Subject: Re: [boost] [graph_parallel] Is a distributed directed edge list possible?
From: Aydin Buluc (abuluc_at_[hidden])
Date: 2011-01-05 15:44:01
Dave Abrahams <dave <at> boostpro.com> writes:
> At Wed, 5 Jan 2011 13:35:59 -0500 (EST),
> Jeremiah Willcock wrote:
> > On Mon, 3 Jan 2011, caustik wrote:
> > > Is it currently possible to represent a highly connected graph where any
> > > given vertex may contain more directed edges than can be held in a single
> > > process' OutEdgeListS?
> > >
> > > I see that the distributed version of adjacency_list allows vertices to be
> > > distributed across multiple processes in a process group, but I don't see
> > > the ability to allow directed edges to be distributed across multiple
> > > processes. Is this an intentional limitation?
> > No, it is not possible. PBGL currently distributes graph vertices (a
> > so-called 1-D distribution) and keeps each edge with its source
> > vertex. To do what you are describing, we would need to implement 2-D
> > (edge-based) distributions, which do not fit into the current PBGL
> > design.
> Another thought I had, if graph connectivity is really super-high, is
> that you might be able to use the inverse graph representation.
Caustik, what kind of kernels and/or algorithms
you want to run on the distributed graph?
Combinatorial BLAS has evolved quite a bit (with many more primitives,
more flexible vector distributions, a few more example algorithms, and
numerous bug fixes) from the 1.0 alpha version that is on the web.
Until the next release (likely to happen in a month or two), I can share
the active branch with you if your application seems to fit well. If it
doesn't, it'll still be very useful for us to see what kind of primitives
we do not support so that we can include them in the future.
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk