OilfieldToolsNet/LBPM
4
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Adding function to reorder ranks based on load balancing

Browse Source
This commit is contained in:
Mark Berrill
2020-05-15 13:39:36 -04:00
parent 94156e066e
commit 85cc3363da
3 changed files with 244 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
common/MPI.cpp
View File

@@ -2,6 +2,7 @@
#include "common/MPI.h" #include "common/MPI.h"
#include "common/Utilities.h" #include "common/Utilities.h"
#include "common/Utilities.hpp"
#include "ProfilerApp.h" #include "ProfilerApp.h"
#include "StackTrace/ErrorHandlers.h" #include "StackTrace/ErrorHandlers.h"
@@ -3806,5 +3807,25 @@ void MPI_CLASS::stop_MPI()
} }
/****************************************************************************
* Function to perform load balancing *
****************************************************************************/
MPI MPI::loadBalance( double local, std::vector<double> work )
{
MPI_ASSERT( (int) work.size() == getSize() );
auto perf = allGather( local );
std::vector<int> I( work.size() );
for ( size_t i=0; i<work.size(); i++)
I[i] = i;
auto J = I;
quicksort( perf, I );
quicksort( work, J );
std::vector<int> key( work.size() );
for ( size_t i=0; i<work.size(); i++)
key[J[i]] = I[i];
return split( 0, key[getRank()] );
}
} // namespace Utilities } // namespace Utilities

7
common/MPI.h
View File

@@ -1067,6 +1067,13 @@ public: // Member functions
static void stop_MPI(); static void stop_MPI();
/*!
* \brief Load balance
* \details This function will return a new communicator in which the ranks match
* the performance and the work load.
*/
MPI loadBalance( double localPerformance, std::vector<double> work );
private: // Private helper functions for templated MPI operations; private: // Private helper functions for templated MPI operations;
template<class type> template<class type>
void call_sumReduce( type *x, const int n = 1 ) const; void call_sumReduce( type *x, const int n = 1 ) const;

