank == size-1)
len = len + NX % size;
printf("\trank=%d: NX,NY,len,off = %d\t%d\t%d\t%d\n", rank, NX, NY, len, off);
assert(len > 0);
t = (double *) malloc(sizeof(double) * len * NY);
memset(t, '\0', sizeof(double) * len * NY);
assert(t != NULL);
start[0] = off;
start[1] = 0;
count[0] = len;
count[1] = NY;
// Not working ...
//sel = adios_selection_boundingbox (2, start, count);
sel = malloc(sizeof(ADIOS_SELECTION));
sel->type=ADIOS_SELECTION_WRITEBLOCK;
sel->u.block.index = rank;
#ifdef _USE_GNUPLOT
if ((NX % size) > 0)
{
fprintf(stderr, "Equal distribution is required\n");
return -1;
}
if (rank == 0) {
pipe = popen("gnuplot", "w");
fprintf(pipe, "set view map\n");
fprintf(pipe, "set xrange [0:%d]\n", NX-1);
tmp = (double *) malloc(sizeof(double) * NX * NY);
assert(tmp != NULL);
}
#else
// ADIOS write init
adios_init ("adios.xml", comm);
#endif
//while(adios_errno != err_end_of_stream && adios_errno != err_step_notready)
while(1)
{
steps++;
// Reading
adios_schedule_read (fp, sel, "var_2d_array", 0, 1, t);
adios_perform_reads (fp, 1);
printf("step=%d\trank=%d\tfp->current_step=%d\t[%d,%d]\n",
steps, rank, fp->current_step, len, NY);
/*
// Debugging
for (i=0; i<len; i++) {
printf("%d: rank=%d: t[%d,0:4] = ", steps, rank, off+i);
for (j=0; j<5; j++) {
printf(", %g", t[i*NY + j]);
}
printf(" ...\n");
}
*/
// Do something
#ifdef _USE_GNUPLOT // Option 1: plotting
MPI_Gather(t, len * NY, MPI_DOUBLE, tmp, len * NY, MPI_DOUBLE, 0, comm);
if (rank == 0)
{
fprintf(pipe, "set title 'Soft X-Rray Signal (shot #%d)'\n", steps);
fprintf(pipe, "set xlabel 'Channel#'\n");
fprintf(pipe, "set ylabel 'Timesteps'\n");
fprintf(pipe, "set cblabel 'Voltage (eV)'\n");
# ifndef _GNUPLOT_INTERACTIVE
fprintf(pipe, "set terminal png\n");
fprintf(pipe, "set output 'fig%03d.png'\n", steps);
# endif
fprintf(pipe, "splot '-' matrix with image\n");
//fprintf(pipe, "plot '-' with lines, '-' with lines, '-' with lines\n");
double *sum = calloc(NX, sizeof(double));
for (j = 0; j < NY; j++) {
for (i = 0; i < NX; i++) {
sum[i] += tmp[i * NY + j];
}
}
for (j = 0; j < NY; j++) {
for (i = 0; i < NX; i++) {
fprintf (pipe, "%g ", (-tmp[i * NY + j] + sum[i]/NY)/3276.8);
}
fprintf(pipe, "\n");
}
fprintf(pipe, "e\n");
fprintf(pipe, "e\n");
fflush (pipe);
# ifdef _GNUPLOT_INTERACTIVE
printf ("Press [Enter] to continue . . .");
fflush (stdout);
getchar ();
# endif
free(sum);
}
#else // Option 2: BP writing
snprintf (outfn, sizeof(outfn), "reader_%3.3d.bp", steps);
adios_open (&adios_handle, "reader", outfn, "w", comm);
adios_groupsize = 4 * sizeof(int) + sizeof(double) * len * NY;
adios_group_size (adios_handle, adios_groupsize, &adios_totalsize);
adios_write (adios_handle, "NX", &NX);
adios_write (adios_handle, "NY", &NY);
adios_write (adios_handle, "len", &len);
adios_write (adios_handle, "off", &off);
adios_write (adios_handle, "var", t);
adios_close (adios_handle);
#endif
// Advance
MPI_Barrier (comm);
adios_advance_step(fp, 0, TIMEOUT_SEC);
if (adios_errno == err_end_of_stream)
{
printf("rank %d, Stream terminated. Quit\n", rank);
break; // quit while loop
}
else if (adios_errno == err_step_notready)
{
printf ("rank %d: No new step arrived within the timeout. Quit.\n", rank);
break; // quit while loop
}
