Is libecm thread-safe?

[frmky]

I'm trying to extend Paul Zimmermann's aliq2.c from http://www.loria.fr/~zimmerma/records/aliquot.html to run the ECM on multiple threads. I've done this:

#pragma omp parallel for 
          for (i=0;i<nt;i++) { 
            //printf("Thread: %d\n",i);
            ress[i] = ecm_factor(ps[i], ns[i], B1, NULL); 
          }

where nt is the number of threads running, and ps and ns are arrays of mpz's. If I compile this with openmp and choose only 1 thread at runtime, all works correctly. Selecting more than one thread, however, leads to a lockup, segfault, or trying to free() something already free()'ed. Is ecm_factor() not thread-safe? If not, would it be difficult to make it so?

Thanks,
Greg

[jasonp]

Are you sure OpenMP spawn multiple threads, so that each function call gets its own stack? Maybe you also need an array of ecm_params entries, initialized before running the loop in parallel?

[frmky]

Originally Posted by jasonp

Are you sure OpenMP spawn multiple threads, so that each function call gets its own stack?

It uses pthreads to spawn the threads.

Originally Posted by jasonp

Maybe you also need an array of ecm_params entries, initialized before running the loop in parallel?

I'll figure out how to do that in a few days. I also sometime plan to see if I can figure out how to use libmsieve as well.

Greg

[fivemack]

Originally Posted by frmky

Is ecm_factor() not thread-safe? If not, would it be difficult to make it so?

It is not thread-safe - there is a single global prime-generating object which is not protected by locks (so your multiple-frees come when thread 1 has deleted the sieving array while thread 2 is trying to read from it), and if you add locks you get the obvious problem that thread 1 picks up some primes and thread 2 others when computing the first product of primes. I'm not sure how fiddly it is to fix, but fiddlier than I'd first thought.

[jasonp]

Originally Posted by frmky

I'll figure out how to do that in a few days. I also sometime plan to see if I can figure out how to use libmsieve as well.

libmsieve should be completely thread safe, if you write to different log files and data files with each instance. The library has no writeable globals.

[akruppa]

Most of GMP-ECM has been written without thread-safeness (or indeed a library) in mind, so I'm not surprised it isn't... the prime sieve's state in static variables is obviously a problem, but I think there are some other static variables, too. (Never mind the brilliant idea of putting temp space in the mpmod struct, or mixing mpz/mpmod for temp data... half of the code should be rewritten from scratch. )

Alex