bool
ADIOSFileObject::ReadVariable(const std::string &nm,
int ts,
vtkDataArray **array)
{
debug5<<"ADIOSFileObject::ReadVariable("<<nm<<" time= "<<ts<<")"<<endl;
Open();
varIter vi = variables.find(nm);
if (vi == variables.end())
{
debug5<<"Variable "<<nm<<" not found."<<endl;
return false;
}
ADIOSVar v = vi->second;
int tupleSz = adios_type_size(v.type, NULL);
*array = ADIOSFileObject::AllocateArray(v.type);
int ntuples = 1;
uint64_t start[4] = {0,0,0,0}, count[4] = {0,0,0,0};
v.GetReadArrays(ts, start, count, &ntuples);
(*array)->SetNumberOfTuples(ntuples);
void *data = (*array)->GetVoidPointer(0);
debug5<<"ARR: adios_read_var:"<<endl<<v<<endl;
OpenGroup(v.groupIdx);
uint64_t retval = adios_read_var_byid(gps[v.groupIdx], v.varid, start, count, data);
CloseGroup(v.groupIdx);
return (retval > 0);
}
void
ADIOSFileObject::GetTimes(std::string &varNm, std::vector<double> ×)
{
Open();
times.resize(0);
varIter vi = variables.find(varNm);
if (vi == variables.end())
{
debug5<<"Variable "<<varNm<<" not found."<<endl;
return;
}
ADIOSVar v = vi->second;
int tupleSz = adios_type_size(v.type, NULL);
uint64_t start[4] = {0,0,0,0}, count[4] = {0,0,0,0};
count[0] = v.count[0];
int ntuples = v.count[0];
void *readData = malloc(ntuples*tupleSz);
OpenGroup(v.groupIdx);
uint64_t retval = adios_read_var_byid(gps[v.groupIdx], v.varid, start, count, readData);
CloseGroup(v.groupIdx);
if (retval > 0)
{
times.resize(ntuples);
ConvertTo(×[0], ntuples, v.type, readData);
}
free(readData);
}
bool
ADIOSFileObject::ReadVariable(const std::string &nm,
int ts,
vtkFloatArray **array)
{
debug5<<"ADIOSFileObject::ReadVariable("<<nm<<" time= "<<ts<<")"<<endl;
Open();
varIter vi = variables.find(nm);
if (vi == variables.end())
{
debug5<<"Variable "<<nm<<" not found."<<endl;
return false;
}
ADIOSVar v = vi->second;
int tupleSz = adios_type_size(v.type, NULL);
*array = vtkFloatArray::New();
int ntuples = 1;
uint64_t start[4] = {0,0,0,0}, count[4] = {0,0,0,0};
v.GetReadArrays(ts, start, count, &ntuples);
(*array)->SetNumberOfTuples(ntuples);
float *data = (float *)(*array)->GetVoidPointer(0);
void *readData = (void *)data;
bool convertData = (v.type != adios_real);
if (convertData)
readData = malloc(ntuples*tupleSz);
debug5<<"ARR: adios_read_var:"<<endl<<v<<endl;
OpenGroup(v.groupIdx);
uint64_t retval = adios_read_var_byid(gps[v.groupIdx], v.varid, start, count, readData);
CloseGroup(v.groupIdx);
if (convertData)
{
if (retval > 0)
ConvertTo(data, ntuples, v.type, readData);
free(readData);
}
return (retval > 0);
}
bool
ADIOSFileObject::ReadCoordinates(const std::string &nm,
int ts,
vtkPoints **pts)
{
Open();
varIter vi = variables.find(nm);
if (vi == variables.end())
{
debug5<<"Variable "<<nm<<" not found."<<endl;
return false;
}
ADIOSVar v = vi->second;
*pts = ADIOSFileObject::AllocatePoints(v.type);
int ntuples = 1;
uint64_t start[4] = {0,0,0,0}, count[4] = {0,0,0,0};
v.GetReadArrays(ts, start, count, &ntuples);
ntuples /= v.dim;
(*pts)->SetNumberOfPoints(ntuples);
void *data = (*pts)->GetVoidPointer(0), *data2 = NULL;
void *readData = data;
if (v.dim < 3)
{
if (v.type == adios_real)
