Re: [Quantlib-users] Regarding OpenMP and Open MPI library support

[Peter C. <pca...@gm...>]

Hi,

in the past we saw that a "blind" usage of open mp does more harm than good, see

https://sourceforge.net/p/quantlib/mailman/message/32461242/

so we removed most of the pragmas. It seems that the omp pragma
survived in lattice.hpp contrary to what is said in the mail thread, I
don't remember if there is a good reason for that. Maybe something to
review at some point. There are two more places, zabrsmilesection.hpp
and gaussian1dswaptionengine.cpp where I think enabling openmp gives
_some_ speedup although it's far from what you would hope for if I
remember correctly.

grep --include="*.?pp"  -nH --null -r -e "pragma omp" *
ql/experimental/volatility/zabrsmilesection.hpp209:#pragma omp parallel for
ql/methods/lattices/lattice.hpp169:        #pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp127:
          //#pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp137:
          //#pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp153:
      //#pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp172:
      //#pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp187:
  //#pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp203:
  //#pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp220:
  #pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp238:
  //#pragma omp parallel for
ql/methods/finitedifferences/operators/triplebandlinearop.cpp260:
  //#pragma omp parallel for
ql/methods/finitedifferences/operators/ninepointlinearop.cpp147:
 //#pragma omp parallel for
ql/methods/finitedifferences/operators/ninepointlinearop.cpp191:
 //#pragma omp parallel for
ql/methods/finitedifferences/parallelevolver.hpp51:
//#pragma omp parallel for
ql/methods/finitedifferences/parallelevolver.hpp102:
//#pragma omp parallel for
ql/methods/finitedifferences/stepcondition.hpp48:            //#pragma
omp parallel for
ql/pricingengines/swaption/gaussian1dswaptionengine.cpp118:#pragma omp
parallel for default(shared) firstprivate(p) if(expiry0>settlement)

Thanks
Peter

On Thu, 6 Jan 2022 at 13:18, Jonathan Sweemer <sw...@gm...> wrote:
>
> Hi Vamshi,
>
> OpenMP is only wired into a few places in QuantLib so if your program doesn't use those specific code paths then you probably won't see much difference in terms of speed.
>
> Moreover, QuantLib is not thread safe in general, and wasn't specifically designed for thread-level parallelism. See Luigi's answer on Stack Overflow for more details: https://stackoverflow.com/a/47098133
>
> The good news is that you can use multiprocessing instead of multithreading with your HPC cluster to parallelize your jobs across cores and nodes.
>
>
> On Thu, Jan 6, 2022 at 12:28 PM Vamshi Krishna <kri...@gm...> wrote:
>>
>> Hiii Users
>>
>> I am new in this group and new to quantlib. I come for HPC back ground.
>>
>> I have tried some examples program with openMP but no major performance gained.
>>
>> When I tried FinanceBench program which which have openMP functions and shown good performance on multi cores.
>>
>> I do not, I am do correctly or not.
>>
>> If any user could help me "how to use quantlib on multi cores and multiple node hpc cluster,  I will be thankful.
>>
>> Regards
>> Vamshi Krishna
>> _______________________________________________
>> QuantLib-users mailing list
>> Qua...@li...
>> https://lists.sourceforge.net/lists/listinfo/quantlib-users
>
> _______________________________________________
> QuantLib-users mailing list
> Qua...@li...
> https://lists.sourceforge.net/lists/listinfo/quantlib-users