> On 30/01/23 11:01, John Naylor wrote:
> Since David referred to L3 size as the starting point of a possible
configuration parameter, that's actually cache-conscious.
Okay, makes sense. I am correcting error on my part.
> I'm not close enough to this thread to guess at the right direction
(although I hope related work will help), but I have a couple general
remarks:
> 1. In my experience, academic papers like to test sorting with
register-sized numbers or strings. Our sorttuples are bigger, have
complex comparators, and can fall back to fields in the full tuple.
> 2. That paper is over 20 years old. If it demonstrated something
genuinely useful, some of those concepts would likely be implemented in
the real-world somewhere. Looking for evidence of that might be a good
exercise.
> 3. 20 year-old results may not carry over to modern hardware.
> 4. Open source software is more widespread in the academic world now
than 20 years ago, so papers with code (maybe even the author's github)
are much more useful to us in my view.
> 5. It's actually not terribly hard to make sorting faster for some
specific cases -- the hard part is keeping other inputs of interest from
regressing.
> 6. The bigger the change, the bigger the risk of regressing somewhere.
Thanks John, these inputs are actually what I was looking for. I will do
more research based on these inputs and build up my understanding.
Regards,
Ankit