I would like to parallel the following C++ loop using OpenMP to improve the efficiency. (The value of each element in array2d can be 0 or 1 or 2. The values of array2d are not important for efficiency so I just randomly set each value from 0, 1 and 2. The values in count is initialized to 0.)
int array2d[100][10000];
int count[3][3][3];
//omp_set_num_threads(2);
//#pragma omp parallel for
for (int i = 0; i < 10000; ++i) {
int x = array2d[10][i];
int y = array2d[40][i];
int z = array2d[78][I];
//#pragma omp atomic
count[z][x][y]++;
}
But I cannot get improvements if I use 2, 4, or 8 threads to parallel the loop by #pragma omp parallel for. The execution time of the parallel versions is greater than that of the sequential version. I am curious whether this loop can be improved by OpenMP parallelism? If yes, how can I get a shorter execution time?
If your concern is efficiency, there are other things to do before you try OMP.
Your code is not cache-friendly: the row of 100 ints is 400 bytes, while the cache line is only 64. Since the values are limited to 0..2, single byte (uint8_t) will work better. I would even pack four of those into each byte.
There are 3 effects which can cause that your code is slower in parallel (but I do not know which one is the most impartant in your case):
This code is memory bound, depending on your hardware using more threads may not improve the speed of memory access, therefore the overall speed will not increase.
As already told by #Daniel the workload is very small, so the parallel overhead is big compared to the workload, therefore the runtime will be increased.
As also emphasized by #Daniel count array is small, it has 27 elements only. Continous increment of its elements can cause false sharing, which may reduce the efficiency significantly. You can change it by using reduction (note that in this case you do not need atomic operation, so delete that line):
#pragma omp parallel for reduction(+:count[:3][:3][:3])
If the speed will not increase, unfortunately this code is not worth parallelizing on your hardware. Try to parallelize a bigger part of your program.
Related
I have a piece of code in my full code:
const unsigned int GL=8000000;
const int cuba=8;
const int cubn=cuba+cuba;
const int cub3=cubn*cubn*cubn;
int Length[cub3];
int Begin[cub3];
int Counter[cub3];
int MIndex[GL];
struct Particle{
int ix,jy,kz;
int ip;
};
Particle particles[GL];
int GetIndex(const Particle & p){return (p.ix+cuba+cubn*(p.jy+cuba+cubn*(p.kz+cuba)));}
...
#pragma omp parallel for
for(int i=0; i<cub3; ++i) Length[i]=Counter[i]=0;
#pragma omp parallel for
for(int i=0; i<N; ++i)
{
int ic=GetIndex(particles[i]);
#pragma omp atomic update
Length[ic]++;
}
Begin[0]=0;
#pragma omp single
for(int i=1; i<cub3; ++i) Begin[i]=Begin[i-1]+Length[i-1];
#pragma omp parallel for
for(int i=0; i<N; ++i)
{
if(particles[i].ip==3)
{
int ic=GetIndex(particles[i]);
if(ic>cub3 || ic<0) printf("ic=%d out of range!\n",ic);
int cnt=0;
#pragma omp atomic capture
cnt=Counter[ic]++;
MIndex[Begin[ic]+cnt]=i;
}
}
If to remove
#pragma omp parallel for
the code works properly and the output results are always the same.
But with this pragma there is some undefined behaviour/race condition in the code, because each time it gives different output results.
How to fix this issue?
Update: The task is the following. Have lots of particles with some random coordinates. Need to output to the array MIndex the indices in the array particles of the particles, which are in each cell (cartesian cube, for example, 1×1×1 cm) of the coordinate system. So, in the beginning of MIndex there should be the indices in the array particles of the particles in the 1st cell of the coordinate system, then - in the 2nd, then - in the 3rd and so on. The order of indices within given cell in the area MIndex is not important, may be arbitrary. If it is possible, need to make this in parallel, may be using atomic operations.
There is a straight way: to traverse across all the coordinate cells in parallel and in each cell check the coordinates of all the particles. But for large number of cells and particles this seems to be slow. Is there a faster approach? Is it possible to travel across the particles array only once in parallel and fill MIndex array using atomic operations, something like written in the code piece above?
