1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
|
// Copyright (C) 2004-2006 The Trustees of Indiana University.
// Copyright (C) 2002 Brad King and Douglas Gregor
// Use, modification and distribution is subject to the Boost Software
// License, Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt)
// Authors: Douglas Gregor
// Andrew Lumsdaine
// Brian Barrett
#ifndef BOOST_PARALLEL_GRAPH_PAGE_RANK_HPP
#define BOOST_PARALLEL_GRAPH_PAGE_RANK_HPP
#ifndef BOOST_GRAPH_USE_MPI
#error "Parallel BGL files should not be included unless <boost/graph/use_mpi.hpp> has been included"
#endif
#include <boost/assert.hpp>
#include <boost/graph/overloading.hpp>
#include <boost/graph/page_rank.hpp>
#include <boost/graph/distributed/concepts.hpp>
#include <boost/property_map/parallel/distributed_property_map.hpp>
#include <boost/property_map/parallel/caching_property_map.hpp>
#include <boost/graph/parallel/algorithm.hpp>
#include <boost/graph/parallel/container_traits.hpp>
// #define WANT_MPI_ONESIDED 1
namespace boost { namespace graph { namespace distributed {
namespace detail {
#ifdef WANT_MPI_ONESIDED
template<typename Graph, typename RankMap, typename owner_map_t>
void page_rank_step(const Graph& g, RankMap from_rank, MPI_Win to_win,
typename property_traits<RankMap>::value_type damping,
owner_map_t owner)
{
typedef typename property_traits<RankMap>::value_type rank_type;
int me, ret;
MPI_Comm_rank(MPI_COMM_WORLD, &me);
// MPI_Accumulate is not required to store the value of the data
// being sent, only the address. The value of the memory location
// must not change until the end of the access epoch, meaning the
// call to MPI_Fence. We therefore store the updated value back
// into the from_rank map before the accumulate rather than using
// a temporary. We're going to reset the values in the from_rank
// before the end of page_rank_step() anyway, so this isn't a huge
// deal. But MPI-2 One-sided is an abomination.
BGL_FORALL_VERTICES_T(u, g, Graph) {
put(from_rank, u, (damping * get(from_rank, u) / out_degree(u, g)));
BGL_FORALL_ADJ_T(u, v, g, Graph) {
ret = MPI_Accumulate(&(from_rank[u]),
1, MPI_DOUBLE,
get(owner, v), local(v),
1, MPI_DOUBLE, MPI_SUM, to_win);
BOOST_ASSERT(MPI_SUCCESS == ret);
}
}
MPI_Win_fence(0, to_win);
// Set new rank maps for the other map. Do this now to get around
// the stupid synchronization rules of MPI-2 One-sided
BGL_FORALL_VERTICES_T(v, g, Graph) put(from_rank, v, rank_type(1 - damping));
}
#endif
template<typename T>
struct rank_accumulate_reducer {
BOOST_STATIC_CONSTANT(bool, non_default_resolver = true);
template<typename K>
T operator()(const K&) const { return T(0); }
template<typename K>
T operator()(const K&, const T& x, const T& y) const { return x + y; }
};
} // end namespace detail
template<typename Graph, typename RankMap, typename Done, typename RankMap2>
void
page_rank_impl(const Graph& g, RankMap rank_map, Done done,
typename property_traits<RankMap>::value_type damping,
typename graph_traits<Graph>::vertices_size_type n,
RankMap2 rank_map2)
{
typedef typename property_traits<RankMap>::value_type rank_type;
int me;
MPI_Comm_rank(MPI_COMM_WORLD, &me);
typename property_map<Graph, vertex_owner_t>::const_type
owner = get(vertex_owner, g);
(void)owner;
typedef typename boost::graph::parallel::process_group_type<Graph>
::type process_group_type;
typedef typename process_group_type::process_id_type process_id_type;
process_group_type pg = process_group(g);
process_id_type id = process_id(pg);
BOOST_ASSERT(me == id);
rank_type initial_rank = rank_type(rank_type(1) / n);
BGL_FORALL_VERTICES_T(v, g, Graph) put(rank_map, v, initial_rank);
#ifdef WANT_MPI_ONESIDED
BOOST_ASSERT(sizeof(rank_type) == sizeof(double));
bool to_map_2 = true;
MPI_Win win, win2;
MPI_Win_create(&(rank_map[*(vertices(g).first)]),
sizeof(double) * num_vertices(g),
sizeof(double),
MPI_INFO_NULL, MPI_COMM_WORLD, &win);
MPI_Win_set_name(win, "rank_map_win");
MPI_Win_create(&(rank_map2[*(vertices(g).first)]),
sizeof(double) * num_vertices(g),
sizeof(double),
MPI_INFO_NULL, MPI_COMM_WORLD, &win2);
MPI_Win_set_name(win, "rank_map2_win");
// set initial rank maps for the first iteration...
