A summary from the course, the idea of processing local data is not to reduce latency (as it can be hidden), but to reduce the bandwidth usage.
emt
The first idea here was to just add all the lines to the work queue and have each processor take from the queue. If some node grabbed a block that was on another node, then we would send the block over to that node to be processed. However, the second idea takes advantage of locality by saying if the data is on some node, we will run the computation for that data on the same node. In doing so, we reduce the bandwidth we need so that we don't have to pay the cost of interconnect traffic.
atadkase
Even if the data is processed initially in the same node, a need for communication arises after the first stage of calculations are done.
A summary from the course, the idea of processing local data is not to reduce latency (as it can be hidden), but to reduce the bandwidth usage.
The first idea here was to just add all the lines to the work queue and have each processor take from the queue. If some node grabbed a block that was on another node, then we would send the block over to that node to be processed. However, the second idea takes advantage of locality by saying if the data is on some node, we will run the computation for that data on the same node. In doing so, we reduce the bandwidth we need so that we don't have to pay the cost of interconnect traffic.
Even if the data is processed initially in the same node, a need for communication arises after the first stage of calculations are done.