You probably can't get a compiler to auto-parallelize scalar code for you if you want an algorithm that can work efficiently (without needing atomic RMWs on shared counters which would be a disaster, see below). But you might be able to use OpenMP as a way to start threads and get thread IDs.
Keep per-thread count arrays from the initial histogram, use in 2nd pass
(Update: this might not work: I didn't notice the if(particles[i].ip==3) in the source before. I was assuming that Count[ic] will go as high as Length[ic] in the serial version. If that's not the case, this strategy might leave gaps or something.
But as Laci points out, perhaps you want that check when calculating Length in the first place, then it would be fine.)
Manually multi-thread the first histogram (into Length[]), with each thread working on a known range of i values. Keep those per-thread lengths around, even as you sum across them and prefix-sum to build Begin[].
So Length[thread][ic] is the number of particles in that cube, out of the range of i values that this thread worked on. (And will loop over again in the 2nd loop: the key is that we divide the particles between threads the same way twice. Ideally with the same thread working on the same range, so things may still be hot in L1d cache.)
Pre-process that into a per-thread Begin[][] array, so each thread knows where in MIndex to put data from each bucket.
// pseudo-code, fairly close to actual C
for(ic < cub3) {
// perhaps do this "vertical" sum into a temporary array
// or prefix-sum within Length before combining across threads?
int pos = sum(Length[0..nthreads-1][ic-1]) + Begin[0][ic-1];
Begin[0][ic] = pos;
for (int t = 1 ; t<nthreads ; t++) {
pos += Length[t][ic]; // prefix-sum across threads for this cube bucket
Begin[t][ic] = pos;
}
}
This has a pretty terrible cache access pattern, especially with cuba=8 making Length[t][0] and Length[t+1][0] 4096 bytes apart from each other. (So 4k aliasing is a possible problem, as are cache conflict misses).
Perhaps each thread can prefix-sum its own slice of Length into that slice of Begin, 1. for cache access pattern (and locality since it just wrote those Lengths), and 2. to get some parallelism for that work.
Then in the final loop with MIndex, each thread can do int pos = --Length[t][ic] to derive a unique ID from the Length. (Like you were doing with Count[], but without introducing another per-thread array to zero.)
Each element of Length will return to zero, because the same thread is looking at the same points it just counted. With correctly-calculated Begin[t][ic] positions, MIndex[...] = i stores won't conflict. False sharing is still possible, but it's a large enough array that points will tend to be scattered around.
Don't overdo it with number of threads, especially if cuba is greater than 8. The amount of Length / Begin pre-processing work scales with number of threads, so it may be better to just leave some CPUs free for unrelated threads or tasks to get some throughput done. OTOH, with cuba=8 meaning each per-thread array is only 4096 bytes (too small to parallelize the zeroing of, BTW), it's really not that much.
(Previous answer before your edit made it clearer what was going on.)
Is this basically a histogram? If each thread has its own array of counts, you can sum them together at the end (you might need to do that manually, not have OpenMP do it for you). But it seems you also need this count to be unique within each voxel, to have MIndex updated properly? That might be a showstopper, like requiring adjusting every MIndex entry, if it's even possible.
After your update, you are doing a histogram into Length[], so that part can be sped up.
Atomic RMWs would be necessary for your code as-is, performance disaster
Atomic increments of shared counters would be slower, and on x86 might destroy the memory-level parallelism too badly. On x86, every atomic RMW is also a full memory barrier, draining the store buffer before it happens, and blocking later loads from starting until after it happens.
As opposed to a single thread which can have cache misses to multiple Counter, Begin and MIndex elements outstanding, using non-atomic accesses. (Thanks to out-of-order exec, the next iteration's load / inc / store for Counter[ic]++ can be doing the load while there are cache misses outstanding for Begin[ic] and/or for Mindex[] stores.)
ISAs that allow relaxed-atomic increment might be able to do this efficiently, like AArch64. (Again, OpenMP might not be able to do that for you.)