else if (adios_errno != err_no_error) {
printf("ADIOS returned code=%d, msg:%s\n",
adios_errno, adios_get_last_errmsg());
break; // quit while loop
}
}
//
free(t);
adios_read_close(fp);
//printf("rank %d, Successfully closed stream\n", rank);
adios_read_finalize_method(ADIOS_READ_METHOD_FLEXPATH);
//printf("rank %d, Successfully finalized read method\n", rank);
#ifndef _USE_GNUPLOT
adios_finalize (rank);
//printf("rank %d, Successfully finalized adios\n", rank);
#else
if (rank==0) {
free(tmp);
pclose(pipe);
}
#endif
MPI_Finalize ();
return 0;
}
int main (int argc, char ** argv)
{
char filename [256];
int rank, size, i, j;
int NX = 100, gb, offset; //local/global/offset
double t[NX];
int nblocks = 3;
MPI_Comm comm = MPI_COMM_WORLD;
char g_str[100], o_str[100], l_str[100];
// attributes (from C variables)
int someints[5] = {5,4,3,2,1};
double somedoubles[5] = {5.55555, 4.4444, 3.333, 2.22, 1.1};
/* ADIOS variables declarations for matching gwrite_temperature.ch */
uint64_t adios_groupsize, adios_totalsize;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
gb = nblocks * NX * size;
sprintf (g_str, "%d", gb);
sprintf (l_str, "%d", NX);
strcpy (filename, "no_xml_write_byid.bp");
adios_init_noxml (comm);
adios_set_max_buffer_size (10);
int64_t m_adios_group;
int64_t m_adios_file;
int64_t var_ids[nblocks];
adios_declare_group (&m_adios_group, "restart", "iter", adios_flag_yes);
adios_select_method (m_adios_group, "MPI", "", "");
for (i = 0; i < nblocks; i++)
{
offset = rank * nblocks * NX + i * NX;
sprintf (o_str, "%d", offset);
var_ids[i] = adios_define_var (m_adios_group, "temperature"
,"", adios_double
,l_str, g_str, o_str
);
adios_set_transform (var_ids[i], "none");
// This is here just for fun
uint64_t varsize = adios_expected_var_size(var_ids[i]);
// adios_expected_var_size() works here because the definition of the variable
// does not depend on any dimension variable (but defined with numerical dimensions)
fprintf (stderr, "Temperature block %d is %" PRIu64 " bytes\n", i, varsize);
}
// add some attributes
adios_define_attribute_byvalue (m_adios_group,
"single_string","", adios_string, 1, "A single string attribute");
char *strings[] = {"X","Yy","ZzZ"};
adios_define_attribute_byvalue (m_adios_group,
"three_strings","", adios_string_array, 3, strings);
adios_define_attribute_byvalue (m_adios_group,
"single_int", "", adios_integer, 1, &someints);
adios_define_attribute_byvalue (m_adios_group,
"single_double","", adios_double, 1, &somedoubles);
adios_define_attribute_byvalue (m_adios_group,
"five_ints", "", adios_integer, 5, &someints);
adios_define_attribute_byvalue (m_adios_group,
"five_double", "", adios_double, 5, &somedoubles);
adios_open (&m_adios_file, "restart", filename, "w", comm);
adios_groupsize = nblocks * (4 + 4 + 4 + NX * 8);
adios_group_size (m_adios_file, adios_groupsize, &adios_totalsize);
/* now we will write the data for each sub block */
for (i = 0; i < nblocks; i++)
{
offset = rank * nblocks * NX + i * NX;
for (j = 0; j < NX; j++)
t[j] = offset + j;
adios_write_byid(m_adios_file, var_ids[i], t);
}
adios_close (m_adios_file);
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int main (int argc, char ** argv)
{
char filename [256] = "stream.bp";