data2 = malloc(ntuples*3*sizeof(float));
else if (v.type == adios_double)
data2 = malloc(ntuples*3*sizeof(double));
else if (v.type == adios_integer)
data2 = malloc(ntuples*3*sizeof(int));
readData = data2;
}
OpenGroup(v.groupIdx);
debug5<<"adios_read_var:"<<endl<<v<<endl;
uint64_t retval = adios_read_var_byid(gps[v.groupIdx], v.varid, start, count, readData);
if (retval > 0)
{
if (readData != data)
{
int i, j, n = ntuples*3;
for (i=0, j=0; i < n; i += 3, j += v.dim)
{
int k;
for (k = 0; k < v.dim; k++)
((double *)data)[i+k] = ((double *)data2)[j+k];
for ( ; k < 3; k++)
((double *)data)[i+k] = 0.0;
}
free(data2);
}
}
CloseGroup(v.groupIdx);
return (retval > 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)
{
int gidx, i, j, k,l;
MPI_Comm comm_dummy = 0; /* MPI_Comm is defined through adios_read.h */
void * data = NULL;
uint64_t start[] = {0,0,0,0,0,0,0,0,0,0};
uint64_t count[10];
int64_t bytes_read = 0;
if (argc < 2) {
printf("Usage: %s <BP-file>\n", argv[0]);
return 1;
}
ADIOS_FILE * f;
//int step;
//for (step=0; step < 2; step++) {
f = adios_fopen (argv[1], comm_dummy);
if (f == NULL) {
printf ("%s\n", adios_errmsg());
return -1;
}
/* For all groups */
for (gidx = 0; gidx < f->groups_count; gidx++) {
printf("Group %s:\n", f->group_namelist[gidx]);
ADIOS_GROUP * g = adios_gopen (f, f->group_namelist[gidx]);
if (g == NULL) {
printf ("%s\n", adios_errmsg());
return -1;
}
/* For all variables */
printf(" Variables=%d:\n", g->vars_count);
for (i = 0; i < g->vars_count; i++) {
ADIOS_VARINFO * v = adios_inq_var_byid (g, i);
uint64_t total_size = adios_type_size (v->type, v->value);
for (j = 0; j < v->ndim; j++)
total_size *= v->dims[j];
printf(" %-9s %s", adios_type_to_string(v->type), g->var_namelist[i]);
if (v->ndim == 0) {
/* Scalars do not need to be read in, we get it from the metadata
when using adios_inq_var */
printf(" = %s\n", value_to_string(v->type, v->value, 0));
} else {
/* Arrays have to be read in from the file */
printf("[%" PRIu64,v->dims[0]);
for (j = 1; j < v->ndim; j++)
printf(", %" PRIu64,v->dims[j]);
//printf("] = \n");
if (v->type == adios_integer)
printf("] = min=%d max=%d timedim=%d\n", (*(int*)v->gmin), (*(int*)v->gmax), v->timedim);
else if (v->type == adios_double)
printf("] = min=%lg max=%lg timedim=%d\n", (*(double*)v->gmin), (*(double*)v->gmax), v->timedim);
if (total_size > 1024*1024*1024) {
printf(" // too big, do not read in\n");
} else {
data = malloc (total_size);
if (data == NULL) {
fprintf (stderr, "malloc failed.\n");
return -1;
}
for (j = 0; j < v->ndim; j++)
count[j] = v->dims[j];
bytes_read = adios_read_var_byid (g, i, start, count, data);
if (bytes_read < 0) {
printf ("%s\n", adios_errmsg());
} else if (bytes_read > 1024*1024) {
printf ("Too big to print\n");
} else if (v->ndim == 1) {
printf (" [");
for (j = 0; j < v->dims[0]; j++)
printf("%s ", value_to_string(v->type, data, j));
printf ("]\n");
} else if (v->ndim == 2) {
for (j = 0; j < v->dims[0]; j++) {
printf (" row %d: [", j);
for (k = 0; k < v->dims[1]; k++)
printf("%s ", value_to_string(v->type, data, j*v->dims[1] + k));
printf ("]\n");
}
} else if (v->ndim == 3) {
