Parallel Sorting Algorithms ITCS 41455145 Parallel Programming B

Parallel Sorting Algorithms ITCS 4145/5145, Parallel Programming B. Wilkinson March 11, 2016 1

Sorting Algorithms - Rearranging a list of numbers into increasing (strictly non-decreasing) order. Sorting numbers is important in applications as it can make subsequent operations more efficient. Simple sorting algorithms (Bubble Sort, Insertion Sort, Selection Sort, …) are O(n 2) (n numbers) Lower bound for comparison based algorithms (Merge Sort, Quicksort, Heap Sort, …) is O (n log n) 2

Potential time complexity using parallel programming O(nlogn) optimal for any sequential sorting algorithm (without using special properties of the numbers, see later). Best parallel time complexity we can expect based upon such a sequential sorting algorithm using n processors is: Why not better possible? Has been obtained but constant hidden in order notation extremely large. 3

General notes In-place sorting – sorting the values within the same array Not in-place (or out-of-place) sorting – sorting the values to a new array Stable sorting algorithm – Identical values remain in the same order as the original after sorting 4

General notes-1 Sorting requires moving numbers from one place to another in a list. The following algorithms concentrate upon a message –passing model for moving the numbers. Also parallel time complexities we give assume all processes operate in synchronism. Full treatment of time complexity given in textbook. 5

Compare-and-Exchange Sorting Algorithms “Compare and exchange” -- the basis of several, if not most, classical sequential sorting algorithms. Two numbers, say A and B, are compared. If A > B, A and B are exchanged, i. e. : if (A > B) { temp = A; A = B; B = temp; } 6

Message-Passing Compare and Exchange Version 1 P 1 sends A to P 2, which compares A and B and sends back B to P 1 if A is larger than B (otherwise it sends back A to P 1): 7

Alternative Message Passing Method Version 2 P 1 sends A to P 2 and P 2 sends B to P 1. Then both processes perform compare operations. P 1 keeps the larger of A and B and P 2 keeps the smaller of A and B: 8

Question Version 1 Do both versions give the same answers if A and B are on different computers? Answer Version 2 Usually but not necessarily. Depends upon how A and B are represented on each computer! 9

Note on Precision of Duplicated Computations Previous code assumes if condition, A > B, will return the same Boolean answer in both processors. Different processors operating at different precision could conceivably produce different answers if real numbers are being compared. Situation applies to anywhere computations are duplicated in different processors, which is sometimes done to improve performance (can reduce data movement). 10

Data Partitioning Version 1 p processors and n numbers. n/p numbers assigned to each processor: 11

Version 2 12

Partitioning numbers into groups: p processors and n numbers. n/p numbers assigned to each processor applies to all parallel sorting algorithms to be given as number of processors usually much less than the number of numbers. 13

Parallelizing common sequential sorting algorithms 14

Bubble Sort First, largest number moved to the end of list by a series of compares and exchanges, starting at the opposite end. Actions repeated with subsequent numbers, stopping just before the previously positioned number. In this way, larger numbers move (“bubble”) toward one end. After pass m, the m largest values are in the m last locations in the array. 15

After pass 1, the largest values is in it’s location 16

Time Complexity Number of compare and exchange operations Indicates time complexity of O(n 2) if a single compareand- exchange operation has a constant complexity, O(1). Not good but can be parallelized. 17

Parallel Bubble Sort Iteration could start before previous iteration finished if does not overtake previous bubbling action: 18

Odd-Even (Transposition) Sort Variation of bubble sort. Operates in two alternating phases, even phase and odd phase. Even phase: Even-numbered processes exchange numbers with their right neighbor. Odd phase: Odd-numbered processes exchange numbers with their right neighbor. 19

Odd-Even Transposition Sorting eight numbers compare and exchange 20

Question What is the parallel time complexity? Answer O(n) with n processors and n numbers. 21

Mergesort A classical sequential sorting algorithm using divide-and-conquer approach Unsorted list first divided into half. Each half is again divided into two. Continued until individual numbers obtained. Then pairs of numbers combined (merged) into sorted list of two numbers. Pairs of these lists of four numbers are merged into sorted lists of eight numbers. This is continued until the one fully sorted list is obtained. Takes advantage of merging two sorted lists can be be done efficiently in O(n) time. 22

Merge Sort P 0 P 1 P 2 P 3 P 4 P 5 P 6 P 7 4 2 7 8 5 1 3 6 Tim e 4 2 8 7 8 2 4 7 8 5 1 5 3 6 1 1 5 1 3 5 6 1 2 3 4 5 6 7 8 3 3 6 6 Merge 2 4 5 1 3 6 7 8 7 Divide 4 2 7 8

