If strictly positive, all slices switch to a generation sequentially, and we wait this time in seconds between two slices. This spreads the temporary memory consumption to avoid large memory spikes and swap out.