Collaboration diagram for CPM::SGv::Traff_builder:

Public Member Functions
void	build (int size, int root, int rank, int *counts, Graph &g, Vertex &r, Vertex &u, Vertex &v)

Detailed Description

Model-based builder of Traff communication tree for scatterv/gatherv

Member Function Documentation

void CPM::SGv::Traff_builder::build	(	int	size,
		int	root,
		int	rank,
		int *	counts,
		Graph &	g,
		Vertex &	r,
		Vertex &	u,
		Vertex &	v
	)			`[inline]`

The routine build is based on two phases which take place in a loop:

1. PHASE: in the first phase, the given set S_i is partitioned into subsets in following way:

if S_i is not empty:
- sort the ranks from S_i in decreasing order according to the value of predict_p2p(i,parent(S_i),m_size(i))
- (*) partition S_i into subsets so that for every subset W(S_j1) >= Sum(j2=0..j1-1) W(S_j2) and j1 is minimal. ->W(S):=sum(i in S) predict_p2p(i,parent(S),m_size(i)); 2. PHASE: in the second phase, determine the root of each new subset from 1(*) in the following way:
in a routine getRoot, find the element of the subset to become root. For the optimal case, that routine would be:
- find i from S_i so that predict_p2p(i,parent(S_i),Sum(j in S_i) m_size(j))= MINIMAL
- make i the root of S_i and continue from 1. with the children of S_i (remark: this root may be different than the first element (and so different than Traff's root)

The documentation for this class was generated from the following file: