OpenMP schedule dynamic for parallelForEdges

[angriman]

I recently used parallelForEdges to generate random weights on large graphs (1-2 billion edges). On a 36-cores machine, the performance was very poor (it took >1h) due to bad load balancing (during the last 30 minutes only a few threads were running).
The loop iterates over the nodes in parallel and each thread iterates over the neighbors of the current node, so the workload of a thread depends on the degrees of the nodes it processes (which might vary a lot).

Changing the schedule from guided to dynamic improved the load balancing and the performance significantly (the time dropped to a few minutes), which makes sense because with guided some thread might end up processing many more edges than others. So I think we should use dynamic as the default scheduling strategy for this loop.

[hmeyerhenke]

Have you tried other use cases with benchmarking results as well? Back then, we used guided because it was slightly faster than dynamic for the use case back then.

[angriman]

I don't have detailed benchmark data. I can imagine that, on smaller graphs than the ones I am considering, the dynamic schedule has more synchronization costs so it performs slightly worse (but I am not sure about that).
It's probably worth investigating more on this aspect.

[hmeyerhenke]

I am not opposed to the change in principle. I would simply suggest to investigate and benchmark more use cases. Probably something for Lucas, right?