@Deepak U​ 

It's hard to find an answer here without the proper monitoring of the entire process.

First of all, I would check how the workers behave during runtime (Ganglia - what's the overall usage of the workers, where's the bottleneck. It could be even more efficient to have a smaller number of more powerful nodes than multiple less powerful ones.

Second thing - networking. Maybe there's a bottleneck somewhere in there? Worth checking.

Third thing - Postgres instance - monitoring during the runtime.

Fourth thing - worth considering - exporting the data to Parquet files into storage, then using COPY FROM on Postgres to import the data. It's usually faster than using JDBC.