...
- If user specified partition function, internally, "partitioning job" is ran as a pre-processing step.
- Each task of "partitioning job" reads its assigned data block and rewrite them to particular partition files.
- Job Scheduler assigns partitions to proper task based on partition ID.
- After pre-partitioning done, launch the BSP job.
Other suggestions are:–
- Instead of running a separate job, we inject a partitioning superstep before the first superstep of the job. (This has a dependency on the Superstep API) For graph jobs, we can configure this partitioning superstep class specific to graph partitioning class that partitions and loads vertices. - Suraj Menon
- Since scaling the number of BSP tasks between supersteps in single job, is not possible, the key is "how to launch the number of tasks differently with the number of file blocks". - Edward J. Yoon
- The partitions instead of being written to HDFS, which is creating a copy of input files in HDFS Cluster (too costly I believe), should be written to local files and read from. - Suraj Menon
- If we want to run multiple jobs on same input, for example, friendship graph is input and want to run PageRank job and SSSP job, ...., etc., reuse of partitions should be considered. If partitions are written on local fs, metadata should be managed to reuse them. - Edward J. Yoon
- If we want to run multiple jobs on same input, for example, friendship graph is input and want to run PageRank job and SSSP job, ...., etc., reuse of partitions should be considered. If partitions are written on local fs, metadata should be managed to reuse them. - Edward J. Yoon
...