[Forwarding for archival] -------- Original Message -------- Subject: Subcommittee assignments for Parallel Interface Definition Date: Wed, 03 Oct 2007 14:37:11 -0600 From: Devine, Karen D. <[email protected]> To: Tim Tautges <[email protected]>, Mark Shephard <[email protected]>, Kenneth Jansen <[email protected]>, Lori A. Diachin <[email protected]>, Vitus Leung <[email protected]>, Carl Ollivier-Gooch <[email protected]>, Karen Devine <[email protected]>, Onkar Sahni <[email protected]> Honored selectees: As promised, here are the assignments for draft zero of the parallel interface. Our task due by November 16 is to build the interface for Sections 3.2 (Basic Parallelism) and 3.3 (Partitions) of the requirements document. There are four categories of interface needed. - Global queries (including MPI Communicators) - Global IDs - Partition Model - Partitioned Mesh Here's the plan: Step 0: draft what functions are needed in each category; due Oct 15. Step 1: agree on what functions are needed in each category; due Oct. 18. Step 2: formalize the syntax for those functions; due Oct 25. Step 3: review and revise output from Steps 1 and 2, and write first draft; due Nov. 16. Below is an outline of what should be included in each category. The requirement numbers are those used in CVS Version 1.14 of the requirements document, distributed October 2. Global Queries: - Req-5: MPI Communicators We did not agree at the bootcamp whether MPI communicators should be set once or passed to each function. It seems natural that an MPI communicator be associated with a mesh; that is, a mesh would know which processors contained any part of the mesh. Different communicators could be passed to services like load balancing, but in general, a query like "give me the global number of element in this mesh" would use the mesh's communicator. So should we have one function that associates a communicator with a mesh? Or should both a communicator and a mesh be associated with a partition? Or are there cases where separate communicators would be needed for each function? Or...? Also here, we should probably decide what a root set really is in parallel. Should there be one root set per processor, with functions invoked on the root set returning only info on the processor? If so, should there be an additional layer above the root set representing the global mesh? Just for my understanding, is a root set equivalent to a mesh instance? If, say, a multiphysics application has two meshes, does it have two root sets? - Req-6: Global queries Which queries should we support? Should they all be associated with a mesh? - This category also enforces requirements Req-2, Req-3, and Req-4. Global IDs: - Req-7: Global IDs We agreed at bootcamp that global IDs would be represented as an array of integers. We need to define the data types and functions on global IDs, as well as their immutability. Partition Model - We decided to have a separate iPart handle that would be passed to ITAPS functions. We'll also need some functions that return information about a partition. These functions are required by Req-10 to Req-13. Partitioned Mesh - We decided that iPart handles could be passed to many ITAPS functions in place of Entity handles. We need to now examine whether that decision was correct. So for this section, we need to discuss how easy substitution of iPart handles will be and what additional capability will be needed. This section deals with Req-14 to Req-18. OK, now the exciting part -- the part you've all been waiting for -- your assignments! We'll divide and conquer, having each group take care of one category as follows. Global Queries: Argonne (Tim) Global IDs: Sandia (Vitus, Karen) Partition Model: RPI (Mark, Ken, Onkar) Partitioned Mesh: UBC (Carl) By Oct 15, please complete Step 0 and send the output to this distribution list for review and discussion. Thanks. Karen -- ================================================================ "You will keep in perfect peace him whose mind is steadfast, because he trusts in you." Isaiah 26:3 Tim Tautges Argonne National Laboratory ([email protected]) (telecommuting from UW-Madison) phone: (608) 263-8485 1500 Engineering Dr. fax: (608) 263-4499 Madison, WI 53706