Even on x86, with enough (logical) cores, you might still get some speedup, especially if the Counter accesses are scattered enough they cores aren't constantly fighting over the same cache lines. You'd still get a lot of cache lines bouncing between cores, though, instead of staying hot in L1d or L2. (False sharing is a problem,
Perhaps software prefetch can help, like prefetchw (write-prefetching) the counter for 5 or 10 i iterations later.
It wouldn't be deterministic which point went in which order, even with memory_order_seq_cst increments, though. Whichever thread increments Counter[ic] first is the one that associates that cnt with that i.
Alternative access patterns
Perhaps have each thread scan all points, but only process a subset of them, with disjoint subsets. So the set of Counter[] elements that any given thread touches is only touched by that thread, so the increments can be non-atomic.
Filtering by p.kz ranges maybe makes the most sense since that's the largest multiplier in the indexing, so each thread "owns" a contiguous range of Counter[].
But if your points aren't uniformly distributed, you'd need to know how to break things up to approximately equally divide the work. And you can't just divide it more finely (like OMP schedule dynamic), since each thread is going to scan through all the points: that would multiply the amount of filtering work.
Maybe a couple fixed partitions would be a good tradeoff to gain some parallelism without introducing a lot of extra work.
Re: your edit
You already loop over the whole array of points doing Length[ic]++;? Seems redundant to do the same histogramming work again with Counter[ic]++;, but not obvious how to avoid it.
The count arrays are small, but if you don't need both when you're done, you could maybe just decrement Length to assign unique indices to each point in a voxel. At least the first histogram could benefit from parallelizing with different count arrays for each thread, and just vertically adding at the end. Should scale perfectly with threads since the count array is small enough for L1d cache.
BTW, for() Length[i]=Counter[i]=0; is too small to be worth parallelizing. For cuba=8, it's 8*8*16 * sizeof(int) = 4096 bytes, just one page, so it's just two small memsets.
(Of course if each thread has their own separate Length array, they each need to zero it). That's small enough to even consider unrolling with maybe 2 count arrays per thread to hide store/reload serial dependencies if a long sequence of points are all in the same bucket. Combining count arrays at the end is a job for #pragma omp simd or just normal auto-vectorization with gcc -O3 -march=native since it's integer work.
For the final loop, you could split your points array in half (assign half to each thread), and have one thread get unique IDs by counting down from --Length[i], and another counting up from 0 in Counter[i]++. With different threads looking at different points, this could give you a factor of 2 speedup. (Modulo contention for MIndex stores.)
To do more than just count up and down, you'd need info you don't have from just the overall Length array... but which you did have temporarily. See the section at the top
You are right to make the update Counter[ic]++ atomic, but there is an additional problem on the next line: MIndex[Begin[ic]+cnt]=i; Different iterations can write into the same location here, unless you have mathematical proof that this is never the case from the structure of MIndex. So you have to make that line atomic too. And then there is almost no parallel work left in your loop, so your speed up if probably going to be abysmal.
EDIT the second line however is not of the right form for an atomic operation, so you have to make it critical. Which is going to make performance even worse.
Also, #Laci is correct that since this is an overwrite statement, the order of parallel scheduling is going to influence the outcome. So either live with that fact, or accept that this can not be parallelized.
I have to sort a collection of vectors with OpenMP in C++. I really think this should be a common problem; however, I never found anything about it.
Of course I could just use
#pragma omp parallel for schedule(dynamic,1)
for(int i =0; i<narrays; ++i)
sortSerial(array[i]);
and leave the scheduling to the runtime.
However, one array could take longer than all the others combined. Actually, I might well only have one array.
Or I might, say, have twenty thousands with a hundred elements each, plus two with a million each. In that case, if I have 8 cores, I'd like to assign five thousands sections per thread for the first four threads, and then two threads per section with the last four. So that the total time is constant.
I might then use nested parallelism. But if I do this
#pragma omp parallel for schedule(dynamic,1)
for(int i =0; i<narrays; ++i)
sortParallel(array[i]);
where sortParallel spawns the correct number of threads to balance the workload, OMP will still start processing as many threads as possible; and then the nested parallelism will most probably oversubscribe. This seems bad, especially since the sortParallel routine has a lot of barriers.
The ideal solution would be to tell for how many threds every iteration will need, so that it might avoid starting new tasks if it would oversubscribe, but this doesn't seem possible.
I might have to notice that it should compile under VisualStudio 2013, so, it seems OpenMP 2.0 -and in particular, no tasks.
I am now exploring weird combinations of deep nestings of parallel region. But I'd really rather not do that. Is there some obvious solution I am missing?
My last design starts from the fact that, if the thread division is done correctly, then the big sections never take (too much) longer than total/threads. And the same will do the small sections, if we properly divide them among the remaining threads. So we don't really need to scramble to reuse the threads we used to sort the bigger vectors; since they will finish (alas, roughly) all at the same time anyway.
Therefore, I can just divide long intervals and small intervals in different vectors, put in threads[i] a partition of the number of threads which tries to create optimal balance, and do this
void sort(int i,int threads){
if(i<nbigintervals) //sort one big interval
sortParallel(bigInterval[i],threads); //spawns threads threads
else //sort all the others
#pragma omp parallel for schedule(dynamic,1) num_threads(threads) //nested parallel section, assigning the remaining vectors in dynamic fashion.
for(int j=0;j<nsmallintervals;++j)
sortSerial(smallInterval[j]);
}
int main() {
//....
//assign vectors of intervals
//balance threads
omp_set_nested(true);
#pragma omp parallel for schedule(static,1) num_threads(nbigintervals +1)
for(int i=0; i<=nbigintervals; ++i) //notice the less _or equal_, which adds one iteration (and makes their number match the number of threads requested)
sort(i,threads[i]);
//....
}
This avoids oversubscription, since the threads spawned by each iteration are designed to sum to the available total. Is there a simpler/cleaner/faster solution?
I'm using the following code, which contains an OpenMP parallel for loop nested in another for-loop. Somehow the performance of this code is 4 Times slower than the sequential version (omitting #pragma omp parallel for).
Is it possible that OpenMp has to create Threads every time the method is called? In my test it is called 10000 times directly after each other.
I heard that sometimes OpenMP will keep the threads spinning. I also tried setting OMP_WAIT_POLICY=active and GOMP_SPINCOUNT=INFINITE. When I remove the openMP pragmas, the code is about 10 times faster. Note that the method containing this code will be called 10000 times.
for (round k = 1; k < processor.max; ++k) {
initialise_round(k);
for (std::vector<int> bucket : color_buckets) {
#pragma omp parallel for schedule (dynamic)
for (int i = 0; i < bucket.size(); ++i) {
if (processor.mark.is_marked_item(bucket[i])) {
processor.process(k, bucket[i]);
}
}
processor.finish_round(k);
}
}
You say that your sequential code is much faster so this makes me think that your processor.process function has too few instructions and duration. This leads to the case where passing the data to each thread does not pay off (the data exchange overhead is simply larger than the actual computation on that thread).
Other than that, I think that parallelizing the middle loop won't affect the algorithm but increase the amount of work per thread/
I think you are creating a team of threads on each iteration of the loop... (although I'm not sure what for alone does - I thought it should be parallel for). In this case, it would probably be better to separate the parallel from the for so the work of forking and creating the threads is done just once rather than being repeated in the other loops. So you could try to put a parallel pragma before your outermost loop so the overhead of forking and thread creation is just done once.
The actual problem was not related to OpenMP directly.
As the system has two CPUs, half of the threads where spawned on one and the other half on the other CPU. Therefore there was not shared L3 Cache. This lead in combination that the algorithm doesn't scale well to a performance decrease especially when using 2-4 Threads.
The solution was to use thread pinning for example via the linux tool: taskset
I have a simple question about using OpenMP (with C++) that I hoped someone could help me with. I've included a small example below to illustrate my problem.
