Figure 3c presents performance in terms of number of nodes processed per second. It is interesting to add how the processing of nodes is performed for this problem. In terms of data layout, the N-Queens instance is rather small:17 variables which represents a node size of 136 bytes. On the other hand, the total number of nodes processed is quite
large (757914186), at a very high rate (e.g. around 40 million nodes per second with 256 cores). In fact, looking at the time spent on the main steps of the solving procedure (propagation,splitting and searching) we observe the following distribution:
propagation takes 48%, splitting 10% and restoring takes 42% of the total time. This distribution is constant and independent of the number of cores used. A large portion is spent on retrieving stores from the pool (restoring) which includes
releasing work. Since processing a node is a fast operation, the overhead of releasing is more evident.