[Mpi3-rma] RMA chapter for quick review

Torsten Hoefler htor at illinois.edu
Fri Mar 11 21:55:40 CST 2011


> Re:
>>> [1, 35-38]: In the list: "MPI_PUT (remote write), MPI_GET (remote read),
>>> [and] MPI_ACCUMULATE (remote update), MPI_GET_ACCUMULATE,
>>> (remote read and update), MPI_COMPARE_AND_SWAP (remote atomic swap
>>> operations), MPI_RPUT, MPI_RGET, MPI_RACCUMULATE and
>>> it is odd that a subset of the operations have parenthetical expressions
>>> that indicate the nature of the functionality. Either all operations
>>> should be followed with a parenthetical explanation or none of them.
>>> The right solution is probably to reformat into a list with each
>>> item being the group of operations that provide functionality of the
>>> same nature (e.g., I am guessing that MPI_GET and MPI_RGET are both
>>> remote read operations). Initially I thought the ones that are not
>>> followed by a parenthetical expression were being grouped with the
>>> following one that has the parenthetical expression (so MPI_GET_ACCUMULATE
>>> is a remote read and update operation). However, the list ends with
>>> several operations with no following parenthetical expression. Anyway,
>>> the implicit grouping is unclear. Making it explicit would be better.
>>> It would also simplify how the following sentence ("When a reference is
>>> made to “accumulate” operations in this chapter, it refers to the
>>> following operations: MPI_ACCUMULATE, MPI_GET_ACCUMULATE,
>> Ok, I added the descriptions to the R-operations.
> Hmm. the more I thought about the more I though this would
> be better as a bulleted list, like the one about the types
> of collective operations. I think that instead of "The
> following communication calls are provided:" (etc.) something
> like this would be best:
> ----------------------------------------------------------------
> One-sided operations can be classified as follows:
> * Remote read: MPI_GET, MPI_RGET;
> * Remote write: MPI_PUT, MPI_RPUT;
> * Remote read and update: MPI_GET_ACCUMULATE, MPI_FETCH_AND_OP,
>                           MPI_RGET_ACCUMULATE;
> * Remote atomic swap operations: MPI_COMPARE_AND_SWAP.
> This chapter refers to operations set that includes all remote
> update, remote read and update, and remote atomic swap operations
> as “accumulate” operations.
Ok, this is now in the style of the collectives chapter (slightly
different than you propose but the same gist). 

> -------------------------------------------------------------------
> This version makes it much clearer that grouping is intended
> and makes the definition of accumulate operations simpler.
> It also makes it clearer that the ommission of an MPI_RCOMPARE_AND_SWAP 
> operation is intentional.

> [snip]
>>> [2, 12] Use active voice: "However, support for asynchronous communication
>>> agents in software (handlers, threads, etc.) might be needed, for certain
>>> RMA functions, in a distributed memory environment." => "However, certain
>>> RMA functions might need support for asynchronous communication agents in
>>> software (handlers, threads, etc.) in a distributed memory environment."
>> I'll defer this one for now because it doesn't seem essential and it's a
>> macro-mess :).
> I'm not sure that I understand.
The cost/benefit of this change killed it from my perspective ;).

>> Generally, I fear that it will be a lot of work if we're trying to fix
>> all the little wording things in the OS chapter. Many of them (most of
>> the ones you found) have been there since MPI-2 and are in the places
>> that we didn't want to touch. It might be better to not apply such
>> changes at this point so that reviewers can focus on the semantic
>> changes (and not have colored sections all over the document).
> Hmm. I will need to read through the whole chapter in order
> to vote on the changes. In that process, I will hit these
> types of corrections. How would you prefer that we handle them?
Yes, please record them for now. I think we need to make a decision in
the working group how to deal with them. They will be massive (I found
many of those while reading the chapter).

Thanks & All the Best,

 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Torsten Hoefler         | Performance Modeling and Simulation Lead
Blue Waters Directorate | University of Illinois (UIUC)
1205 W Clark Street     | Urbana, IL, 61801
NCSA Building           | +01 (217) 244-7736

More information about the mpiwg-rma mailing list