You're welcome. It's mostly an adaptation of [cligen/examples/linect.nim](https://github.com/c-blake/cligen/blob/master/examples/linect.nim) or [adix/tests/wf.nim](https://github.com/c-blake/adix/blob/master/tests/wf.nim). Relative to your book's "read whole file, then chunk" idea, my `nSplit` does one less "pass over the data" by leveraging mmap/random access to the file. Both approaches in the large rely on statistical regularity of record sizes to avoid switching overhead from more fine-grained parallel dispatch.
The parsing in general (from Linux `perf`) seems ~50% just `memchr` (avx2 vectorized). That is usually close to the best one can get, though in this case it is only about 4 GB/sec (about 10% of my DIMM bw). It probably bears re-mention here that pre-parsing into random access binary formats pays more perf dividends for re-analysis of the same data, making "parsing performance" a less than perfect example. DB systems have done binary formats for 60 years with query languages "for non-programmers" (that "for" has always struck me as a real stretch..LOL). For programmer consumers, you can get most/all that perf with little work from something like [nio](https://github.com/c-blake/nio) or Vindaar's [nimHDF5](https://github.com/Vindaar/nimhdf5). Irregularities of DNA strings (that began this discussion) do tend to limit those benefits.
