Have you looked at graph processing for Hadoop? Like Hama ( http://hama.apache.org/) or Giraph (http://incubator.apache.org/giraph/). I can't say for sure it would help you but it seems to be in the same problem domain.
With regard to the chaining reducer issue this is indeed a general implementation decision of Hadoop 1. >From a purely functional point of view, regardless of performance, I guess it could be shown that a map/reduce/map can be done with a reduce only and that a sequence of map can be done with a single map. Of course, with Hadoop the picture is bit more complex due to the sort phase. map -> sort -> reduce : operations in map/reduce can not generally be transferred due to the sort 'blocking' them when they are related to the sort key reduce -> map : all operations can be performed in the reduce So map -> sort -> reduce -> map -> sort -> reduce -> map -> sort -> reduce can generally be implemented as map -> sort -> reduce -> sort -> reduce -> sort -> reduce if you are willing to let the possibility of having different scaling options for maps and reduces And that's what you are asking. But with hadoop 1 the map phase is not an option (even though you could use the identify but that's not a wise option with regards to performance like you said). The picture might be changing with Hadoop 2/YARN. I can't provide the details but it may be worth it to look at it. Regards Bertrand On Fri, Oct 5, 2012 at 8:02 PM, Jim Twensky <[email protected]> wrote: > Hi Harsh, > > The hidden map operation which is applied to the reduced partition at > one stage can generate keys that are outside of the range covered by > that particular reducer. I still need to have the many-to-many > communication from reduce step k to reduce step k+1. Otherwise, I > think the ChainReducer would do the job and apply multiple maps to > each isolated partition produced by the reducer. > > Jim > > On Fri, Oct 5, 2012 at 12:54 PM, Harsh J <[email protected]> wrote: > > Would it then be right to assume that the keys produced by the reduced > > partition at one stage would be isolated to its partition alone and > > not occur in any of the other partition outputs? I'm guessing not, > > based on the nature of your data? > > > > I'm trying to understand why shuffling is good to be avoided here, and > > if it can be in some ways, given the data. As I see it, you need > > re-sort based on the new key per partition, but not the shuffle? Or am > > I wrong? > > > > On Fri, Oct 5, 2012 at 11:13 PM, Jim Twensky <[email protected]> > wrote: > >> Hi Harsh, > >> > >> Yes, there is actually a "hidden" map stage, that generates new > >> <key,value> pairs based on the last reduce output but I can create > >> those records during the reduce step instead and get rid of the > >> intermediate map computation completely. The idea is to apply the map > >> function to each output of the reduce inside the reduce class and emit > >> the result as the output of the reducer. > >> > >> Jim > >> > >> On Fri, Oct 5, 2012 at 12:18 PM, Harsh J <[email protected]> wrote: > >>> Hey Jim, > >>> > >>> Are you looking to re-sort or re-partition your data by a different > >>> key or key combo after each output from reduce? > >>> > >>> On Fri, Oct 5, 2012 at 10:01 PM, Jim Twensky <[email protected]> > wrote: > >>>> Hi, > >>>> > >>>> I have a complex Hadoop job that iterates over large graph data > >>>> multiple times until some convergence condition is met. I know that > >>>> the map output goes to the local disk of each particular mapper first, > >>>> and then fetched by the reducers before the reduce tasks start. I can > >>>> see that this is an overhead, and it theory we can ship the data > >>>> directly from mappers to reducers, without serializing on the local > >>>> disk first. I understand that this step is necessary for fault > >>>> tolerance and it is an essential building block of MapReduce. > >>>> > >>>> In my application, the map process consists of identity mappers which > >>>> read the input from HDFS and ship it to reducers. Essentially, what I > >>>> am doing is applying chains of reduce jobs until the algorithm > >>>> converges. My question is, can I bypass the serialization of the local > >>>> data and ship it from mappers to reducers immediately (as soon as I > >>>> call context.write() in my mapper class)? If not, are there any other > >>>> MR platforms that can do this? I've been searching around and couldn't > >>>> see anything similar to what I need. Hadoop On Line is a prototype and > >>>> has some similar functionality but it hasn't been updated for a while. > >>>> > >>>> Note: I know about ChainMapper and ChainReducer classes but I don't > >>>> want to chain multiple mappers in the same local node. I want to chain > >>>> multiple reduce functions globally so the data flow looks like: Map -> > >>>> Reduce -> Reduce -> Reduce, which means each reduce operation is > >>>> followed by a shuffle and sort essentially bypassing the map > >>>> operation. > >>> > >>> > >>> > >>> -- > >>> Harsh J > > > > > > > > -- > > Harsh J > -- Bertrand Dechoux