216
common/Utilities.hpp Normal file
View File

@@ -0,0 +1,216 @@
#ifndef included_Utilities_hpp
#define included_Utilities_hpp
#include "Utilities.h"
#include <vector>
namespace Utilities {
/************************************************************************
* templated quicksort routines *
************************************************************************/
template<class T>
void quicksort( std::vector<T> &x )
{
if ( x.size() <= 1u )
return;
T *arr = &x[0];
bool test;
long int i, ir, j, jstack, k, l, istack[100];
T a, tmp_a;
jstack = 0;
l = 0;
ir = x.size() - 1;
while ( 1 ) {
if ( ir - l < 7 ) { // Insertion sort when subarray small enough.
for ( j = l + 1; j <= ir; j++ ) {
a = arr[j];
test = true;
for ( i = j - 1; i >= 0; i-- ) {
if ( arr[i] < a ) {
arr[i + 1] = a;
test = false;
break;
}
arr[i + 1] = arr[i];
}
if ( test ) {
i = l - 1;
arr[i + 1] = a;
}
}
if ( jstack == 0 )
return;
ir = istack[jstack]; // Pop stack and begin a new round of partitioning.
l = istack[jstack - 1];
jstack -= 2;
} else {
k = ( l + ir ) / 2; // Choose median of left, center and right elements as partitioning
// element a. Also rearrange so that a(l) < a(l+1) < a(ir).
tmp_a = arr[k];
arr[k] = arr[l + 1];
arr[l + 1] = tmp_a;
if ( arr[l] > arr[ir] ) {
tmp_a = arr[l];
arr[l] = arr[ir];
arr[ir] = tmp_a;
}
if ( arr[l + 1] > arr[ir] ) {
tmp_a = arr[l + 1];
arr[l + 1] = arr[ir];
arr[ir] = tmp_a;
}
if ( arr[l] > arr[l + 1] ) {
tmp_a = arr[l];
arr[l] = arr[l + 1];
arr[l + 1] = tmp_a;
}
// Scan up to find element > a
j = ir;
a = arr[l + 1]; // Partitioning element.
for ( i = l + 2; i <= ir; i++ ) {
if ( arr[i] < a )
continue;
while ( arr[j] > a ) // Scan down to find element < a.
j--;
if ( j < i )
break; // Pointers crossed. Exit with partitioning complete.
tmp_a = arr[i]; // Exchange elements of both arrays.
arr[i] = arr[j];
arr[j] = tmp_a;
}
arr[l + 1] = arr[j]; // Insert partitioning element in both arrays.
arr[j] = a;
jstack += 2;
// Push pointers to larger subarray on stack, process smaller subarray immediately.
if ( ir - i + 1 >= j - l ) {
istack[jstack] = ir;
istack[jstack - 1] = i;
ir = j - 1;
} else {
istack[jstack] = j - 1;
istack[jstack - 1] = l;
l = i;
}
}
}
}
template<class T1, class T2>
void quicksort( std::vector<T1> &x, std::vector<T2> &y )
{
if ( x.size() <= 1u )
return;
T1 *arr = &x[0];
T2 *brr = &y[0];
bool test;
long int i, ir, j, jstack, k, l, istack[100];
T1 a, tmp_a;
T2 b, tmp_b;
jstack = 0;
l = 0;
ir = x.size() - 1;
while ( 1 ) {
if ( ir - l < 7 ) { // Insertion sort when subarray small enough.
for ( j = l + 1; j <= ir; j++ ) {
a = arr[j];
b = brr[j];
test = true;
for ( i = j - 1; i >= 0; i-- ) {
if ( arr[i] < a ) {
arr[i + 1] = a;
brr[i + 1] = b;
test = false;
break;
}
arr[i + 1] = arr[i];
brr[i + 1] = brr[i];
}
if ( test ) {
i = l - 1;
arr[i + 1] = a;
brr[i + 1] = b;
}
}
if ( jstack == 0 )
return;
ir = istack[jstack]; // Pop stack and begin a new round of partitioning.
l = istack[jstack - 1];
jstack -= 2;
} else {
k = ( l + ir ) / 2; // Choose median of left, center and right elements as partitioning
// element a. Also rearrange so that a(l) ? a(l+1) ? a(ir).
tmp_a = arr[k];
arr[k] = arr[l + 1];
arr[l + 1] = tmp_a;
tmp_b = brr[k];
brr[k] = brr[l + 1];
brr[l + 1] = tmp_b;
if ( arr[l] > arr[ir] ) {
tmp_a = arr[l];
arr[l] = arr[ir];
arr[ir] = tmp_a;
tmp_b = brr[l];
brr[l] = brr[ir];
brr[ir] = tmp_b;
}
if ( arr[l + 1] > arr[ir] ) {
tmp_a = arr[l + 1];
arr[l + 1] = arr[ir];
arr[ir] = tmp_a;
tmp_b = brr[l + 1];
brr[l + 1] = brr[ir];
brr[ir] = tmp_b;
}
if ( arr[l] > arr[l + 1] ) {
tmp_a = arr[l];
arr[l] = arr[l + 1];
arr[l + 1] = tmp_a;
tmp_b = brr[l];
brr[l] = brr[l + 1];
brr[l + 1] = tmp_b;
}
// Scan up to find element > a
j = ir;
a = arr[l + 1]; // Partitioning element.
b = brr[l + 1];
for ( i = l + 2; i <= ir; i++ ) {
if ( arr[i] < a )
continue;
while ( arr[j] > a ) // Scan down to find element < a.
j--;
if ( j < i )
break; // Pointers crossed. Exit with partitioning complete.
tmp_a = arr[i]; // Exchange elements of both arrays.
arr[i] = arr[j];
arr[j] = tmp_a;
tmp_b = brr[i];
brr[i] = brr[j];
brr[j] = tmp_b;
}
arr[l + 1] = arr[j]; // Insert partitioning element in both arrays.
arr[j] = a;
brr[l + 1] = brr[j];
brr[j] = b;
jstack += 2;
// Push pointers to larger subarray on stack, process smaller subarray immediately.
if ( ir - i + 1 >= j - l ) {
istack[jstack] = ir;
istack[jstack - 1] = i;
ir = j - 1;
} else {
istack[jstack] = j - 1;
istack[jstack - 1] = l;
l = i;
}
}
}
}
}
#endif