问题
I try to use OpenMP to parallel quicksort in partition part and quicksort part. My C code is as follows:
#include "stdlib.h"
#include "stdio.h"
#include "omp.h"
// parallel partition
int ParPartition(int *a, int p, int r) {
int b[r-p];
int key = *(a+r); // use the last element in the array as the pivot
int lt[r-p]; // mark 1 at the position where its element is smaller than the key, else 0
int gt[r-p]; // mark 1 at the position where its element is bigger than the key, else 0
int cnt_lt = 0; // count 1 in the lt array
int cnt_gt = 0; // count 1 in the gt array
int j=p;
int k = 0; // the position of the pivot
// deal with gt and lt array
#pragma omp parallel for
for ( j=p; j<r; ++j) {
b[j-p] = *(a+j);
if (*(a+j) < key) {
lt[j-p] = 1;
gt[j-p] = 0;
} else {
lt[j-p] = 0;
gt[j-p] = 1;
}
}
// calculate the new position of the elements
for ( j=0; j<(r-p); ++j) {
if (lt[j]) {
++cnt_lt;
lt[j] = cnt_lt;
} else
lt[j] = cnt_lt;
if (gt[j]) {
++cnt_gt;
gt[j] = cnt_gt;
} else
gt[j] = cnt_gt;
}
// move the pivot
k = lt[r-p-1];
*(a+p+k) = key;
// move elements to their new positon
#pragma omp parallel for
for ( j=p; j<r; ++j) {
if (b[j-p] < key)
*(a+p+lt[j-p]-1) = b[j-p];
else if (b[j-p] > key)
*(a+k+gt[j-p]) = b[j-p];
}
return (k+p);
}
void ParQuickSort(int *a, int p, int r) {
int q;
if (p<r) {
q = ParPartition(a, p, r);
#pragma omp parallel sections
{
#pragma omp section
ParQuickSort(a, p, q-1);
#pragma omp section
ParQuickSort(a, q+1, r);
}
}
}
int main() {
int a[10] = {5, 3, 8, 4, 0, 9, 2, 1, 7, 6};
ParQuickSort(a, 0, 9);
int i=0;
for (; i!=10; ++i)
printf("%d\t", a[i]);
printf("\n");
return 0;
}
I'm new to OpenMG and parallel programming. For the example in the main function, the sorting result is:
0 9 9 2 2 2 6 7 7 7
I used gdb to debug. In the early recursion, all went well. But in some recursions, it suddenly messed up to begin duplicate elements. Then generate the above result.
Can someone help me figure out where the problem is? Thank you very much!
回答1:
I feel sorry for my first comment.It does not matter with your problem.I have not found the true problem of your question(Maybe your move element has the problem).According to your opinion, I wrote a similar program, it works fine.(I am also new on OpenMP).
#include <stdio.h>
#include <stdlib.h>
int partition(int * a, int p, int r)
{
int lt[r-p];
int gt[r-p];
int i;
int j;
int key = a[r];
int lt_n = 0;
int gt_n = 0;
#pragma omp parallel for
for(i = p; i < r; i++){
if(a[i] < a[r]){
lt[lt_n++] = a[i];
}else{
gt[gt_n++] = a[i];
}
}
for(i = 0; i < lt_n; i++){
a[p + i] = lt[i];
}
a[p + lt_n] = key;
for(j = 0; j < gt_n; j++){
a[p + lt_n + j + 1] = gt[j];
}
return p + lt_n;
}
void quicksort(int * a, int p, int r)
{
int div;
if(p < r){
div = partition(a, p, r);
#pragma omp parallel sections
{
#pragma omp section
quicksort(a, p, div - 1);
#pragma omp section
quicksort(a, div + 1, r);
}
}
}
int main(void)
{
int a[10] = {5, 3, 8, 4, 0, 9, 2, 1, 7, 6};
int i;
quicksort(a, 0, 9);
for(i = 0;i < 10; i++){
printf("%d\t", a[i]);
}
printf("\n");
return 0;
}
回答2:
I've implemented parallel quicksort in a production environment, although with concurrent processes (i.e. fork() and join()) and not OpenMP. I also found a pretty good pthread solution, but a concurrent process solution was the best in terms of worst-case runtime. Let me start by saying that it doesn't seem like you're making copies of your input array for each thread, so you'll definitely encounter race conditions which can corrupt your data.
Essentially, what is happening is you have created an array N in shared memory, and when you do a #pragma omp parallel sections, you're spawning as many worker threads as there are #pragma omp section's. Each time a worker thread tries to access and modify elements of a, it will execute a series of instructions: "read the n'th value of N from the given address", "modify the n'th value of N", "write the n'th value of N back to the given address". Since you have multiple threads with no locking or synchronization, the read, modify, and write instructions may be executed in any order by multiple processors, so the threads may overwrite each other's modifications or read a non-updated value.
The best solution that I found (after many weeks of testing and benchmarking many solutions that I came up with) is to subdivide the list log(n) times, where n is the number of processors. For example, if you have a quad core machine (n = 4), subdivide the list 2 times (log(4) = 2) choosing pivots that are the medians of the data set. It is important that the pivots are medians, because otherwise you can end up with a case where a poorly chosen pivot causes the lists to be distributed unevenly amongst processes. Then each process does quicksort on its local subarray, then merges its results with the results of other processes. This is called "hyperquicksort", and from an initial github search, I found this. I can't vouch for the code in there, and can't publish any of the code that I wrote since it is protected under an NDA.
By the way, one of the best parallel sorting algorithm is PSRS (Parallel Sorting by Regular Sampling), which keeps list sizes more balanced amongst processes, doesn't unnecessarily communicate keys between processes, and can work on an arbitrary number of concurrent processes (they don't necessarily have to be a power of 2).
回答3:
This seems wrong my op is looking like this
[student@localhost ~]$ g++ -openmp QuickSort1.cpp
[student@localhost ~]$ ./a.out
Enter the size of array:8
Enter element at 0:77
Enter element at 1:99
Enter element at 2:22
Enter element at 3:44
Enter element at 4:11
Enter element at 5:66
Enter element at 6:88
Enter element at 7:55
=====================After Sorting===================
11 22 44 99 55 66 77 88
来源：https://stackoverflow.com/questions/16007640/openmp-parallel-quicksort