I have a simple program that I am using for physics simulation. I want to know how to implement a certain threading paradigm in OpenMP.
int main() { #define steps (100000) for (int t = 0;t < steps; t++) { firstParallelLoop(); secondParallelLoop(); if (!(t%100)) { checkpoint(); } } } void firstParallelLoop() {// In another file.c #pragma omp parallel for for (int i = 0; i < sizeOfSim;i++) { //Some atomic floating point ops. } }
Formerly, I was using pthreads and got a 1.7 speedup on my dualcore laptop. I can't seem to get any speedup when using OpenMP. I suspect the problem is that the thread groups/pools are rapidly being created and destroyed with disasterous effect.
In my pthreads implementations I needed to ensure that no new threads were created, and that my program behaved as a client-server. In the pthreads scheme, the main() was a server and calls to firstParallelLoop would release mutexes/semaphores that triggered the thread to reprocess the data.
When I look at CPU utilization I expect it to be over the 30% mark (4 core, 2 are HT), but it stays around 27...
How do I get OpenMP to do something similar? How can I tell OpenMP to reuse my threads?
The GCC OpenMP run-time libgomp implements thread teams on POSIX systems by something akin to a thread pool - threads are only created when the first parallel region is encountered, with each thread running an infinite work loop. Entering and exiting a parallel region is implemented with barriers. By default libgomp uses a combination of busy-waiting and sleeping to implement barriers. The amount of busy-waiting is controlled by the OMP_WAIT_POLICY environment variable. If it is not specified, threads that wait on a barrier would busy-wait for 300000 spins (3 ms at 100000 spins/msec) and then would go into sleeping state. If OMP_WAIT_POLICY is set to active, then the busy-wait time is increased to 30000000000 spins (5 mins at 100000 spins/sec). You can fine tune the busy-waiting time by setting the GOMP_SPINCOUNT variable to the number of busy cycles (libgomp assumes about 100000 spins/msec but it could vary by a factor of 5 depending on the CPU). You can fully disable sleeping like this:
OMP_WAIT_POLICY=active GOMP_SPINCOUNT=infinite OMP_NUM_THREADS=... ./program
This would somehow improve the thread team starting time, but at the expense of CPU time as idle threads would not idle but rather busy-wait.
In order to remove the overhead you should rewrite your program in more OpenMP-friendly way. Your example code could be rewritten like this:
int main() { #define steps (100000) #pragma omp parallel { for (int t = 0; t < steps; t++) { firstParallelLoop(); secondParallelLoop(); if (!(t%100)) { #pragma omp master checkpoint(); #pragma omp barrier } } } } void firstParallelLoop() {// In another file.c #pragma omp for for (int i = 0; i < sizeOfSim; i++) { //Some atomic floating point ops. } }
Note the following two things:
- A parallel region is inserted in the main program. It is not a
parallel for though. All threads in the team would execute the outer loop steps times. - The
for loop in firstParallelLoop is made parallel by using omp for only. Thus it will execute as a serial loop if called outside an OpenMP parallel and as parallel when called from inside a parallel region. The same should be done for the loop in secondParallelLoop.
The barrier in the main loop is used to ensure that other threads would wait for the checkpoint to finish before starting the next iteration.
翻译