Thread: [GSoC 2018] Proposal Draft

On Sat, Mar 17, 2018 at 5:34 PM, Kefan Yang <starordust@gmail.com> wrote:
> I am Kefan Yang, a third-year Computing Science student from Simon Fraser
> University, Canada. I am very interested in the sorting algorithm
> benchmarking and implementation issue you mentioned on the idealist of
> Google Summer of Code 2018.

> I've attached a proposal draft to this email. It's a brief one but I guess
> it's enough to show my current ideas:)

The proposal reads:

"""
Industrial implementation of selected sorting algorithm:
The industrial version is basically an optimization based on the benchmark
implementation. I plan to use optimizations like checking if input
array is already sorted
or applying insertion sort directly for short arrays to see if the
performance can be
improved. I am still looking for other common optimization methods.

"""

But our qsort implementation, which is the Bentley & McIlroy
implementation with a couple of customizations, already does
everything you mention (I refer to the precheck algorithmic
customization, and the way that we check to see which side of a pivot
to use real recursion with to avoid stack overflows). I suggest that
you read the paper [1] -- the code that we use is almost directly
lifted from the paper. The opaque sounding variable names are the
same, and the C code from the paper is structured in exactly the same
way.

I think that this won't be a particularly easy project to get
committed. I suggest that if you go forward with it that you
specifically look into integrating "Dual-Pivot Quicksort" [2] as a
whole cloth replacement for the B&M implementation. It seems like this
has some chance of working out, because it's more or less acknowledged
by Bentley himself to be a kind of successor to his own industrial
quicksort implementation [3] -- it's derived from it. Note that Java 7
uses dual pivot quicksort when sorting integers.

In general, we've had a lot of success with optimizing sorting in the
past few years by focusing on things like avoiding cache misses in
comparators. There has been much less success with algorithmic
improvements, and no success at all with novel algorithmic
enhancements. PostgreSQL development just isn't a great way to do that
sort of thing.

BTW, I noticed that you go on to say this:

"""
However,
since disk operation is much expensive than in-memory sorting, I am
not sure if we can
observe a significant difference in this way.

"""

I think that you'd be surprised at how often this isn't true these
days, at least when sorting enough data for spilling to disk to be
relevant. The main reason for that is that the cost of writing out
runs increases linearly, and therefore eventually becomes very small
compared to the costs of sorting itself, which increases at a
linearithmic rate.

[1] http://cs.fit.edu/~pkc/classes/writing/samples/bentley93engineering.pdf
[2] http://codeblab.com/wp-content/uploads/2009/09/DualPivotQuicksort.pdf
[3] http://mail.openjdk.java.net/pipermail/core-libs-dev/2009-September/002630.html
-- 
Peter Geoghegan

On Sat, Mar 17, 2018 at 7:00 PM, Kefan Yang <starordust@gmail.com> wrote:
> What I am trying to say here is that similar optimizations can be applied to
> novel algorithms or other implementations of quicksort.

A novel algorithm is something to avoid here, because novel techniques
tend to only work out for specific datasets and data distributions. In
general, we care about a wide variety of cases, and are very keen to
avoid regressions.

> The paper about "Dual-Pivot Quicksort" is really helpful and it has a Java
> implementation already. I can definitely make use of that.

Cool. Please be careful to pick source material that has a license
that is compatible with the PostgreSQL license.

> Also, I am wondering what's the normal approach to testing if a certain
> sorting implementation brings performance gain in this project?

It varies quite a bit. You can search the lists archives for some of
this. For example, Tomas Vondra often tests my sorting patches by
subjecting them to a variety of tests with varied distributions and
data types. His results are then collated in a spreadsheet for easy
analysis. Maybe you can replicate that.

The only obvious trend is that everything is tested using real SQL
statements (including CREATE INDEX statements). In the past,
enhancements that were successfully incorporated tended to either
obviously have little or no downside, or have a very significant
upside that made it worth talking about accepting a small regression
in a minority of less representative cases.

> More
> specifically,  You mentioned that little progress has been made with novel
> algorithmic enhancement. How can we get that conclusion?

That's simply been the trend. It isn't particularly likely that
Postgres will be a project that pioneers some kind of algorithmic
enhancement that has broad applicability, that could just as easily
benefit a general purpose library sort routine. If you're interested
in algorithmic enhancements in the sense that I understand the term,
then that's the high bar that would have to be met -- that would
amount to a new insight in an area that has already been researched in
enormous detail by many people, most of whom don't know much about
database systems in particular.

On the other hand, techniques like abbreviated keys work well because
they effectively exploit things like the first normal form, and the
fact that a high start up cost for the sort is already very likely. It
is a technique specifically tailored for a database system, and modern
hardware characteristics.

-- 
Peter Geoghegan