Re: [Boost-users] Parallel BGL status in Boost

On Sep 4, 2009, at 8:24 AM, Mathieu Malaterre wrote:

On Thu, Sep 3, 2009 at 10:37 PM, Nick Edmonds wrote:

...

Yep, PBGL as in the distributed data structures, MPI infrastructure, and distributed algorithms is now in Boost.

Cool, I did not see it in the changelog, so I was confused.

...

Yep, that webpage is desperately out of date, I'm a terrible webmaster. Updating it (and the attendant performance numbers) has been on my todo list for ages. I'd make more promises, but they'd probably sound hollow now. So when I actually do it, it will be a pleasant surprise.

So that it serve as reference to anyone reading this. The documentation have been updated, simply not uploaded neither on boost.org page nor on the old PBGL page. It is available locally as:

file:///path/to/boost/libs/graph_parallel/doc/html/index.html

Yep, 1.40 is unfortunately missing the proper redirect in libs/ graph_parallel to libs/graph_parallel/doc/html/index.html. I've fixed this in the trunk. When 1.41 goes out hopefully the docs will show up on boost.org, this was due to my ignorance w.r.t. how docs got to the webpage during the release process.

...

Structural information about the graph is only available on the process that owns that portion of the graph. Currently PBGL distributes the graph using a row-wise decomposition of the adjacency matrix. This means that calling vertices(g) returns the local vertex set owned by the calling process. The result of calling edges(g) depends on the directed category of the edge, but basically iterates over incident edges. Attempting to access structural information that is not local to the calling process, i.e. calling out_edges(v, g) where 'v' is not owned by the calling process is invalid and results in undefined behavior.

ok that clears things quite a bit !

...

If you want more details, to know why things work the way they do now, or how they might work in the future, feel free to ask away.

'in the future', what do you mean ? What are the actual plans ?

Well, there are lots of plans... Hybrid parallelism (distributed memory + threads), 2-dimensional decomposition for graphs (as opposed to the current row-wise decomposition), and alternates to MPI as a communication infrastructure amongst many less important plans (and the thesis project I'm keeping under my hat because I haven't published enough about it to claim it yet). There's also a very nascent project to do multi-core/highly multi-threaded BGL, starting from the BGL sources as opposed to PBGL. That's the short list, as (if?) we get more people working on the project there's lots more to be done as well. Thanks, Nick

Thanks again, -- Mathieu _______________________________________________ Boost-users mailing list Boost-users@lists.boost.org http://lists.boost.org/mailman/listinfo.cgi/boost-users