During each Verlet list rebuild, particle positions of neighboring cells of every cell are first copied into contiguous buffers (c_pos). The buffers are then iterated over to determine the particle pairs that are within cutoff distance. This results in more cache-friendly access since the particle positions are stored contiguously instead of being stored in different memory locations (in the case of particles in std::vector< Cell >) and separated by other particle information. The cost of copying the data initially is amortized by the number of times the buffer is iterated during the cutoff check (equal to the number of particles in the cell).

I've added a test and comparison script that runs a Lennard-Jones simulation (testsuite/verlet_list_buffer/verlet_list_buffer.py). With one process on a Broadwell node of Mogon 2, compiled with gcc 7.3.0 and -O3 optimization, the method results in around 2x speedup for the Resort part of the simulation.