Cygwin sparse checkout degrades performance

Counter-intuitively, using sparse checkout in Cygwin degrades "status" times as status appears to "stat" non-existent files and directories.

To demonstrate, I created a repo with 100k random files in a dir/dir/dir/file structure (on a linux box -- to do this in Cygwin requires piping the result of "openssl rand" to "dos2unix" as the output contains "\r") and cloned in a Cygwin shell:

git init test cd test git commit --allow-empty -m 'Empty first commit' for i in {1..10}; do for j in {1..10000}; do file=$( openssl rand -hex 32 | sed 's,^\(.\)\(.\)\(.\),\1/\2/\3/,'); mkdir -p $( dirname $file ); echo $file > $file ; done & done; wait git add . git commit -m '100000 files' git gc --prune=now --aggressive

I then timed and plotted "git status" as sparse checkout step-wisely reduced the number of files in the working tree using the folllowing command:

( ( git status >& /dev/null; time -p git status > /dev/null ) |& sed -n '/real/{s/real/100000/p}'; git config core.sparseCheckout true; for i in $( seq 90000 -10000 10000 ) 1; do git ls-files | head -n $i | sed 's,^,/,' > .git/info/sparse-checkout; git read-tree -u -m HEAD; git status >& /dev/null; ( time -p git status > /dev/null ) |& sed -n "/real/{s/real/$i/p}"; done; echo '*' > .git/info/sparse-checkout; git read-tree -u -m HEAD; rm .git/info/sparse-checkout; git config --unset core.sparseCheckout ) | gnuplot -p -e "set terminal dumb; set xrange[] reverse; set style data dots; set nokey; plot '-' using 1:2"

Vertical bar is time in seconds, horizontal the number of files in the working tree after the sparse checkout.

Linux results (v2.1.0):
   0.45 
.+-----+------+-----+------+------+------+------+-----+------+-----++
        +      +      +     +      +      +      +      + +      +      +
| |
    0.4 ++ ++
| |
   0.35 ++ ++
| |
| |
    0.3 ++ .                                                           ++
| |
        | .                                                     |
   0.25 ++                  . .                                       ++
        |                                 . .                          |
| |
    0.2 ++                                              . .            ++
| |
   0.15 ++                                                           . +.
| |
        +      +      +     +      +      +      +      + +      +      +
    0.1 
++-----+------+-----+------+------+------+------+-----+------+-----++
      100000 90000  80000 70000  60000  50000  40000  30000 20000 10000    0

Cygwin results (v2.1.1):
   10 
++-----+------+------+------+------+------+------+------+------+-----++
      +      +      +      +      +      +      +      + +      +      +
| .
    9 ++ ++
| .      |
| |
    8 ++ .            ++
| |
| |
    7 ++ .                   ++
| |
| |
      |                                  . .                           |
    6 ++ ++
| |
| |
    5 ++ .                                                             ++
      .                    . .                                         |
      +      +      .      +      +      +      +      + +      +      +
    4 
++-----+------+------+------+------+------+------+------+------+-----++
    100000 90000  80000  70000  60000  50000  40000  30000  20000 10000    0

Linux times do what I expect/want (they get better as the number of working tree files decrease), but Cygwin does the opposite: the worst times are in a working tree with only 1 (sparse) file, and it's double where I started with no sparse checkout! I'd hoped sparse checkout would improve the too-slow status times when all files are present...

Looking at strace with a working tree consisting of a single (sparse) file suggests Cygwin is attempting to access the non-existent files and directories whereas Linux does not appear to do so. In fact, if I do nothing more than "mkdir -p $( git ls-files | cut -c1-5 | sort -u )" when looking at a single (sparse) file, I can drop status times below 3s, a 3-fold improvement and something at least better than where I started!

Is there a way I can get improved status times using sparse checkout with Cygwin?

Huh. The graphs (somehow) ended up incoherently reformatted... Sorry about that!

Here's the raw data after a second run:

Linux:
100000 0.49
  90000 0.27
  80000 0.27
  70000 0.28
  60000 0.23
  50000 0.21
  40000 0.21
  30000 0.19
  20000 0.19
  10000 0.16
      1 0.14

Cygwin:
100000 4.72
  90000 4.28
  80000 4.41
  70000 4.43
  60000 4.67
  50000 5.04
  40000 6.24
  30000 7.28
  20000 7.88
  10000 8.96
      1 9.43

On 12/24/2014 12:30 PM, Brian Ericson wrote: