On Wed, Sep 7, 2011 at 9:26 PM, Avery Ching <ach...@apache.org> wrote:
> Haha, this really is turning into a movie =). I'll start warming up the
Yeah, I've got my co-workers wondering if I'm going to ship any actual
production code *inside* the company this week, at this rate (*shhhhh*
Dmitriy don't tell!)
> On 9/7/11 12:51 PM, Jake Mannix wrote:
> Maybe a few more examples would help? Cases where you want to do a BSP
> computation where the total sort (both the vertexes, and the edges for each
> vertex) is required, as is the random access nature of the Map?
> I think the range based examples are the ones that immediately come to
> mind. BSP for graph processing is still pretty new, and I have no idea what
> kind of interesting algorithms will be tried out on this platform. We are
> still exploring many possible algorithms to run.
Ok, cool. I can see how wanting flexibility is important.
> I think the idea was that after returning the map, users could directly
> manipulate the map of edges or use the interfaces, there should probably be
> a removeEdge() too. I'm starting to feel that we should remove Edge from
> the user perspective, just keep it internally only for the add edge
> requests. It just makes things a little more complex to the user (too many
> ways to do the same thing). Perhaps the interfaces you specified could hit
> most of the use cases (getTargetVertices(), getEdgeValue(), addEdge(), and
> removeEdge()). If there turns out to be a big need, we can always change it
> back to a SortedMap or something else more appropriate.
getTargetVertices(), getEdgeValue(), addEdge(), and removeEdge()
sound like the right level of flexibility while keeping the data
encapsulated (so you can try your block compression idea, I can try out
primitives, etc, but the interface remains the same).
Memory consumption (see above), these are aggregate members for all the
> Ok, I'll see what it looks like if this data is moved to something like a
> VertexState object attached to the GraphMapper, which all the vertexes can
> have a reference to.
> As I've thought more about primitives vs objects, I think the object
> flexibility is quite important. The page rank example could probably get
> away with primitives, but other algorithms will likely require objects for
> edge values, message values, and vertex values (i.e. maybe storing the
> inlinks, or a bunch of different values i.e. multiple personalized page
> ranks run simultaneously). I guess you're thinking that Giraph will have
> two separate implementations? One that is primitives based and the other
> that is object based?
I'm thinking that with the right interface (like discussed above), you can
have the same base interface, but yeah, for the particular case of
implementers of BaseVertex<I,V,E,M> where all of I,V, and E are wrappers of
primitives, that there is some nice memory savings that can be done by
keeping them primitive (and only instantiating objects / autoboxing when
accessing via the generic methods, when this is transiently done).
PageRank isn't the only example where I'd want to get the (suspected) perf
boost of using primitives, as most cases where I've dealt with graphs
everything gets normalized at some point - the input features all get
eventually turned into an edge "weight" of some kind, the vertexes
themselves maybe keep some small data with them, but the edges just look
like a target vertex id and an edge weight. For example, for social graphs,
you can imagine lots and lots of fancy data you associate with users (geo,
language, account freshness, recent text, topical interests, etc...), and
lots of things to associate with the edges (there are many ways users can
interact, beyond the explicit "x is connected to y" way), but when you want
to run some big monstrous computation, this data is condensed into some
fixed final "connection strength" combination of weights. On the other
hand, to actually *compute* that connection weight, maybe non-local
information gleaned from a graph algorithm would be nice. Similarly,
computing a nice big topic-sensitive pagerank might require a bunch of
topic-weight metadata at the vertexes.
I don't know why I'm arguing this - I agree with you, keeping the *ability*
to do object stuff is important, yes. I'm not advocating completely
primitivizing all of the base implementations. I'm just suggesting that it
be added, as that's a pretty common use case which could benefit from some
low-hanging fruit memory savings.