Re: pthreads and UPC ....

From: Tahar Amari (amari_at_cpht.polytechnique.fr)
Date: Sat Feb 14 2009 - 08:16:37 PST

  • Next message: Tahar Amari: "shared data and operator"
    Kathy,
    
    Thanks a lot, this is very helpful  and makes me feel
    much more comfortable.
    
    Tahar
    
    
    --------------------------------------------
    T. Amari
    Centre de Physique Theorique
    Ecole Polytechnique
    91128 Palaiseau Cedex France
    tel : 33 1 69 33 42 52
    fax: 33 1 69 33 30 08
    email: <mailto:amari@cpht.polytechnique.fr>
    URL : http://www.cpht.polytechnique.fr/cpht/amari
    
    Le 14 févr. 09 à 15:57, Kathy Yelick a écrit :
    
    > Tahar,
    >
    > There are multiple UPC implementations, so it's often the case that  
    > even if one does have a bug, another one will work.  Here's a list  
    > of the compilers -- I hope someone will correct me if I've missed one:
    >
    > Berkeley UPC, bupc (upc.lbl.gov)
    > Intrepid, gcc-upc (intrepid.com)
    > HP UPC (h30097.www3.hp.com/upc)
    > IBM UPC (http://www.alphaworks.ibm.com/tech/upccompiler)
    > Cray UPC (www.cray.com)
    >
    > The first two are open source, so in principle the community at  
    > large can also fix bugs in them.  Both also have active support  
    > teams and there are regular (at least annual) releases of both.    
    > The other three are closed source commercial products and also have  
    > regular releases and ongoing support.
    >
    > I think the quote you mentioned was only said once and later  
    > retracted.  In particular, the Berkeley team does have people fixing  
    > bugs,  so this kind of thing will be fixed, but as with most open  
    > source projects, not necessarily within a few days.  This is true  
    > with vanilla gcc today, so you should not feel less comfortable  
    > about programming UPC than with C or C++.  In both cases you can buy  
    > commercial compilers that come with higher level support guarantees  
    > or use open source ones in which problems get fixed, but not  
    > necessarily on a particular time scale.  The Berkeley and Intrepid  
    > compilers are also highly portable and should run on any machine  
    > with shared memory (pthreads, as you're discussing) or distributed  
    > memory (using GASNet).
    >
    >    Kathy
    >
    > ------------------------------------------------
    > Katherine Yelick
    > NERSC Division Director
    > Lawrence Berkeley National Laboratory
    > One Cyclotron Rd., MS 50B-4230
    > Berkeley, CA  94720
    > Email: kayelick_at_lbl_dot_gov
    > Phone: 510-495-2431
    >
    > On Feb 14, 2009, at 4:58 AM, Tahar Amari wrote:
    >
    >> Hello Paul,
    >>
    >> Many thanks for this detailed reply.
    >>
    >> SInce I am  new with UPC, it is not a criticism at all but simply  
    >> some thoughts before starting.
    >> UPC seems certainlbut to be much easier to use than MPI.
    >>
    >> I worry a little bit about bugs, and so how reliable it would be to  
    >> put our projects to UPC.
    >> I saw on the FAQs that there were some bugs and sometime some reply  
    >> like " thanks for this bug
    >> , we do not have anybody working on this right now , but will  
    >> consider it the future".
    >>
    >> Is is the same for all implementation, IBM, HP ?
    >>
    >> Please understand again that I just want to enquire before making  
    >> the "big jump" because
    >> I feel very interested by UPC, but do not want to face compiler  
    >> problems which I do not control,
    >> when I hardly yet control the numerical or physical problems.
    >>
    >> Thank you very much for your impression,
    >>
    >> Tahar
    >>
    >>
    >>
    >>
    >> URL : http://www.cpht.polytechnique.fr/cpht/amari
    >>
    >> Le 13 févr. 09 à 22:24, Paul H. Hargrove a écrit :
    >>
    >>> Tahar,
    >>> I regret that nobody appears to have replied to your questions yet  
    >>> (or did not cc:ed the list if they did).
    >>> The answer to your question is not a simple "Yes", but more a  
    >>> "probably".
    >>>
    >>> Out runtime can optionally use pthreads internally for  
    >>> implementing UPC threads within a single node, both with and  
    >>> without a network.  So, there is no fundamental problem with  
    >>> linking the pthreads libraries.  Additionally, I am aware of at  
    >>> least on project that uses pthreads in *addition* to UPC threads  
    >>> (as I suspect you are asking).  In that case, each individual UPC  
    >>> thread retains its own identity and can spawn pthreads to execute  
    >>> C (or C++ or FORTRAN, I suppose) code, but the pthreads spawned by  
    >>> the user code cannot execute UPC code because they do not  
    >>> correspond to any MYTHREAD value.
    >>> If you have more specific information about how you wish to mix  
    >>> pthreads with UPC, let us know and we might be able to clarify  
    >>> further.
    >>>
    >>> -Paul
    >>>
    >>> Tahar Amari wrote:
    >>>> Hello,
    >>>>
    >>>> One of my code use Pthreads.
    >>>> Is it possible to still use this piece of code with UPC ?
    >>>>
    >>>> Many thanks
    >>>>
    >>>> Tahar
    >>>
    >>>
    >>> -- 
    >>> Paul H. Hargrove                          PHHargrove_at_lbl_dot_gov
    >>> Future Technologies Group                 Tel: +1-510-495-2352
    >>> HPC Research Department                   Fax: +1-510-486-6900
    >>> Lawrence Berkeley National Laboratory
    >>
    

  • Next message: Tahar Amari: "shared data and operator"