Why are the "Busy" portions of the bars higher on the right?
thunder
Correct me if I'm wrong. When applying cost-zones assignment which is a semi-static assignment, in order to achieve better temporal locality, the particles need to be reassigned to the cores periodically. I think it is the reassignment of all the particles as well as the corresponding data redistribution that causes the "Busy" bars higher than static assignment.
jmc
@thunder Shouldn't that be recorded as "Synch" rather than "Busy" time? I thought "Busy" was only the time you're spending performing the math of the actual algorithm, and "Synch" would include any work reassignment.
paracon
@jmc, I think sync just means that you are ensuring all threads have completed certain set of operations. Workload distribution should be considered in busy time since the processors are actively performing some operation to divide tasks.
Levy
@jmc, Performing tree traversal, reducing and repartition the calculation is much more examples than a single randomized shuffle. This is the execution overheads introduced on the right graph.
Why are the "Busy" portions of the bars higher on the right?
Correct me if I'm wrong. When applying cost-zones assignment which is a semi-static assignment, in order to achieve better temporal locality, the particles need to be reassigned to the cores periodically. I think it is the reassignment of all the particles as well as the corresponding data redistribution that causes the "Busy" bars higher than static assignment.
@thunder Shouldn't that be recorded as "Synch" rather than "Busy" time? I thought "Busy" was only the time you're spending performing the math of the actual algorithm, and "Synch" would include any work reassignment.
@jmc, I think sync just means that you are ensuring all threads have completed certain set of operations. Workload distribution should be considered in busy time since the processors are actively performing some operation to divide tasks.
@jmc, Performing tree traversal, reducing and repartition the calculation is much more examples than a single randomized shuffle. This is the execution overheads introduced on the right graph.