Feature #6256

Slightly improve ruby_qsort performance

Added by Martin Bosslet about 2 years ago. Updated over 1 year ago.

[ruby-core:44136]
Status:Assigned
Priority:Low
Assignee:Martin Bosslet
Category:core
Target version:next minor

Description

Hi all,

I think I may have found a way to slightly improve the performance of ruby_qsort.
Quicksort running time is slightly decreased if the recursion depth (or in our
case, rather iteration depth) is bounded by leaving sub problems larger than or
equal to some cutoff bound k untouched and running Insertion Sort on these small
sub problems to finalize the sorting.

I experimented with this, but to no avail, only marginal improvements if any. Then
I remembered that instead of running Insertion Sort on each sub problem, it is
equivalent in terms of running time to run one single Insertion Sort on the whole
nearly sorted array as a final step. And in practice, this turns out to run faster
than without the optimization. In my tests, execution time dropped to about 95% on
average with an optimal cutoff (64-bit Fedora 15) [1].

Now this ain't the world - but it is faster, and I could very well imagine that there
is still room for improving my code. In my tests, the optimal cutoff seems to be ~13
for Integers and ~8 for Strings and Symbols. I imagine the more costly the comparisons,
the lower will be the optimal cutoff. I've tested only on 64 Bit yet, but I will do so
for 32 Bit, too.

If it turns out that this runs faster regardless of the architecture in use, with an
optimal cutoff yet to be determined, do you think this would be a useful addition?

I have attached a C extension for testing and discussing, it's mostly a one-to-one copy of
the code in util.c. I just added mmassign and insertionsort plus the few lines that respect
the cutoff in rqsort (had to rename it, otherwise it would collide with the real "ruby
qsort").

-Martin

[1] https://github.com/emboss/hybridsort/blob/master/hybrid-sort-results

hybridsort.tar.gz (3.93 KB) Martin Bosslet, 04/05/2012 10:02 AM

rubyqisort.patch Magnifier (5.04 KB) Martin Bosslet, 04/07/2012 07:36 AM

rubyqisort4.patch Magnifier (3.66 KB) Martin Bosslet, 04/07/2012 10:35 AM

rubyqisort5.patch Magnifier (3.69 KB) Martin Bosslet, 04/07/2012 11:01 AM

History

#1 Updated by Martin Dürst about 2 years ago

I haven't looked at the sources, but this exact optimization is very well known in the literature, and I'm showing it to my students in a lecture on algorithms and data structures every year (see e.g. http://www.sw.it.aoyama.ac.jp/2011/DA/programs/6qsort.rb, sorry, comments are in Japanese). I'm quite a bit surprised that this optimization isn't yet made in Ruby. The only potential problem is that the final insertion sort may mask errors in the quicksort, but as Ruby's current quicksort is working well, that should not be a problem. And a cutoff of about 10 is what's generally claimed to be a good value. The speed gains/losses when shifting the cutoff value by just a little bit shouldn't be very big.

So I'm definitely encouraging adoption of this improvement.

#2 Updated by Yusuke Endoh about 2 years ago

  • Status changed from Open to Assigned
  • Assignee set to Martin Bosslet

Sound good. I agree, too.

Please create a patch and show benchmark.

Yusuke Endoh mame@tsg.ne.jp

#3 Updated by Martin Bosslet about 2 years ago

I added a patch, together with benchmarks.

Here's the results on my machine without the optimization:

$ ruby driver.rb -e ruby -p bm_sort -r 10

benchmark results:
name ruby 2.0.0dev (2012-04-05 trunk 35241) [x8664-linux]
sort
int 0.378
sortintcustom 0.327
sortintrev 0.111
sortintsorted 0.100
sortstrrnd 0.107

and here with the patch applied:

$ ruby driver.rb -e ruby -p bm_sort -r 10

benchmark results:
name ruby 2.0.0dev (2012-04-05 trunk 35241) [x8664-linux]
sort
int 0.367
sortintcustom 0.332
sortintrev 0.229
sortintsorted 0.218
sortstrrnd 0.106

From running them several times, my impression is that it's definetely
faster for sorting Fixnums, but there's no real improvement as soon
as we compare Strings or use a custom comparison, as I assumed before,
the costlier the comparison becomes the less advantage we get, at least
that's how it seemed in my tests.

Another fact is that we definitely lose in the "already ordered" cases.
The existing Quicksort does a linear check for "ordered regions" already,
to prevent worst case behavior. I actually had to take those out in
order to get the improvement with Insertion Sort. If I put them back in,
then the whole advantage is gone, and the "optimized" version runs
slightly slower than the old version.

Applying just one final Insertion Sort instead of sorting each sub
problem below the cut off with it also turned out to be slightly
slower unlike in my test extension.

I'm confused to see only minimal improvements, when implementing a
"textbook" Quicksort with the same optimization, the benefits are much
more obvious.

I assume there's more optimization in Ruby's current algorithm
that would no longer be needed with additional Insertion Sort,
maybe you have some ideas? Or maybe there's also room to
improve my Insertion Sort?

This is as fast as I could get it right now, at least in theory it
should be possible to improve it.

#4 Updated by Martin Bosslet about 2 years ago

I could improve the Insertion Sort a bit by using memcpy/memmove.
It runs a bit faster even with the "worst case detection" of
Quicksort turned on, things look a little brighter now:

$ ruby driver.rb -p bm_sort -r 10


benchmark results:
name ruby 2.0.0dev (2012-04-05 trunk 35241) [x8664-linux]
sort
int 0.343
sortintcustom 0.296
sortintrev 0.101
sortintsorted 0.090
sortstrrnd 0.099

I also checked whether Binary Insertion Sort would bring
additional improvement, it did not.

This version performs slightly better now on all fronts
and the patch needs a lot less additional code.

#5 Updated by Martin Bosslet about 2 years ago

Updated patch - I forgot that chklim may be set to 0,
so I separated chklim and cutoff.

#6 Updated by Koichi Sasada over 1 year ago

ping. status?

#7 Updated by Martin Bosslet over 1 year ago

  • Target version changed from 2.0.0 to next minor

It's a tiny bit faster. But it should be noticeably faster. I need to get into the details of the current qsort implementation and do some tests/calculations in order to really know where to optimize appropriately. OK if I bump to "next minor"? I'd prefer to do some more analysis.

Also available in: Atom PDF