for (j = 0; j < v->dims[0]; j++) {
printf (" block %d: \n", j);
for (k = 0; k < v->dims[1]; k++) {
printf (" row %d: [", k);
for (l = 0; l < v->dims[2]; l++) {
printf("%s ", value_to_string(v->type, data, j*v->dims[1]*v->dims[2] + k*v->dims[1] + l));
}
printf ("]\n");
}
printf ("\n");
}
} else {
printf (" cannot print arrays with >3 dimensions\n");
}
free (data);
}
}
adios_free_varinfo (v);
} /* variables */
/* For all attributes */
printf(" Attributes=%d:\n", g->attrs_count);
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);
printf(" %-9s %s = %s\n", adios_type_to_string(atype),
g->attr_namelist[i], value_to_string(atype, adata, 0));
free(adata);
} /* attributes */
adios_gclose (g);
} /* groups */
adios_fclose (f);
//} /* loop 'step' */
return 0;
}
int readVar(ADIOS_GROUP *gp, ADIOS_VARINFO *vi, const char * name)
{
int i,j;
uint64_t start_t[MAX_DIMS], count_t[MAX_DIMS]; // processed <0 values in start/count
uint64_t s[MAX_DIMS], c[MAX_DIMS]; // for block reading of smaller chunks
hsize_t h5_start[MAX_DIMS], h5_count[MAX_DIMS], h5_stride[MAX_DIMS];
hsize_t h5_dims[MAX_DIMS];
uint64_t nelems; // number of elements to read
int elemsize; // size in bytes of one element
uint64_t st, ct;
void *data;
uint64_t sum; // working var to sum up things
int maxreadn; // max number of elements to read once up to a limit (10MB of data)
int actualreadn; // our decision how much to read at once
int readn[MAX_DIMS]; // how big chunk to read in in each dimension?
int64_t bytes_read; // retval from adios_get_var()
int incdim; // used in incremental reading in
hid_t grp_id, space_id, dataset, global_memspace, dataspace;
hid_t local_memspace, h5_ndim ;
hid_t h5_err;
hid_t h5_type_id;
if (getTypeInfo(vi->type, &elemsize)) {
fprintf(stderr, "Adios type %d (%s) not supported in bpls. var=%s\n",
vi->type, adios_type_to_string(vi->type), name);
return 10;
}
h5_err = bp_getH5TypeId (vi->type, &h5_type_id);
h5_ndim = (hsize_t) vi->ndim;
for (j=0;j<h5_ndim;j++)
h5_dims[j] = vi->dims[j];
// create the hdf5 dataspace.
// create the counter arrays with the appropriate lengths
// transfer start and count arrays to format dependent arrays
for (j=0; j<vi->ndim; j++) {
icount[j]=-1;
h5_stride[j]= (hsize_t) 1;
}
nelems = 1;
for (j=0; j<vi->ndim; j++) {
if (istart[j] < 0) // negative index means last-|index|
st = vi->dims[j]+istart[j];
else
st = istart[j];
if (icount[j] < 0) // negative index means last-|index|+1-start
ct = vi->dims[j]+icount[j]+1-st;
else
ct = icount[j];
if (verbose>2)
printf(" j=%d, st=%llu ct=%llu\n", j, st, ct);
start_t[j] = st;
count_t[j] = ct;
nelems *= ct;
if (verbose>1)
printf(" s[%d]=%llu, c[%d]=%llu, n=%llu\n", j, start_t[j], j, count_t[j], nelems);
}
if (verbose>1) {
printf(" total size of data to read = %llu\n", nelems*elemsize);
}
maxreadn = MAX_BUFFERSIZE/elemsize;
if (nelems < maxreadn)
maxreadn = nelems;
// special case: string. Need to use different elemsize
if (vi->type == adios_string) {
if (vi->value)
elemsize = strlen(vi->value)+1;
maxreadn = elemsize;
}
// allocate data array
data = (void *) malloc (maxreadn*elemsize); // SAK: don't want the +8....
//+8 for just to be sure
// determine strategy how to read in:
// - at once
// - loop over 1st dimension
// - loop over 1st & 2nd dimension
// - etc
if (verbose>1) printf("Read size strategy:\n");
sum = (uint64_t) 1;
actualreadn = (uint64_t) 1;
for (i=vi->ndim-1; i>=0; i--) {
if (sum >= (uint64_t) maxreadn) {
readn[i] = 1;
} else {
readn[i] = maxreadn / (int)sum; // sum is small for 4 bytes here
// this may be over the max count for this dimension
if (readn[i] > count_t[i])
readn[i] = count_t[i];
}
if (verbose>1) printf(" dim %d: read %d elements\n", i, readn[i]);
sum = sum * (uint64_t) count_t[i];
actualreadn = actualreadn * readn[i];
}
if (verbose>1) printf(" read %d elements at once, %lld in total (nelems=%lld)\n", actualreadn, sum, nelems);
// init s and c
for (j=0; j<vi->ndim; j++) {
s[j]=start_t[j];
c[j]=readn[j];
h5_count[j] = (hsize_t) c[j];
h5_start[j] = (hsize_t) s[j];
}
// read until read all 'nelems' elements
sum = 0;
while (sum < nelems) {
// how many elements do we read in next?
actualreadn = 1;
for (j=0; j<vi->ndim; j++)
actualreadn *= c[j];
if (verbose>2) {
printf("adios_read_var name=%s ", name);
//PRINT_DIMS(" start", s, vi->ndim, j);
//PRINT_DIMS(" count", c, vi->ndim, j);
printf(" read %d elems\n", actualreadn);
}
// read a slice finally
bytes_read = adios_read_var_byid (gp, vi->varid, s, c, data);
if (bytes_read < 0) {
fprintf(stderr, "Error when reading variable %s. errno=%d : %s \n", name, adios_errno, adios_errmsg());
free(data);
return 11;
}
// now we must place this inside the hdf5 file... we know the offset (s)
// we know the count, c
// we know the global rank v->ndim
// we know the global dimensions v->dims
// get the hdf5 calls for writing the hyperslab.....
dataset = H5Dopen(HDF5_FILE,name);
if (dataset> 0) {
global_memspace = H5Dget_space(dataset);
hid_t rank_old = H5Sget_simple_extent_ndims(global_memspace);
hsize_t *maxdims = (hsize_t *) malloc (h5_ndim * sizeof (hsize_t));
h5_err = H5Sget_simple_extent_dims(global_memspace,maxdims,NULL);
free(maxdims);
} else {
global_memspace = H5Screate_simple (h5_ndim, h5_dims, NULL);
hid_t cparms = H5Pcreate(H5P_DATASET_CREATE);
h5_err = H5Pset_chunk(cparms,h5_ndim,h5_count);
dataset = H5Dcreate(HDF5_FILE,name,h5_type_id, global_memspace,cparms);
H5Pclose(cparms);
}
local_memspace = H5Screate_simple (h5_ndim, h5_count, NULL);
for (j=0;j<vi->ndim;j++)
h5_err = H5Sselect_hyperslab (global_memspace, H5S_SELECT_SET
,h5_start, h5_stride, h5_count, NULL);
h5_err = H5Dwrite(dataset,h5_type_id,local_memspace,global_memspace,H5P_DEFAULT, data);
H5Sclose(local_memspace);
H5Dclose(dataset);
//H5Tclose(h5_type_id);
if (verbose>2) printf(" read %lld bytes\n", bytes_read);
// print slice
// prepare for next read
sum += actualreadn;
incdim=1; // largest dim should be increased
for (j=vi->ndim-1; j>=0; j--) {
if (incdim==1) {
if (s[j]+c[j] == start_t[j]+count_t[j]) {
// reached the end of this dimension
s[j] = start_t[j];
c[j] = readn[j];
h5_count[j] = (hsize_t) c[j];
h5_start[j] = (hsize_t) s[j];
incdim = 1; // next smaller dim can increase too
} else {
// move up in this dimension up to total count
s[j] += readn[j];
h5_start[j] = (hsize_t) s[j];
if (s[j]+c[j] > start_t[j]+count_t[j]) {
// do not reach over the limit
c[j] = start_t[j]+count_t[j]-s[j];
h5_count[j] = (hsize_t) c[j];
}
incdim = 0;
}
}
}
} // end while sum < nelems
H5Sclose(global_memspace);
free(data);
return 0;
}