Re: [Boost-users] mpi/serialization: broadcast function and the value argument
-----Original Message----- From: "Matthias Troyer" [troyer_at_[hidden]] Date: 09/12/2010 02:39 PM To: boost-users_at_[hidden] Subject: Re: [Boost-users] mpi/serialization: broadcast function and the value argument
On Dec 8, 2010, at 5:50 PM, Hicham Mouline wrote:
Following a suggestion from another thread, I have tried broadcasting in a polymorphic way.
I have a hierarchy of polymorphic struct (they have 2 virtual functions). I was wondering if it was possible/allowed to
root process: base* b; /// b points to an actual derived struct but may also point the base struct (not abstract) broadcast( ..., b, ...);
slave processes: base* b; broadcast( ..., b, ...); /// broadcast constructs the appropriate struct and makes b point to where it did that
assuming all the conditions for pointer to class conditions are satisfied for the serialization library.
If serialization works on your polymorphic class then this should work as well. Matthias
It appears the implementation of mpi::broadcast(), according to this post, http://lists.boost.org/boost-users/2010/12/64802.php and I confirm that expects an already built object to call the (de)serialization procedure on. Therefore, this seems a specific issue that mpi::broadcast() may want to handle differently, could you confirm/correct? thanks,
I would very much appreciate extra opinions about this issue and the route I have taken to solve it. Let me sum up: mpi::broadcast() appears to be incapable of sending the argument "value" in a polymorphic way, ie: sender------------------- Base* base; If I call boost::mpi::broadcast(..., base, ...) receiver----------------- Base* base; boost::mpi::broadcast(..., base, ...) This does not construct the most derived type of base on the heap and make base point there, as serialization would do. I therefore used a variant made of all the classes of the hierarchy and broadcast it instead. sender------------------- variant<...> v; boost::mpi::broadcast(..., v, ...) receiver----------------- variant<...> v; boost::mpi::broadcast(..., v, ...) This works but before I replicate it in other parts of the code, I would appreciate your opinions about this solution provided by Riccardo. Regards,
On Dec 13, 2010, at 9:31 PM, Hicham Mouline wrote:
-----Original Message----- From: "Matthias Troyer" [troyer_at_[hidden]] Date: 09/12/2010 02:39 PM To: boost-users_at_[hidden] Subject: Re: [Boost-users] mpi/serialization: broadcast function and the value argument
On Dec 8, 2010, at 5:50 PM, Hicham Mouline wrote:
Following a suggestion from another thread, I have tried broadcasting in a polymorphic way.
I have a hierarchy of polymorphic struct (they have 2 virtual functions). I was wondering if it was possible/allowed to
root process: base* b; /// b points to an actual derived struct but may also point the base struct (not abstract) broadcast( ..., b, ...);
slave processes: base* b; broadcast( ..., b, ...); /// broadcast constructs the appropriate struct and makes b point to where it did that
assuming all the conditions for pointer to class conditions are satisfied for the serialization library.
If serialization works on your polymorphic class then this should work as well. Matthias
It appears the implementation of mpi::broadcast(), according to this post, http://lists.boost.org/boost-users/2010/12/64802.php and I confirm that expects an already built object to call the (de)serialization procedure on. Therefore, this seems a specific issue that mpi::broadcast() may want to handle differently, could you confirm/correct? thanks,
I would very much appreciate extra opinions about this issue and the route I have taken to solve it. Let me sum up:
mpi::broadcast() appears to be incapable of sending the argument "value" in a polymorphic way, ie:
sender------------------- Base* base; If I call boost::mpi::broadcast(..., base, ...)
receiver----------------- Base* base; boost::mpi::broadcast(..., base, ...) This does not construct the most derived type of base on the heap and make base point there, as serialization would do.
Did you register the most derived type? Did you test whether pointer deserialization works with other serialization archives? Matthias
participants (2)
-
Hicham Mouline
-
Matthias Troyer