Re: [Boost-users] [Boost.MPI] SendingReceiving an array "view"

Subject: Re: [Boost-users] [Boost.MPI] SendingReceiving an array "view"
From: Matthias Troyer (troyer_at_[hidden])
Date: 2009-05-27 11:24:12

On 11 May 2009, at 06:34, Joel Falcou wrote:

> Joel Falcou wrote:
>> I have a small class view<T> that containes a T* and a size. When I  
>> send/receive view using Boost.MPI, I want transfert the actual n  
>> data pointed by the view pointer and send the size. On receiving, i  
>> will receive the size, checks if they are compatible and receive  
>> the data into the zone pointed by the pointer of the receiving  
>> view. Of course, I may do something more complex than send/Receive,  
>> like all_2_all etc. So i was thinking to give view<T> a proper  
>> serialize function so boost.serialization may kick in. As the  
>> behavior of sending and receiving is different, I use the load/save  
>> idiom to separate the behavior. The question is then, how to  
>> serialize the n elements poitned by a raw pointer to those data in  
>> an efficient way ?
> Back to this issue, now here is my use case.
>
> On my N MPI process, I have an array of float (let's say any  
> fundamental types if we want).
> All processor build a few "view" of this array to map a small subset  
> of contiguous data.
>
> float tab[]= {1,2,3,4,5,6,7,8,9,10};
> view<float> v0(&tab[0],5);
> view<float> v1(&tab[5],5);
>
> Then I have an array of optional<view<T>> that will contain v0 and v1.
>
> optionall< view<float> > data[2] = {v0,v1};
>
> What I want next is to use mpi::all_to_all to send those data across.
>
> all_to_all(data,data);
>
> In this example, what I want to have is that the data from v1 to be  
> send from P0 to P1 and data
> from v0 to be send from P1 to P0 *without copies if possible* by  
> hoping that the serialization will
> forward the data through the view pointer. view serialization is  
> done using make_array.
>
> If I won't use optional, it works. Once I put those in optional, the  
> serialization of optional actually overwrite the data
> by calling the in-place default constructor of view, which reset the  
> pointer to 0 and prevent proper serialization (no error but, at  
> runtime, the data just aren't transfered).
>
> Is there anything I missed or should I abandon the idea to use  
> make_array and roll on my own serialization routine ?

What you should do is use the skeleton+content mechanism, or just send  
a contents object of your view once you have created both views.

Matthias