Implementing Sorting Refinements

From: Manolo _ <mac_man2005(at)hotmail(dot)it>
To: <pgsql-hackers(at)postgresql(dot)org>
Subject: Implementing Sorting Refinements
Date: 2008-01-01 20:09:49
Message-ID: BAY112-W7305A293CC276B888367AE6510@phx.gbl
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers


Hi to all.

This mail is aimed at asking some suggestion to face PostgreSQL (PG) development to implement a refinement to PG source code. I'll briefly summarize the idea of the "2-Way Replacement Selection" (2WRS) refinement, just in case. If you already remember what 2WRS is, you can please jump directly to the IMPLEMENTATION ISSUES part of this mail.

2WRS.
Refinement of the Replacement Selection (RS) technique currently used by PG in src/backend/utils/sort/tuplesort.c .
The 2WRS uses two heaps instead of just one in order to create the current (logical) run. Here there are some fundamental points of the 2WRS technique:
- 'qsort' the initial unsorted 'memtuples' array
- divide the 'memtuples' array into two parts and each of those will be managed as a heap
- one of the heaps will arrange its elements in ascending order, while the other heap in descending order
- each heap will spill its current root in its corresponding run (i.e.: we have a run per each of those two heaps), so we are actually creating 2 physical current runs
- those 2 current physical runs could "theoretically" be merged into the same logical run, actually we can make 'mergesort' think they do belong to the same physical run. That reduces the number of comparisons 'mergesort' has to do at each merge step (that means less seek delay time on mass storage). We can also think the average length of logical runs produced by 2WRS will probably be greater than the average length produced by RS (again less seek delay time: the longer each run the less number of final runs produced, that means the less number of comparisons at each 'mergesort' step).

IMPLEMENTATION ISSUES.
Where to place those heaps?
1) I think that both heaps could be arranged on the same 'memtuples' array. That allows easily subsequent resizing those heaps according to their effective use or according to some heuristic, without reallocating memory.

How to arrange those heaps?
2a) It's convenient to arrange those heaps "root to root". That is arranging those heaps with their roots toward the center of 'memtuples' (in a way we can say they lay "face to face", or "root to root" as said before) while their leaves lay towards the extreme indexes of the 'memtuples' array (that is the last leaf of one heap will lay at index 0, the last leaf of the other heap laying at index memtupsize-1. This arrangement prevents overlapping elements between those physical runs associated to the same current logical run.
PRO: once we qsort memtuples and divide it into 2 parts we already get those two heaps, no need to build them.
CONTRA: ???

2b) As in 2a) but arranging heaps "leaf to leaf", that is their roots will lay at the extreme indexes of 'memtuples' while their leaves towards the center of the 'memtuples' array.
Or even start building heaps as soon as we get initial elements, instead of qsort the whole 'memtuples' array.
Any PRO/CONTRA compared to 2a)???

Current run numbers
I think I should duplicate the 'int currentRun' variable in the Tuplesortstate struct. I'll replace it with a 'int currentRunUP' and 'int currentRunDOWN' variables in order to distinguish those two physical runs associated to those 2 heaps. In this case I will give a run number (max{currentRunUP,currentRunDOWN} + 1) to elements not belonging to the current logical run. I suppose no need to touch 'long availMem' nor 'long allowedMem' variables nor any others.

Heap functions
I will duplicate all the heap management functions in order to adapt them to the kind of heap they should be applied to (for example, the tuplesort_heap_siftup function should be replaced with tuplesort_heap_siftupUP and tuplesort_heap_siftupDOWN functions).

Merge Plan
This technique would use a sort of "merge plan" to instruct mergesort on how to use those physical runs. Actually mergesort should consider at first "odd runs" before "pair runs". That is, for example, mergesort should start merging runs with run number 1,3,5,7,... and when run number X terminates start considering run number X+1. Obviously that doesn't need any merge plan, but I remember someone else (as Simon Riggs) was interested in sorting improvements so it could be a good thing to know if I should consider any conventions or paramethers in order to possibly create that merge plan.

DEVELOPMENT CONTEXT
I preferred to use the "last stable release" at the moment, that is 8.2.

Any comment/suggestion/advice ?

Thanks for your attention and for your time.
Regards, Manolo.
_________________________________________________________________
Express yourself instantly with MSN Messenger! Download today it's FREE!
http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Simon Riggs 2008-01-01 20:55:58 Re: Index Page Split logging
Previous Message Tom Lane 2008-01-01 19:02:01 Re: Index Page Split logging