Re: [Boost-users] [graph][parallel] Problem distributed property map synchronization

Subject: Re: [Boost-users] [graph][parallel] Problem distributed property map synchronization
From: Borja Miñano (bminyano_at_[hidden])
Date: 2013-03-11 06:07:35

Note that what I am trying to synchronize is an edge property map.
Maybe it is different and I am doing something wrong.
Are the edges crossing processors synchronized? For a bidirectional 
graph it is necessary if one want to use in_edges.

El 11/03/13 09:20, Borja Miñano escribió:
> Hello Nick,
>
> Oops, I attach it again.
> Anyway, I am using an bidirectional adjacency list, because I need to 
> know the in_edges.
> typedef boost::adjacency_list<vecS, distributedS<mpi_process_group, 
> vecS>, bidirectionalS, property<vertex_index_t, int>, EdgeProperties> 
> Graph;
>
> I have been looking in more detail and it seems that before the first 
> synchronize call, when the ghost values are required they are filled 
> correctly, but after then they never synchronize their values again. 
> Reading the documentation I saw that the first time one do get for a 
> ghost cell it is automatically synchronized. But, unfortunately, the 
> following times I try to get the values they don't do the 
> synchronization again.
>
> If using a bidirectional directed graph was a problem (as it seems for 
> undirected ones), is there another way to perform a manual 
> synchronization?
>
> El 08/03/13 19:18, Nick Edmonds escribió:
>> On Mar 8, 2013, at 10:27 AM, bminano wrote:
>>
>>> Hello,
>>>
>>> I need help. I have a distributed graph with a distributed property 
>>> map for
>>> an edge value.
>>> The problem is that the synchronize procedure does not synchronize 
>>> my ghost
>>> edges values.
>>>
>>> I have a loop that reads from the property, synchronizes, modifies 
>>> its value
>>> and synchronizes it.
>>> What I see is that the value is not being synchronized for the ghost 
>>> edges.
>>>
>>> for (int t = 0; t < 2; t++) {
>>>     myfile<<rank&lt;&lt;&quot;:Time &quot;&lt;&lt;t&lt;&lt;endl;
>>>
>>>     BGL_FORALL_VERTICES(v, g, Graph) {
>>>             for (boost::tie(ei_i,edge_end_i) = in_edges(v, g); ei_i 
>>> != edge_end_i;
>>> ++ei_i) {
>>>             remote_key_e k(rank, *ei_i);
>>>             myfile&lt;&lt;&quot;value 
>>> &quot;&lt;&lt;d_index[source(*ei_i,
>>> g)]&lt;&lt;&quot; -> "<<d_index[target(*ei_i, g)]&lt;&lt;&quot; valor
>>> &quot;&lt;&lt;get(d_ecash, k)&lt;&lt;endl;
>>>             }
>>>     }
>>>
>>>     synchronize(d_ecash);
>>>     BGL_FORALL_VERTICES(v, g, Graph) {
>>>             for (boost::tie(ei,edge_end) = out_edges(v, g); ei != 
>>> edge_end; ++ei)
>>> {
>>>             remote_key_e k(rank, *ei);
>>>             put(d_ecash, k, t + 1);
>>>         }
>>>     }
>>>               synchronize(d_ecash);
>>> }
>>>
>>> Could anyone tell me what I am doing wrong?
>>> I attach the full code if someone want to look at them.
>>>
>>> &lt;nabble_a href=&quot;sync.cpp&quot;>sync.cpp
>> Your mail client seems to have mangled your code pretty badly and I 
>> don't see an attachment, but the first question would be what graph 
>> type are you using?  If you're using the undirected graph then 
>> there's no support in the PBGL for automatically synchronizing the 
>> values between the two local copies of the edge stored at the owner 
>> of the source and target vertex respectively.  It could be added but 
>> would require O(E) additional storage to map from an edge to it's 
>> corresponding anti-edge as well as some additional communication to 
>> update both copies (the complexity of which would vary depending on 
>> the consistency model you wanted to support).
>>
>> Cheers,
>> Nick
>> _______________________________________________
>> Boost-users mailing list
>> Boost-users_at_[hidden]
>> http://lists.boost.org/mailman/listinfo.cgi/boost-users
>>
>
>

-- 
Un saludo,
Borja Miñano
IAC3 - Universitat de les Illes Balears
ParcBit - Edifici 17 (Disset); Local D7
Cra. Valldemossa km. 7,4
E-07121 Palma de Mallorca. Balears. Spain.
Phone: [phone]