#include<iostream>
#include<vector>
#include<ctime>
#include<omp.h>
using namespace std;
int main(){
srand(time(NULL));//Seed random number generator
vector<int>v;//Create vector to hold random numbers in interval [0,9]
vector<int>d(10,0);//Vector to hold counts of each integer initialized to 0
for(int i=0;i<1e9;++i)
v.push_back(rand()%10);//Push back random numbers [0,9]
clock_t c=clock();
#pragma omp parallel for
for(int i=0;i<v.size();++i)
d[v[i]]+=1;//Count number stored at v[i]
cout<<"Seconds: "<<(clock()-c)/CLOCKS_PER_SEC<<endl;
for(vector<int>::iterator i=d.begin();i!=d.end();++i)
cout<<*i<<endl;
return 0;
}
The above code creates a vector v that contains 1 billion random integers in the range [0,9]. Then, the code loops through v counting how many instances of each different integer there is (i.e., how many ones are found in v, how many twos, etc.)
Each time a particular integer is encountered, it is counted by incrementing the appropriate element of a vector d. So, d[0] counts how many zeroes, d[6] counts how many sixes, and so on. Make sense so far?
My problem is when I try to make the counting loop parallel. Without the #pragma OpenMP statement, my code takes 20 seconds, yet with the pragma it takes over 60 seconds.
Clearly, I've misunderstood some concept relating to OpenMP (perhaps how data is shared/accessed?). Could someone explain my error please or point me in the direction of some insightful literature with appropriate keywords to help my search?
Your code exibits:
race conditions due to unsyncronised access to a shared variable
false and true sharing cache problems
wrong measurement of run time
Race conditions arise because you are concurrently updating the same elements of vector d in multiple threads. Comment out the srand() line and run your code several times with the same number of threads (but with more than one thread). Compare the outputs from different runs.
False sharing occurs when two threads write to memory locations that are close to one another as to result on the same cache line. This results in the cache line constantly bouncing from core to core or CPU to CPU in multisocket systems and excess of cache coherency messages. With 32 bytes per cache line 8 elements of the vector could fit in one cache line. With 64 bytes per cache line the whole vector d fits in one cache line. This makes the code slow on Core 2 processors and slightly slower (but not as slow as on Core 2) on Nehalem and post-Nehalem (e.g. Sandy Bridge) ones. True sharing occurs at those elements that are accesses by two or more threads at the same time. You should either put the increment in an OpenMP atomic construct (slow), use an array of OpenMP locks to protect access to elements of d (faster or slower, depending on your OpenMP runtime) or accumulate local values and then do a final synchronised reduction (fastest). The first one is implemented like this:
#pragma omp parallel for
for(int i=0;i<v.size();++i)
#pragma omp atomic
d[v[i]]+=1;//Count number stored at v[i]
The second is implemented like this:
omp_lock_t locks[10];
for (int i = 0; i < 10; i++)
omp_init_lock(&locks[i]);
#pragma omp parallel for
for(int i=0;i<v.size();++i)
{
int vv = v[i];
omp_set_lock(&locks[vv]);
d[vv]+=1;//Count number stored at v[i]
omp_unset_lock(&locks[vv]);
}
for (int i = 0; i < 10; i++)
omp_destroy_lock(&locks[i]);
(include omp.h to get access to the omp_* functions)
I leave it up to you to come up with an implementation of the third option.
You are measuring elapsed time using clock() but it measures the CPU time, not the runtime. If you have one thread running at 100% CPU usage for 1 second, then clock() would indicata an increase in CPU time of 1 second. If you have 8 threads running at 100% CPU usage for 1 second, clock() would indicate an increate in CPU time of 8 seconds (that is 8 threads times 1 CPU second per thread). Use omp_get_wtime() or gettimeofday() (or some other high resolution timer API) instead.
EDIT
Once your race condition is resolved via correct synchronization, then the following paragraph applies, before that your data race conditions unfortunately make speed comparisons mute:
Your program is slowing down because you have 10 possible outputs during the pragma section which are being accessed randomly. OpenMP cannot access any of those elements without a lock (which you would need to provide via synchronization) as a result and locking will cause your threads to have a higher overhead than you gain from counting in parallel.
