[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [mpi-21] Call for MPI 2.2 and 3.0 agenda items for the Jan mee ting
- To: mpi-21@xxxxxxxxxxxxx
- Subject: Re: [mpi-21] Call for MPI 2.2 and 3.0 agenda items for the Jan mee ting
- From: "William Yu" <wyu@xxxxxxxxxx>
- Date: Mon, 3 Dec 2007 06:21:40 +0800
- Delivered-to: mpifrm-mpi-21-outgoing@mailbouncer.mcs.anl.gov
- Delivered-to: mpifrm-mpi-21@mailbouncer.mcs.anl.gov
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=domainkey-signature:received:received:from:reply-to:to:subject:date:message-id:x-mailer:mime-version:content-language:content-type:content-transfer-encoding:sender; bh=oWMNVctEmRbCY5Wkxyi6jJ5AcK975id+X2SlwmwJlfk=; b=fs/ZqElNupOWCfMBR0fwPg33f0HTAqFp6yz1AUcQOzoj8v8fnPCftrj0NkZVT6l+ZMW3dC8UFDVoxSWea5x7kbxPRQqfkCNlC/0YWwr951Qq5J44Xht6zXW5vp60nnr6Nr+v7WV16ZtmtcA2+djPSg+9hVg78CCUMY/ah5jqilM=
- Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=received:from:reply-to:to:subject:date:message-id:x-mailer:mime-version:content-language:content-type:content-transfer-encoding:sender; b=IkqBjPapEeivCssPNj6xhrNu4eHxaTMiYircOlgwax8wXG5Byk9ZNwfPQT1jn6c1nzYJ25KUtDmWbkIcpnR0gBgcbhK/v3gVkGHhpmVzcch7YjUD8uK1GW+SgraYtTs8sOLX6GJldxHgAqj321e8St9+wP/+57wVHb4QYXirJRM=
- Reply-to: mpi-21@xxxxxxxxxxxxx
- Sender: owner-mpi-21@xxxxxxxxxxxxx
This is true. Maybe interest is presently not enough. Even the existing java mpi projects seem to be in hibernation.
In particular, i like how boost deals with custom data types (its own serialization mechanism). Also it does have some hooks that allow "no copy" efficiency.
These are currently not present with Java MPI bindings that use JNI. It is my understanding, based on my experience with JNI, that all data is "copied" at the interface. This makes JNI really expensive in terms of space.
________________ Reply Header ________________
Subject: Re: [mpi-21] Call for MPI 2.2 and 3.0 agenda items for the Jan mee ting
Author: Douglas Gregor <dgregor@xxxxxxxxxx>
Date: December 02nd 2007 4:05 pm
On Dec 2, 2007, at 8:43 AM, Jeff Squyres wrote:
> On Dec 1, 2007, at 8:25 PM, William Yu wrote:
>
>> On the boost front, I liked how they implemented their MPI
>> interface that takes advantage of C++ language constructs over and
>> above a simple overlay or wrapper.
>>
>> However, boost's implementation is boost specific right? I can't
>> compile the same code to another library because it is not
>> standardized. Fortunately, boost is open source.
>
> Boost is a standalone system that (as I understand it) works on top
> of most/any underlying MPI implementation/s.
Yes, it's standalone and builds on top of the C MPI bindings, so it
should work with any MPI 1.1-compliant implementation.
> It is not a standard; I do not believe that it was designed to be
> interoperable with anything.
I'm not sure what "interoperable" means. It interoperates with C and
Fortran for anything that is described with an MPI datatype (either
primitive or derived); for types that require serialization (linked
lists and such), the Boost implementation does its own thing. It
would be relatively simple to update the interface to eliminate the
explicit dependencies on Boost (there are few of these), but all this
is only relevant if the committee is interested in reconsidering the
official C++ bindings to MPI.
- Doug