On 4/16/11 12:50 PM, "William Gropp" <[email protected]> wrote:

I found a few changes from MPI 2.2 that I missed the first time. I've also change the handling of the few ticket o items to use consistent colors (green is delete, red is inserted). A few of the items from Jim's notes were updated; someone needs to look at the others and at least propose specific updates.

All - I have Jeff's changes and some of my own that we need to get inserted. Since I don't have access to the tex file, do you want me to send the long list? I'd really rather just modify the tex directly, as it will take less time, even if someone else wants to continue managing commits to the svn. As for the examples, it's pretty obvious that after the store in Example 11.7, we'd need an MPI_FOO() call to act as a memory barrier. The other way to handle the issue in the current definition would be to be in a lock shared epoch and call win_sync, in which case the example is valid for both the unified and separate models. Less clear is what to do with example 11.9. Similar to the store in Example 11.7, some synchronization is necessary for the load in Process B. One could insert a MPI_FOO call before the load and give FOO both load and store synchronization, of course. One could also be in a lock shared epoch and call win sync before the load, in which case the example is valid for both the unified and separate models. I'm actually leaning towards the second option; it doesn't require us delving into language dependent semantics, it synchronizes the local and remote views of the window, which is a semantic that we know how to provide and is fairly language independent. Brian -- Brian W. Barrett Dept. 1423: Scalable System Software Sandia National Laboratories