Perhaps we could combines Ideas 1 and 2. Specifically, all CPUs first try to read local data, if there is any left for processing. If the CPU fails to get more local work (and is thus idle) it can begin stealing work from oversubscribed nodes.
yuel1
@scedarbaum that also depends on how the data is distributed and whether the mapper takes uniform time on each line.
Perhaps we could combines Ideas 1 and 2. Specifically, all CPUs first try to read local data, if there is any left for processing. If the CPU fails to get more local work (and is thus idle) it can begin stealing work from oversubscribed nodes.
@scedarbaum that also depends on how the data is distributed and whether the mapper takes uniform time on each line.