Parallelizing Mergesort Using tree allocation of processes 24

Analysis Sequential time complexity is O(nlogn). Parallel 2 logn steps but each merge step takes longer and longer. Turns out to be O(n) with n processor and n numbers, see text. 25

Quicksort Very popular sequential sorting algorithm that performs well with average sequential time complexity of O(nlogn). First list divided into two sublists. All numbers in one sublist arranged to be smaller than all numbers in other sublist. Achieved by first selecting one number, called a pivot, against which every other number is compared. If number less than pivot, it is placed in one sublist, otherwise, placed in other sublist. Pivot could be any number in list, but often first number chosen. Pivot itself placed in one sublist, or separated and placed in its final position. 26

Quicksort Sequential time complexity • Works well when the pivot is approximately in the middle • If the pivot is chosen poorly, quicksort resorts to

Parallelizing Quicksort Using tree allocation of processes 28

With the pivot being withheld in processes: 29

Analysis Fundamental problem with all tree constructions – initial division done by a single processor, which will seriously limit speed. Tree in quicksort will not, in general, be perfectly balanced. Pivot selection very important for reasonably balanced tree and make quicksort operate fast. 30

Batcher’s Parallel Sorting Algorithms • Odd-even Mergesort • Bitonic Mergesort Originally derived in terms of switching networks. Both well balanced and have parallel time complexity of O(log 2 n) with n processors. 31

Odd-Even Mergesort Uses odd-even merge algorithm that merges two sorted lists into one sorted list. 32

Odd-Even Merging of Two Sorted Lists 33

Odd-Even Merge Sort 47 98 14 23 37 84 15 30 Phase 1 47 98 14 23 37 84 15 30

Odd-Even Merge Sort Phase 2 47 98 14 23 37 84 15 30 14 47 23 98 15 37 30 84 14 23 47 98 15 30 37 84

Odd-Even Merge Sort Phase 3 14 23 47 98 15 30 37 84 14 15 37 47 23 30 84 98 14 15 23 30 37 47 84 98

Odd-Even Mergesort Apply odd-even megersort(n/2) recursively to the two halves a 0 … an/2 -1 and subsequence an/2 …an-1. Leads to a time complexity of O(n log n 2) with one processor and O(log n 2) with n processors More information see: http: //www. iti. fh-flensburg. de/lang/algorithmen/sortieren/networks/oemen. htm 37

Bitonic Mergesort Bitonic Sequence A monotonic increasing sequence is a sequence of increasing numbers. A bitonic sequence has two sequences, one increasing and one decreasing. e. g. a 0 < a 1 < a 2, a 3, …, ai-1 < ai > ai+1, …, an-2 > an-1 for some value of i (0 <= i < n). A sequence is also bitonic if the preceding can be achieved by shifting the numbers cyclically (left or right). 38

Bitonic Sequences 39

“Special” Characteristic of Bitonic Sequences If we perform a compare-and-exchange operation on ai with ai+n/2 for all i, where there are n numbers, get TWO bitonic sequences, where the numbers in one sequence are all less than the numbers in the other sequence. 40

Creating two bitonic sequences from one bitonic sequence Example: Starting with bitonic sequence 3, 5, 8, 9, 7, 4, 2, 1, get: All numbers less than other bitonic sequence 41

Sorting a bitonic sequence Given a bitonic sequence, recursively performing operations will sort the list. 42

Sorting To sort an unordered sequence, sequences merged into larger bitonic sequences, starting with pairs of adjacent numbers. By compare-and-exchange operations, pairs of adjacent numbers formed into increasing sequences and decreasing sequences. Two adjacent pairs form a bitonic sequence. Bitonic sequences sorted using previous bitonic sorting algorithm. By repeating process, bitonic sequences of larger and larger lengths obtained. Finally, a single bitonic sequence sorted into a single increasing sequence. 43

Bitonic Mergesort 44

Bitonic mergesort with 8 numbers 45

Phases The six steps (for eight numbers) divided into three phases: Phase 1 (Step 1) compare/exchange pairs of numbers into increasing/decreasing sequences and merge into 4 -bit bitonic sequences. Phase 2 (Steps 2/3) Sort each 4 -bit bitonic sequence (alternate directions) and merge into 8 -bit bitonic sequence. Phase 3 (Steps 4/5/6) Sort 8 -bit bitonic sequence. 46

Number of Steps In general, with n = 2 k, there are k phases, each of 1, 2, 3, …, k steps. Hence total number of steps given by: 47