BGL_FORALL_VERTICES_T(v, g, Graph) put(rank_map2, v, rank_type(1 - damping));
MPI_Win_fence(0, win);
MPI_Win_fence(0, win2);
while ((to_map_2 && !done(rank_map, g)) ||
(!to_map_2 && !done(rank_map2, g))) {
if (to_map_2) {
graph::distributed::detail::page_rank_step(g, rank_map, win2, damping, owner);
to_map_2 = false;
} else {
graph::distributed::detail::page_rank_step(g, rank_map2, win, damping, owner);
to_map_2 = true;
}
}
synchronize(boost::graph::parallel::process_group(g));
MPI_Win_free(&win);
MPI_Win_free(&win2);
#else
// The ranks accumulate after each step.
rank_map.set_reduce(detail::rank_accumulate_reducer<rank_type>());
rank_map2.set_reduce(detail::rank_accumulate_reducer<rank_type>());
rank_map.set_consistency_model(boost::parallel::cm_flush | boost::parallel::cm_reset);
rank_map2.set_consistency_model(boost::parallel::cm_flush | boost::parallel::cm_reset);
bool to_map_2 = true;
while ((to_map_2 && !done(rank_map, g)) ||
(!to_map_2 && !done(rank_map2, g))) {
/**
* PageRank can implemented slightly more efficiently on a
* bidirectional graph than on an incidence graph. However,
* distributed PageRank requires that we have the rank of the
* source vertex available locally, so we force the incidence
* graph implementation, which pushes rank from source to
* target.
*/
typedef incidence_graph_tag category;
if (to_map_2) {
graph::detail::page_rank_step(g, rank_map, rank_map2, damping,
category());
to_map_2 = false;
} else {
graph::detail::page_rank_step(g, rank_map2, rank_map, damping,
category());
to_map_2 = true;
}
using boost::graph::parallel::process_group;
synchronize(process_group(g));
}
rank_map.reset();
#endif
if (!to_map_2)
BGL_FORALL_VERTICES_T(v, g, Graph) put(rank_map, v, get(rank_map2, v));
}
template<typename Graph, typename RankMap, typename Done, typename RankMap2>
void
page_rank(const Graph& g, RankMap rank_map, Done done,
typename property_traits<RankMap>::value_type damping,
typename graph_traits<Graph>::vertices_size_type n,
RankMap2 rank_map2
BOOST_GRAPH_ENABLE_IF_MODELS_PARM(Graph, distributed_graph_tag))
{
(page_rank_impl)(g, rank_map, done, damping, n, rank_map2);
}
template<typename MutableGraph>
void
remove_dangling_links(MutableGraph& g
BOOST_GRAPH_ENABLE_IF_MODELS_PARM(MutableGraph,
distributed_graph_tag))
{
typename graph_traits<MutableGraph>::vertices_size_type old_n;
do {
old_n = num_vertices(g);
typename graph_traits<MutableGraph>::vertex_iterator vi, vi_end;
for (boost::tie(vi, vi_end) = vertices(g); vi != vi_end; /* in loop */) {
typename graph_traits<MutableGraph>::vertex_descriptor v = *vi++;
if (out_degree(v, g) == 0) {
clear_vertex(v, g);
remove_vertex(v, g);
}
}
} while (num_vertices(g) < old_n);
}
} // end namespace distributed
using distributed::page_rank;
using distributed::remove_dangling_links;
} } // end namespace boost::graph
#endif // BOOST_PARALLEL_GRAPH_PAGE_RANK_HPP
|