int rank, size;
int NX, NY;
int len, off;
double *t = NULL;
MPI_Comm comm = MPI_COMM_WORLD;
int64_t adios_handle;
uint64_t adios_groupsize, adios_totalsize;
uint64_t start[2], count[2];
ADIOS_SELECTION *sel;
int steps = 0;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
// ADIOS read init
adios_read_init_method (ADIOS_READ_METHOD_BP, comm, "verbose=3");
ADIOS_FILE* fp = adios_read_open_file ("kstar.bp",
ADIOS_READ_METHOD_BP,
comm);
assert(fp != NULL);
ADIOS_VARINFO* nx_info = adios_inq_var( fp, "N");
ADIOS_VARINFO* ny_info = adios_inq_var( fp, "L");
NX = *((int *)nx_info->value);
NY= *((int*)ny_info->value);
len = NX / size;
off = len * rank;
if (rank == size-1)
len = len + NX % size;
printf("\trank=%d: NX,NY,len,off = %d\t%d\t%d\t%d\n", rank, NX, NY, len, off);
assert(len > 0);
t = (double *) malloc(sizeof(double) * len * NY);
memset(t, '\0', sizeof(double) * len * NY);
assert(t != NULL);
start[0] = off;
start[1] = 0;
count[0] = len;
count[1] = NY;
sel = adios_selection_boundingbox (2, start, count);
// ADIOS write init
adios_init ("adios.xml", comm);
remove (filename);
//int ii;
//for(ii = 0; ii<10; ii++){
// for (i = 0; i < len * NY; i++)
// t[i] = ii*1000 + rank;
while(adios_errno != err_end_of_stream && adios_errno != err_step_notready)
{
steps++;
// Reading
adios_schedule_read (fp, sel, "var", 0, 1, t);
adios_perform_reads (fp, 1);
// Debugging
//for (i = 0; i < len*NY; i++) t[i] = off * NY + i;
printf("step=%d\trank=%d\t[%d,%d]\n", steps, rank, len, NY);
// Writing
adios_open (&adios_handle, "writer", filename, "a", comm);
adios_groupsize = 4*4 + 8*len*NY;
adios_group_size (adios_handle, adios_groupsize, &adios_totalsize);
adios_write (adios_handle, "NX", &NX);
adios_write (adios_handle, "NY", &NY);
adios_write (adios_handle, "len", &len);
adios_write (adios_handle, "off", &off);
adios_write (adios_handle, "var_2d_array", t);
adios_close (adios_handle);
// Advance
MPI_Barrier (comm);
adios_advance_step(fp, 0, TIMEOUT_SEC);
}
free(t);
MPI_Barrier (comm);
adios_read_close(fp);
if (rank==0)
printf ("We have processed %d steps\n", steps);
MPI_Barrier (comm);
adios_read_finalize_method(ADIOS_READ_METHOD_BP);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
int worker(int argc, char* argv[]) {
TAU_PROFILE_TIMER(timer, __func__, __FILE__, TAU_USER);
TAU_PROFILE_START(timer);
static bool announced = false;
my_printf("%d of %d In worker A\n", myrank, commsize);
/* validate input */
validate_input(argc, argv);
my_printf("Worker A will execute %d iterations.\n", iterations);
/* ADIOS: These declarations are required to match the generated
* gread_/gwrite_ functions. (And those functions are
* generated by calling 'gpp.py adios_config.xml') ...
*/
uint64_t adios_groupsize;
uint64_t adios_totalsize;
uint64_t adios_handle;
char adios_filename[256];
MPI_Comm adios_comm;
/* ADIOS: Can duplicate, split the world, whatever.
* This allows you to have P writers to N files.
* With no splits, everyone shares 1 file, but
* can write lock-free by using different areas.
*/
//MPI_Comm_dup(MPI_COMM_WORLD, &adios_comm);
adios_comm = MPI_COMM_WORLD;
int NX = 10;
int NY = 1;
double t[NX];
double p[NX];
/* ADIOS: Set up the adios communications and buffers, open the file.
*/
if (send_to_b) {
sprintf(adios_filename, "adios_a_to_b.bp");
adios_init("adios_config.xml", adios_comm);
}
int index, i;
for (index = 0 ; index < iterations ; index++ ) {
/* Do some exchanges with neighbors */
do_neighbor_exchange();
/* "Compute" */
compute(index);
/* Write output */
//my_printf("a");
for (i = 0; i < NX; i++) {
t[i] = index*100.0 + myrank*NX + i;
}
for (i = 0; i < NY; i++) {
p[i] = index*1000.0 + myrank*NY + i;
}
if (send_to_b) {
TAU_PROFILE_TIMER(adiostimer, "ADIOS send", __FILE__, TAU_USER);
TAU_PROFILE_START(adiostimer);
if (index == 0) {
adios_open(&adios_handle, "a_to_b", adios_filename, "w", adios_comm);
} else {
adios_open(&adios_handle, "a_to_b", adios_filename, "a", adios_comm);
}
/* ADIOS: Actually write the data out.
* Yes, this is the recommended method, and this way, changes in
* configuration with the .XML file will, even in the worst-case
* scenario, merely require running 'gpp.py adios_config.xml'
* and typing 'make'.
*/
#include "gwrite_a_to_b.ch"
/* ADIOS: Close out the file completely and finalize.
* If MPI is being used, this must happen before MPI_Finalize().
*/
adios_close(adios_handle);
TAU_PROFILE_STOP(adiostimer);
#if 1
if (!announced) {
SOS_val foo;
foo.i_val = NX;
SOS_pack(example_pub, "NX", SOS_VAL_TYPE_INT, foo);
SOS_announce(example_pub);
SOS_publish(example_pub);
announced = true;
}
#endif
}
MPI_Barrier(MPI_COMM_WORLD);
}
MPI_Barrier(MPI_COMM_WORLD);
if (send_to_b) {
adios_finalize(myrank);
}
my_printf("Worker A exting.\n");
//MPI_Comm_free(&adios_comm);
TAU_PROFILE_STOP(timer);
/* exit */
return 0;
}
int main (int argc, char ** argv)
{
char filename [256];
int rank, size, i, block;
int NX = 100, Global_bounds, Offsets;
double t[NX];
int sub_blocks = 3;
MPI_Comm comm = MPI_COMM_WORLD;
/* ADIOS variables declarations for matching gwrite_temperature.ch */
uint64_t adios_groupsize, adios_totalsize;
MPI_Init (&argc, &argv);
MPI_Comm_rank (comm, &rank);
MPI_Comm_size (comm, &size);
Global_bounds = sub_blocks * NX * size;
strcpy (filename, "adios_global_no_xml.bp");
adios_init_noxml (comm);
adios_allocate_buffer (ADIOS_BUFFER_ALLOC_NOW, 10);
int64_t m_adios_group;
int64_t m_adios_file;
adios_declare_group (&m_adios_group, "restart", "iter", adios_flag_yes);
adios_select_method (m_adios_group, "MPI", "", "");
adios_define_var (m_adios_group, "NX"
,"", adios_integer
,0, 0, 0);
adios_define_var (m_adios_group, "Global_bounds"
,"", adios_integer
,0, 0, 0);
for (i=0;i<sub_blocks;i++) {
adios_define_var (m_adios_group, "Offsets"
,"", adios_integer
,0, 0, 0);
int64_t varid;
varid = adios_define_var (m_adios_group, "temperature"
,"", adios_double
,"NX", "Global_bounds", "Offsets");
adios_set_transform (varid, "none");
}
adios_open (&m_adios_file, "restart", filename, "w", comm);
adios_groupsize = sub_blocks * (4 + 4 + 4 + NX * 8);
adios_group_size (m_adios_file, adios_groupsize, &adios_totalsize);
adios_write(m_adios_file, "NX", (void *) &NX);
adios_write(m_adios_file, "Global_bounds", (void *) &Global_bounds);
/* now we will write the data for each sub block */
for (block=0;block<sub_blocks;block++) {
Offsets = rank * sub_blocks * NX + block*NX;
adios_write(m_adios_file, "Offsets", (void *) &Offsets);
for (i = 0; i < NX; i++)
t[i] = Offsets + i;
adios_write(m_adios_file, "temperature", t);
}
adios_close (m_adios_file);
MPI_Barrier (comm);
adios_finalize (rank);
MPI_Finalize ();
return 0;
}