Sorting Conclusions so far Computational time complexity using n processors • Odd-even transposition sort - O(n) • Parallel mergesort - O(n) but unbalanced processor load and communication • Parallel quicksort - O(n) but unbalanced processor load, and communication, can degenerate to O(n 2) • Odd-even Mergesort and Bitonic Mergesort O(log 2 n) Bitonic mergesort has been a popular choice for parallel sorting. 48

Sorting on Specific Networks Algorithms can take advantage of underlying interconnection network of the parallel computer. Two network structures have received specific attention: • Mesh • Hypercube because parallel computers have been built with these networks. Of less interest nowadays because underlying architecture often hidden from user. There are MPI features for mapping algorithms onto meshes. Can always use a mesh or hypercube algorithm even if the underlying architecture is not the same. 49

Mesh - Two-Dimensional Sorting The layout of a sorted sequence on a mesh could be row by row or snakelike. Snakelike: 50

Shearsort Alternate row and column sorted until list fully sorted. Row sorting in alternative directions to get snake-like sorting: 51

Shearsort 52

Using Transposition Causes the elements in each column to be in positions in a row. Can be placed between the row operations and column operations 53

Rank Sort A simple sorting algorithm. Does not achieve a sequential time of O(nlogn), but can be parallelized easily. Leads us onto algorithms which can be parallelized to achieve O(logn) parallel time. 54

Rank Sort Number of numbers that are smaller than each selected number is counted. This count provides the position of selected number in sorted list; that is, its “rank. ” • First number a[0] is read and compared with each of the other numbers, a[1] … a[n-1], recording the number of numbers less than a[0]. • Suppose this number is x. This is the index of the location in the final sorted list. The number a[0] is copied into the final sorted list b[0] … b[n-1], at location b[x]. • Actions repeated with the other numbers. Overall sequential sorting time complexity of O(n 2) (not exactly a good sequential sorting algorithm!). 55

Sequential Code for (i = 0; i < n; i++) { /* for each number */ x = 0; for (j = 0; j < n; j++) /* count number less than it */ if (a[i] > a[j]) x++; b[x] = a[i]; /* copy number into correct place */ } This code will fail if duplicates exist in the sequence of numbers. Easy to fix. (How? ) 56

Sequential Code handling duplicates for (i = 0; i < n; i++) { /* for each number */ x = 0; for (j = 0; j < n; j++) /* count number less than it */ if (a[i] > a[j] || (a[i] == a[j] && j < i) ) x++; b[x] = a[i]; /* copy number into correct place */ } 57

Parallel Code Using n Processors One processor allocated to each number. Finds final index in O(n) steps. With all processors operating in parallel, parallel time complexity O(n) with n processors. In forall notation, code would look like forall (i = 0; i < n; i++) { /* for each no in parallel*/ x = 0; for (j = 0; j < n; j++) /* count number less than it */ if (a[i] > a[j]) x++; b[x] = a[i]; /* copy no into correct place */ } Easy to write in Open. MP. Parallel time complexity, O(n), as good as some sorting algorithms so far. Can do even better if we have more processors. 58

Using n 2 Processors Comparing one number with other numbers in list using multiple processors: n - 1 processors used to find rank of one number. With n numbers, (n - 1)n processors or (almost) n 2 processors needed. Incrementing counter done sequentially and requires maximum of n steps. Total number of steps = 1 + n, still O(n). 59

Reducing steps in final count Tree to reduce number of steps involved in incrementing counter: O(logn) algorithm with n 2 processors. Processor efficiency relatively low. 60

Rank Sort using n 2 processors #pragma omp parallel private(i, j, x) { #pragma omp for (i = 0; i < n; i++) { x = 0; #pragma omp for reduction(+: x) for (j = 0; j < n; j++) if (a[i] > a[j]) x++; #pragma omp master { b[x] = a[i]; } } } n threads for each a[i] Original Open. MP didn’t support nested foralls Each of n thread counts a partial sum of array elements less than a[i] Reduction on x gives us the total sum of array elements less than a[i]

Parallel Rank Sort Conclusions Easy to do as each number can be considered in isolation. Rank sort can sort in: O(n) with n processors or O(logn) using n 2 processors. In practical applications, using n 2 processors prohibitive. Theoretically possible to reduce time complexity to O(1) by considering all increment operations as happening in parallel since they are independent of each other. (Look up about PRAMs) 62

Summary Sequential Complexity Parallel Complexity with n processors Bubble Sort O(n 2) O(n) Even-Odd Sort O(n 2) O(n) Merge Sort O(nlogn) O(n) Quicksort O(nlogn) O(n) Odd-Even Merge Sort O(log 2 n) Sorting Algorithm Shearsort Rank Sort O(√n(logn+1)) O(n 2) O(n)

