|| at Mar 2, 2011 at 10:08 pm
Could someone throw some light as to how intuitively fixed-type slots in
Hadoop have a negative impact of cluster utilization as mentioned in Arun's
It's pretty simple: in the course of time, a real cluster with a mix of
job types and submission times will have periods of high map-slot usage
(with reduce-slots available), and vice versa. Being able to assign machine
resources to whatever kind of task needs them at any given time means you're
not wasting those resources.
In short: map- vs. reduce-slots is an artificial and unnecessary restriction.