BlockQuicksort

Code of BlockQuicksort and other algorithms tested in the paper "BlockQuicksort: How Branch Mispredictions don't affect Quicksort" by Stefan Edelkamp and Armin Weiß (available at http://arxiv.org/abs/1604.06697).

This package consists of the following files:

• makefile

• driver.cpp : generates test cases and runs the sorting algorithms

• quicksort.h : contains different Quicksort main loops

• partition.h : contains different versions of the block partitioner and other partitioners

• median.h : functions for choosing pivot elements

• insertionsort.h : copy of insertion_sort from the GCC 4.7.2 implementation of std::sort

• interfaces for BlockQuicksort with different pivot selection methods / partitioners: blocked.h++, blocked_double_pivot_check.h++, blocked_double_pivot_check_mosqrt.h++, blocked_hoare_finish.h++, blocked_mo3_mo3.h++, blocked_mo3_mo5.h++, blocked_mo5_mo5.h++, blocked_Mo5.h++, blocked_mo23.h++, blocked_mosqrt.h++, blocked_simple.h++, hoare.h++

The method blocked_double_pivot_check_mosqrt.h++ performed in all benchmarks close to the optimum and therefore is shown in most plots in the paper. For comparison also the method blocked_simple.h++ is shown in most plots in the paper.

• lomuto_katajainen.h++ : interface for Tuned Quicksort by Elmasry, Katajainen and Stenmark http://www.diku.dk/~jyrki/Myris/Kat2014S.html

• ssssort.h++ : Super Scalar Sample Sort implemented by Timo Bingmann and Lorenz Hübschle-Schneider https://github.com/lorenzhs/ssssort/blob/b931c024cef3e6d7b7e7fd3ee3e67491d875e021/ssssort.h

• stl_gcc : copy of std::sort from the GCC 4.7.2 implementation

• qsort3_aumueller.h++, rotations.h, inssort.h : three pivot Quicksort implemented by Timo Bingmann and Martin Aumüller http://eiche.theoinf.tu-ilmenau.de/quicksort-experiments/

• Yaroslavskiy.h++ : dual pivot Quicksort by Vladimir Yaroslavkiy http://codeblab.com/wp-content/uploads/2009/09/DualPivotQuicksort.pdf

• For running time experiments with one single algorithm: make .time

• For running time tests with different sets of algorithms and test cases (output written to .csv file): make timetest make allAlgtimetest make pivotmethodtest make time-tests-data-all make blocksizetest-data

• For comparison and move experiments with one single algorithm: make .comp make .move These tests are not implemented for all algorithms.

How to build single programs

Build a single algorithm can be done through:

	make build ALGNAME="algorithm_name"

Build a single algorithm with a custom type:

	make build ALGNAME="algorithm_name" TYPE="custom_type"	

#How to run

Run single programs

	./a.out input_size distribution_type seed

distribution_type is given by 1 letter, see driver.cpp for further details, it has quite a few!.

Run tests

There are quite a few testing methods, I encourage you to go exploring yourself, however the most notable are listed below.

newtimetest

Runs all algorithms with the permutations defined in the variable: newsmalldata. We tested with the variable existingsmalldata

make newtimetest

timetest

Existing method we do not implement. Runs a lot of tests with different types and permutations, takes a while!

make timetest

perftimetest

Runs the perf state command on all algorithms, on all input sizes with random permutation. Used to see instructions count, branches and branch misses.

make perftimetest

newblocksizetest

Runs blocksize tests on both blocked and multi-pivot blocked algorithms.

make newblocksizetest