Other Sorting Algorithms We began by giving lower bound for the time complexity of a sequential sorting algorithm based upon comparisons as O(nlogn). Consequently, time complexity of best parallel sorting algorithm based upon comparisons is O(logn) with n processors (or O(nlogn/p) with p processors). Sequential sorting algorithms can achieve better than O(nlogn) if they assume certain properties of the numbers being sorted (e. g. they are integers in a given range). These algorithms very attractive candidates for parallelization. 64

Counting Sort If numbers to be sorted are integers in a given range, can encode rank sort algorithm to reduce sequential time complexity from O(n 2) to O(n). Method called as Counting Sort. Suppose unsorted numbers stored in array a[ ] and final sorted sequence is stored in array b[ ] Algorithm uses an additional array, say c[ ], having one element for each possible value of the numbers. Suppose range of integers is from 1 to m. Array c has elements c[1] through c[m] inclusive. 65

Counting Sort • Counting Sort is a stable sort (identical values remain in same relative position as in original array) • However, it is not an “in-place” sort, we need extra space

Step 1 First, c[ ] will be used to hold the histogram of the sequence, that is, the number of each number. This can be computed in O(m) time with code such as: for (i = 1; i <= m; i++) c[i] = 0; for (i = 1; i <= m; i++) c[ a[i] ]++; 67

Step 2 The number of numbers less than each number found by performing a prefix sum operation on array c[ ]. In the prefix sum calculation, given a list of numbers, x 0, …, xn-1, all the partial summations, i. e. , x 0 + x 1, x 0 + x 1 + x 2 + x 3, … computed. Prefix sum computed using histogram originally held in c[ ] in O(m) time as described below: for (i = 2; i <= m; i++) c[i] = c[i] + c[i-1]; 68

Step 3 (Final step) The numbers are placed in sorted order as described below: for (i = n; i >= 1; i--) { b[c[a[i]]] = a[i] c[ a[i] ]--; /* ensures stable sorting */ } Complete code has O(n + m) sequential time complexity. If m is linearly related to n as it is in some applications, code has O(n) sequential time complexity. 69

Counting Sort 1 2 3 4 5 6 7 8 Original sequence a[] 5 2 3 7 5 6 4 1 Step 3 moves backwards thro a[] Step 1. Histogram c[] 1 1 2 1 1 Step 2. Prefix sum c[] 1 2 3 4 6 7 8 Step 3. Sort b[] 1 2 3 4 5 5 6 7 Final sorted sequence Move 5 to position 6. Then decrement c[5]

Parallel Counting Sort (using n processors) If take into account duplicate numbers • Step 1 – Creating the Histogram: O(logn). Use a tree construction. Updating c[a[i]]++ is a critical section when there are duplicate values in a[ ]. • Step 2 – Parallel version of prefix sum using n processors is O(logn) • Step 3 – Placing number in place: O(logn) again because of possible duplicate values (c[a[i]]– is a critical section) • All 3 steps: O(logn)

Radix Sort Assumes numbers to sort represented in a positional digit representation such as binary and decimal numbers. The digits represent values and position of each digit indicates their relative weighting. Radix sort starts at least significant digit and sorts numbers according to their least significant digits. Sequence then sorted according to next least significant digit and so on until the most significant digit, after which sequence is sorted. For this to work, necessary that order of numbers with the same digit is maintained, that is, one must use a stable sorting algorithm. 72

Radix sort using decimal digits 73

Radix sort using binary digits 74

Radix sort can be parallelized by using a parallel sorting algorithm in each phase of sorting on bits or groups of bits. Neat way of doing is using prefix sum algorithm (for binary digits). 75

Using prefix sum to sort binary digits When prefix sum calculation applied to a column of bits, it gives number of 1’s up to each digit position because all digits can only be 0 or 1 and prefix calculation will simply add number of 1’s. Prefix calculation on the digits inverted (diminished prefix sum) give the number of 0’s up to each digit. When digit considered being a 0, diminished prefix sum calculation provides new position for number. When digit considered being a 1, result of prefix sum calculation plus largest diminished prefix calculation gives final position for number. Prefix sum calculation leads to O(logn) time with n - 1 processors and constant b and r. 76

Summary Sequential Complexity Parallel Complexity with n processors Bubble Sort O(n 2) O(n) Even-Odd Sort O(n 2) O(n) Merge Sort O(nlogn) O(n) Quicksort O(nlogn) O(n) Odd-Even Merge Sort O(log 2 n) Sorting Algorithm Bitonic Merge Sort O(log 2 n) Shearsort O(√n(logn+1)) Rank Sort O(n 2) O(n) Counting Sort O(n) O(logn) O(nlog. N) O(logn*log. N) Radix Sort

Questions