Boost图形库：在并行模式下从adjacency_list读取时访问冲突

Question

I am having trouble with accessing edge properties from a boost adjacency_list in parallel mode. I am using boost 1.54.0 and OpenMP within C++ code. My problem boils down to the following sandbox example:

#include <boost/graph/adjacency_list.hpp>

struct Knoten {int Knoten_Property};
struct Pfeil {int Pfeil_Property};
typedef boost::adjacency_list<boost::vecS, boost::vecS, boost::bidirectionalS, Knoten, Pfeil> Graph;
typedef boost::graph_traits<Graph>::edge_descriptor Edge;
typedef boost::graph_traits<Graph>::edge_iterator Edge_iter;

Graph G;

// […] 
// some initializations for the graph. 
// Test graph contains about  20,000 edges

#pragma omp parallel num_threads(4) 
{
    #pragma omp for private(i) 
    for(int i = 0; i < 100; i++) 
    {
        for(pair<Edge_iter, Edge_iter> ei = edges(G); ei.first != ei.second; ++ei.first)
        {
            for (int i = 0; i < 100000; ++i)
            {
                Edge E = *ei.first;
                int my_test = G[E].Pfeil_Property; // (*)
            }
        }
    }
}

When running this code (parallel mode, 4 threads), I obtain access violations in reading at line (*). It seems like I cannot access the property of an edge simultaneously from different threads, though this is a read-only access. Any help on this issue will be appreciated. Thank you very much!

Here is the entire error message (in German); it reads: "Exception (first chance) at ... Access Violation when reading at position ... Unhandled exception at ..."

Eine Ausnahme (erste Chance) bei 0x000000014042237f in MyApplication.exe: 0xC0000005: Zugriffsverletzung beim Lesen an Position 0x000000000257ac58.
Unbehandelte Ausnahme bei 0x000000014042237f in MyApplication.exe: 0xC0000005: Zugriffsverletzung beim Lesen an Position 0x000000000257ac58.

Answer 1

There must be something else that is tripping you up that isn't visible in the snippet. I was able to get this to run with boost graph and openMP

Update: I changed to code to explicitly show independent work on each iteration of the outer loop and to access cout only once in each loop.

#include <omp.h>
#include <iostream>
#include <boost/graph/adjacency_list.hpp>
#include <vector>
#include <string>

using namespace std;

struct Knoten {int Knoten_Property;};
struct Pfeil {int Pfeil_Property;};
typedef boost::adjacency_list<boost::vecS, boost::vecS, boost::bidirectionalS, Knoten, Pfeil> Graph;
typedef boost::graph_traits<Graph>::edge_descriptor Edge;
typedef boost::graph_traits<Graph>::vertex_descriptor Vertex;
typedef boost::graph_traits<Graph>::edge_iterator Edge_iter;

string convertInt(int number)
{
   //http://www.cplusplus.com/forum/beginner/7777/
   stringstream ss;//create a stringstream
   ss << number;//add number to the stream
   return ss.str();//return a string with the contents of the stream
}


int main()
{
    Graph G;

    vector<Vertex> verts;

    //add 7 vertices
    for(size_t i = 0; i < 7; ++i){
        Vertex v = add_vertex(G);
        verts.push_back(v);
    }

    add_edge(verts.at(0),verts.at(1),G);
    add_edge(verts.at(0),verts.at(3),G);
    add_edge(verts.at(1),verts.at(2),G);
    add_edge(verts.at(1),verts.at(3),G);
    add_edge(verts.at(1),verts.at(4),G);
    add_edge(verts.at(2),verts.at(4),G);
    add_edge(verts.at(3),verts.at(4),G);
    add_edge(verts.at(3),verts.at(5),G);
    add_edge(verts.at(4),verts.at(5),G);
    add_edge(verts.at(4),verts.at(6),G);
    add_edge(verts.at(5),verts.at(6),G);

    #pragma omp parallel num_threads(4) 
    {
        #pragma omp for private(i) 
        for(int i = 0; i < 10; i++) 
        {
                int threadId = omp_get_thread_num();
        string greeting = "hello from ";
              greeting += convertInt(threadId);
        greeting += " on loop variable ";
        greeting += convertInt(i);
        greeting += " graph edges ";
        greeting += convertInt(num_edges(G));
        greeting += "\n";
        cout << greeting;
        //cout << "edges " << num_edges(G) << endl;
                for(pair<Edge_iter, Edge_iter> ei = edges(G); ei.first != ei.second; ++ei.first)
                {
                    for (int i = 0; i < 10; ++i)
                    {
                     Edge E = *ei.first;
                     int my_test = G[E].Pfeil_Property; // (*)

                  }
              }
          }
    }

    cout << "after join " << num_edges(G) << endl;
    cout << "after join " << num_vertices(G) << endl;

    return 0;
}

Compile with this: BOOST_INC is the root to boost, /usr/lib/gcc ... is the location of omp.h

g++ -O -fopenmp -I $BOOST_INC -I . -I /usr/lib/gcc/x86_64-redhat-linux/4.4.7/include/ -c -g  graphOmp.cpp
g++ -g graphOmp.o -o graphOmp -fopenmp

Exporting OPM_NUM_THREADS

export OMP_NUM_THREADS=4

output looks like:

hello from 3 on loop variable 9 graph edges 11
hello from 2 on loop variable 6 graph edges 11
hello from 0 on loop variable 0 graph edges 11
hello from 0 on loop variable 1 graph edges 11
hello from 0 on loop variable 2 graph edges 11
hello from 1 on loop variable 3 graph edges 11
hello from 1 on loop variable 4 graph edges 11
hello from 1 on loop variable 5 graph edges 11
hello from 2 on loop variable 7 graph edges 11
hello from 2 on loop variable 8 graph edges 11
after join 11
after join 7

Here we see that each thread does part of the for iteration. After the threads are joined there is only one thread of execution. You may need to post the rest of the code.