Would reducing the output from the map tasks solve the problem ? i.e. are reducers slowing down because a lot of data is being shuffled ?
If that's the case, you could see if the map output size will reduce by using the framework's combiner or an in-mapper combining technique.
On Wed, Oct 3, 2012 at 6:34 AM, Huanchen Zhang <email@example.com>
I have a small portion of map tasks whose output is much larger than others (more spills). So the reducer is mainly waiting for these a few map tasks. Is there a good solution for this problem ?