Hello, I apologize if this discuss should be directed to a hadoop map reduce forum, however, I have some concern that my problem may be with my use of accumulo.
I have a map reduce job that I want to run over data in a table. I have an index table and a support table which contains a subset of the data in the index table. I would like to map reduce over the support table on my small 4 node cluster. I have written a map reduce job that uses the AccumuloRowInputFormat class and sets the support table as its input table. In my mapper, I read in a row of the support table, and make a call to a static function which pulls information out of the index table. Next, I use the data pulled back from the function call as input to a call to an external .so file that is stored on the name node. I then make another static function call to ingest the new data back into the index table. (I know I could emit this in the reduce step, but what I'm ingesting is formatted in a somewhat complex java object and I already had a static function that ingested it the way I needed it.) My reduce step is completely empty. I output print statements from my mapper to see my progress. The problem that I'm getting is that my entire job appears to run in sequence not in parallel. I am running it from the accumulo master on the 4 node system. I realized that my support table is very small and was not being split across any tables. I am now presplitting this table across all 4 nodes. Now, when I run the map reduce job it appears that 4 separate map reduce jobs run one after each other. The first map reduce job runs, gets to 100%, then the next map reduce job runs, etc. The job is only called once, why are there 4 jobs running? Why won't these jobs run in parallel? Is there any way to set the number of tasks that can run? This is possible from the hadoop command line, is it possible from the java API? Also, could my problem stem from the fact that during my mapper I am making static function calls to another class in my java project, accessing my accumulo index table, or making a call to an exteral .so library? I could restructure the job to avoid making static function calls and I could write directly to the Accumulo table from my map reduce job if that would fix my problem. I can't avoid making the external .so library call. Any help would be greatly appreciated. Thanks, Duane
