Comment by feelin_googley | Hacker News Reader

Comment by feelin_googley | Hacker News Reader

Parent Full thread

When sort is used on really large files, it will automatically attempt to use disk, putting temp files in TMPDIR. This can be really slow.

To overcome the slowdown of disk I/O, perhaps a workaround could be to use mfs or tmpfs, maybe something like:

   mkdir /dir
   mount -t tmpfs tmpfs /dir
   TMPDIR=/dir sort -k2 -k1 -k3 contacts.tsv
   TMPDIR=/dir sort -k1 -k2 -k3 contacts.tsv

Personally, I gave up on sort for large files and use k/kdb+. I suspect it is faster for sorting than sort or the Go libraries, but I could be wrong.

For a dataset larger than physical memory, using a memory filesystem like tmpfs for the merge stage will either swap (|tmpfs| < |ram|) or deadlock (|tmpfs| >= |ram|).

Instead, your best bet in that case is to give sort as much physical memory as you can spare:

    sort -S 95% -k1 huge.tsv

Extra disk I/O is inevitable since your dataset doesn't fit in memory. At least during a merge sort your disk reads will be O(N) and sequentially ordered.

Note: in the special case that your dataset is slightly larger than physical memory, splitting it up in advance such that one of the `sort -m` input files lives on a tmpfs should indeed be faster.

Other things to check out if you need Very Fast Large Sorts:

- Use `sort --parallel=N` to use multiple cores. By default it only uses 1.

- Use `sort --batch-size=NMERGE` to increase the number of files merged at once. Otherwise you may be doing more mergesort stages than are necessary.

Reply on news.ycombinator.com