OPAL (Object Oriented Parallel Accelerator Library)  2.2.0
OPAL
VRB.h
Go to the documentation of this file.
1 // -*- C++ -*-
2 /***************************************************************************
3  *
4  * The IPPL Framework
5  *
6  *
7  * Visit http://people.web.psi.ch/adelmann/ for more details
8  *
9  ***************************************************************************/
10 
11 #ifndef VRB_H
12 #define VRB_H
13 
15 //
16 // Vnode Recursive Bisection package
17 //
18 // This package figures out how to distribute rectangular arrays of
19 // vnodes onto processors with an attempt to minimize communication
20 // and unbalance.
21 //
22 // There one function with external linkage in this package:
23 //
24 // VnodeRecurseiveBisection(int dim, const int* sizes, int nprocs, int *vprocs);
25 //
26 // Input:
27 // dim : The number of dimensions.
28 // sizes: The number of vnodes in each dimension.
29 // nprocs: The number of procs to distribute them over.
30 //
31 // Output:
32 // procs: the proc number for each vnode, in the range 0..nprocs-1.
33 //
35 
37 //
38 // The problem:
39 //
40 // Given a set of vnodes in D dimensions, we need to allocate them
41 // to processors in a way that satisfies three criteria:
42 //
43 // 1. Layouts with the same number of vnodes in each direction should
44 // be aligned independent of the size of the array in each dimension.
45 // That means that the distribution is independent of whether it is
46 // for a vert or cell centered Field.
47 //
48 // We do this by considering two things:
49 //
50 // a. Consider each vnode to have the same weight. That is, don't
51 // try to consider the fact that different vnodes may have slightly
52 // different numbers of cells in them.
53 //
54 // b. Each face between vnodes has the same computational cost. That is,
55 // don't try to consider the fact that the faces between vnodes will
56 // have different sizes if the vnodes are not square.
57 //
58 //
59 // 2. Minimize communication.
60 //
61 // This means that you try to minimize the number of vnode faces
62 // that must be communicated.
63 //
64 // 3. Minimize unbalance
65 //
66 // This means that you try to put the same number of vnodes on
67 // each processor.
68 //
70 //
71 // The Solution: Generalized Recursive Bisection
72 //
73 // The algorithm operates in a series of stages.
74 // In each stage you split the work to be done approximately in half
75 // and the processor pool approximately in half, and giving the split
76 // work to the split processor pools.
77 //
78 // There are multipple ways of doing this.
79 // 1. You can split along coordinate directions.
80 // 2. You can renumber the vnodes with some space filling curve, and then
81 // split along that curve.
82 //
83 // After trying both, we use number 1.
84 //
86 
87 void vnodeRecursiveBisection(int dim, const int* sizes, int nprocs, int *vprocs);
88 
89 #endif // VRB_H
90 
91 /***************************************************************************
92  * $RCSfile: VRB.h,v $ $Author: adelmann $
93  * $Revision: 1.1.1.1 $ $Date: 2003/01/23 07:40:27 $
94  * IPPL_VERSION_ID: $Id: VRB.h,v 1.1.1.1 2003/01/23 07:40:27 adelmann Exp $
95  ***************************************************************************/
void vnodeRecursiveBisection(int dim, const int *sizes, int nprocs, int *procs)
Definition: VRB.cpp:69