I thought that since the purely recursive implementation of mergesort didn’t show weird, abrupt performance drops, traversing linked lists (no matter what order the nodes appeared in memory) also did not show abrupt performance drops, and that node size in memory caused different performance oddities, that the iterative implementation’s memory access patterns might be the cause.

When my kids were small, I would tell them about famous philosophical or mathematical paradoxes.

I checked the effect of list element size on my implementation of mergesort of a linked list.

A museum and exhibit you should visit

Denver Museum of Nature and Science currently has an absolutely fascinating, traveling exhibit about Orcas.

It occurred to me that benchmarking traversing very long linked lists end-to-end might provide some clues.

Hailstones from 2016-07-15T17:14:00, Baker neighborhood, Denver, Colorado.

Based on the effects of linked list re-use, I checked if using a linked list with in-memory-address-order made a difference.

After I wrote a recursive merge sort, I reviewed it to ensure it wasn’t doing extra work. From that perspective, I noticed that for each of the 10 list sorts my benchmark program made for a given length linked list, it created a new linked list.

I thought that I could make the overall benchmark run faster by not creating a new, very large, linked list for every run, but rather re-using the existing list. I hoped to have the benchmark program use less time between the ten measured sort algorithm executions, not make the sorting faster. I surprised myself by getting the opposite effect.

I wrote a traditional, top-down, recursive merge sort. The idea was to see if a different implementation exhibited the same dramatic, abrupt slowdowns. If the different implementation showed the same jumps in timings, the problem is with the operating system, the language run time, or the memory hardware. I know I said I’d ruled out the underlying operating system previously, but I was still suspicious.