I'm not going to load a database, because the results of the sort will be used just once, and as a practical matter may be passed in a pipe without ever hitting the filesystem. For testing, there's an output file, but just for testing, and to make the results more generally relevant to anyone else who might read this.

My sort times are generally within a factor of 2 of the cost of copying the file to temp and then to the output. So thrashing and computing are not horrible and I'm not going to write a separate sort or any part of it because it will take too long to get it right, and I'm not going to use multiple invocations of sort(1) because the disk I/O will clearly eat any benefits.

I've re-run my timing scripts on the small test file, and some comments I made earlier have to be corrected. The differences between runs are not that alarming after all, and are easily explained by differences in other competing activities on the same machine. My speedups so far are more modest than I thought, but --parallel=4 really does give me 20%, and there's about another 20% available from jiggering parameters.

Running after a fresh boot, I noticed some things that surprised me, though perhaps they should not have. By the time testing is done, the kernel has filled 64GB of memory, mostly with "cached" blocks, and has swapped out a little over 3 MB of memory. I presume it's swapping idle daemons. So these results will not scale up for files large enough to do a complete cache wipe.

I've pretty much determined that the main thing to avoid is getting more than one merge pass on the temporaries. I think it's time to try just a few things with my TB-sized things, because I know those were doing at least 2 extra passes with the default parameters. It took forever. I think the defaults are 4GB buffers (1/8 real memory) and merges of 16 files. The buffers seem to have a lot of overhead, so the temp files are smaller than you might expect. On a 1TB file, that will be roughly 500 2-GB temp files, and 3 levels of merge. The question is: given a choice, is it better to use a bigger buffer, or a wider merge? I'm betting 4GB buffers are too big, but I need to do some testing.