A solution to make this speed up, is to instead make a local variable for each OpenMP thread which counts all of the 0-10 values that a particular thread has seen. Then sum those up in the master count vector. This will be easily parallelized and much faster as the threads don't need to lock on a shared write vector. I would expect a close to Nx speed up where N is the number of threads from OpenMP as there should be very limited locking required. This solution also avoids a lot of the race conditions currently in your code.
See http://software.intel.com/en-us/articles/use-thread-local-storage-to-reduce-synchronization/ for more details on thread local OpenMP
From what I understand, #pragma omp parallel and its variations basically execute the following block in a number of concurrent threads, which corresponds to the number of CPUs. When having nested parallelizations - parallel for within parallel for, parallel function within parallel function etc. - what happens on the inner parallelization?
I'm new to OpenMP, and the case I have in mind is probably rather trivial - multiplying a vector with a matrix. This is done in two nested for loops. Assuming the number of CPUs is smaller than the number of elements in the vector, is there any benefit in trying to run the inner loop in parallel? Will the total number of threads be larger than the number of CPUs, or will the inner loop be executed sequentially?
(1) Nested parallelism in OpenMP:
http://docs.oracle.com/cd/E19205-01/819-5270/aewbc/index.html
You need to turn on nested parallelism by setting OMP_NESTED or omp_set_nested because many implementations turn off this feature by default, even some implementations didn't support nested parallelism fully. If turned on, whenever you meet parallel for, OpenMP will create the number of threads as defined in OMP_NUM_THREADS. So, if 2-level parallelism, the total number of threads would be N^2, where N = OMP_NUM_THREADS.
Such nested parallelism will cause oversubscription, (i.e., the number of busy threads is greater than the cores), which may degrade the speedup. In an extreme case, where nested parallelism is called recursively, threads could be bloated (e.g., creating 1000s threads), and computer just wastes time for context switching. In such case, you may control the number of threads dynamically by setting omp_set_dynamic.
(2) An example of matrix-vector multiplication: the code would look like:
// Input: A(N by M), B(M by 1)
// Output: C(N by 1)
for (int i = 0; i < N; ++i)
for (int j = 0; j < M; ++j)
C[i] += A[i][j] * B[j];
In general, parallelizing inner loops while outer loops are possible is bad because of forking/joining overhead of threads. (though many OpenMP implementations pre-create threads, it still requires some to dispatch tasks to threads and to call implicit barrier at the end of parallel-for)
Your concern is the case of where N < # of CPU. Yes, right, in this case, the speedup would be limited by N, and letting nested parallelism will definitely have benefits.
However, then the code would cause oversubscription if N is sufficiently large. I'm just thinking the following solutions:
Changing the loop structure so that only 1-level loop exists. (It looks doable)
Specializing the code: if N is small, then do nested parallelism, otherwise don't do that.
Nested parallelism with omp_set_dynamic. But, please make it sure how omp_set_dynamic controls the number of threads and the activity of threads. Implementations may vary.
For something like dense linear algebra, where all the potential parallelism is already lain bare in one place in nice wide for loops, you don't need nested parallism -- if you do want to protect against the case of having (say) really narrow matricies where the leading dimension might be smaller than the number of cores, then all you need is the collapse directive which notionally flattens the multiple loops into one.
Nested parallelism is for those cases where the parallelism isn't all exposed at once -- say you want to do 2 simultaneous function evaluations, each of which could usefully utilize 4 cores, and you have an 8 core system. You call the function in a parallel section, and within the function definition there is an additional, say, parallel for.
At the outer level use the NUM_THREADS(num_groups) clause to set the number of threads to use. If your outer loop has a count N, and the number of processors or cores is num_cores, use num_groups = min(N,num_cores). At the inner level, you need to set the number of sub-threads for each thread group so that the total number of subthreads equals the number of cores. So if num_cores = 8, N = 4, then num_groups = 4. At the lower level each sub-thread should use 2 threads (since 2+2+2+2 = 8) so use the NUM_THREADS(2) clause. You can collect the number of sub-threads into an array with one element per outer region thread (with num_groups elements).
This strategy always makes optimal use of your cores. When N < num_cores some nested parallelisation occurs. When N >= num_cores the array of subthread counts contains all 1s and so the inner loop is effectively serial.