[mvapich-discuss] Re: data placement in-order guarantee ?

[Håkon Bugge]

On Feb 9, 2010, at 19:08 , अनुज wrote:
> On Tue, Feb 2, 2010 at 9:35 PM, Håkon Bugge <Haakon.Bugge at sun.com> wrote:
> [snip]
> Now, most MPI implementation does not obey this; IIRC MVAPICH does rely on the last byte (word?) being written after all other payload,
> 
> But MVAPICH assumes last byte is written in last. What if that is not the case?

I would assume you in that case copy incorrect data from your IB receive buffer to your application receive buffer (assuming a one-copy receive). This can happen with a less conservative IB implementation and/or weak ordering of I/O transactions.
>  
> Scali MPI relies on 4-byte write atomicy and uses a special encoding/decoding which sustains any reordering of the payload (as long as each 4-byte word is written as an atomic unit). The benefit, besides robustness, is the capability of a worm-hole message consumption (the message can be decoded and copied into the MPI user buffer _before_ it has been fully received). This method is currently a patent application, so I do not recommend you to copy-cat this method.
> 
> Can u provide a link for this method? Atleast I can go through some details of this technique. 

Go to http://portal.uspto.gov/external/portal/pair, have a look at application 11/256752 and 11/437902. As stated above, you should understand that these techniques are protected by patent legislation.



Håkon

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20100209/d6a7dd7c/attachment.html