On 1/29/08 9:54 AM, "Sebastian Gesemann" <s.gesemann@gmail.com> wrote:

You may want to try http://threadpool.sourceforge.net/

I was looking into that, but was wondering if anyone had experience to suggest that creation/destruction of threads was a significant overhead before I implement pools. James

Sorry, I have to amend my statement. I just ran an experiment which launched 10000 threads vs launching 4 threads and doing the same amount of work. Results seem to indicate that the former is quite expensive on Linux too -- about 2 orders of magnitude different in my test. Note that some of this effect could also be from side effects from time slicing, etc... in the kernel scheduler. Brian On Jan 29, 2008 11:25 AM, Brian Budge <brian.budge@gmail.com> wrote:

Hi James -

My impression was that creating new threads on Windows can be quite expensive, but I don't have a lot of experience there. Creating threads in Linux using posix_threads (also via boost) seems to be quite fast.

However, I have noticed that thread_groups are much slower (factor of 4) than using my own vector<boost::thread*> and joining each individually. This is probably because I have only one controlling thread for my vector, so I don't need mutex.

Brian

On Jan 29, 2008 10:28 AM, James Sutherland <James.Sutherland@utah.edu> wrote:

...

On 1/29/08 9:54 AM, "Sebastian Gesemann" <s.gesemann@gmail.com> wrote:

...

You may want to try http://threadpool.sourceforge.net/

I was looking into that, but was wondering if anyone had experience to suggest that creation/destruction of threads was a significant overhead before I implement pools.

James _______________________________________________ Boost-users mailing list Boost-users@lists.boost.org http://lists.boost.org/mailman/listinfo.cgi/boost-users

I have tried threadpool. It is my understanding that the code fragment below should run in constant time independent of the number of threads (assuming that the ³Test::operator()² method runs in constant time). However, I observe an increase in runtime as I increase the pool size from say 1 to 5. Any thoughts on this? James

int main() { const int nt = 5; const int nit = 10000;

time_t t1 = clock();

boost::threadpool::pool p(1); for( int j=0; j<nit; ++j ){ for( int i=0; i<nt; ++i ){ p.schedule( Test(i) ); } p.wait(); }

std::cout << "t=" << difftime(clock(), t1) << std::endl; return 0; }