Re: [boost] Multicore-BGL

Subject: Re: [boost] Multicore-BGL
From: Jeremiah Willcock (jewillco_at_[hidden])
Date: 2010-05-30 19:56:39

On Sun, 30 May 2010, Giorgio Zoppi wrote:

> 2010/5/30 Jeremiah Willcock <jewillco_at_[hidden]>:
>> On Sun, 30 May 2010, Giorgio Zoppi wrote:
>>
>>> 2010/5/30 Adam Merz <adammerz_at_[hidden]>:
>>>>
>>>> Giorgio Zoppi <giorgio.zoppi <at> gmail.com> writes:
>>>>>
>>>>> we're intersted in how to optimize BGL on Multicore for adding our
>>>>> standard Bellman-Ford algorithms for optical networks.
>>>>> How do the current boost work on Multicore? BGL is thread-safe? Which
>>>>> parts are required to optimize?
>>>>
>>>> I believe what you're interested in is the Parallel BGL. See
>>>>
>>>> http://www.boost.org/doc/libs/release/libs/graph_parallel/doc/html/index.html
>>>> and
>>>> http://osl.iu.edu/research/pbgl/
>>>
>>> I skimmed the doc, however we're more fit in a multicore solution than
>>> a distributed process solution. I cannot find something about that in
>>> the doc. Any hints?
>>
>> We are in the process of adding shared-memory/hybrid parallelism to PBGL,
>> but that's still in a very early stage.  You can run multiple MPI processes
>> on the same machine (running them on different cores) and they will
>> communicate using shared memory.  One issue that comes up is that memory is
>> often the limitation for graph algorithms, not CPU performance; in that
>> case, multicores don't really do much for performance.  Are you using more
>> expensive algorithms that aren't memory intensive?  If so, there might be
>> ways to do simple things with existing BGL algorithms to make them run with
>> shared memory (for example, adding OpenMP pragmas).
>
> So if the graph is with sparse matrix, you could provide a more
> compact implementation
> and modify (parallelize) your algorithms.

I don't understand what you're saying.  What kinds of algorithms are you 
running on your graphs that you would like to parallelize?

-- Jeremiah Willcock