Note that I've made a number of changes to the test suite, including ones related to Kevin's suggestions. I was also careful not to make formatting changes that would make it difficult to synchronize with other peoples' updates to this code (I don't think I'll be able to use the one that was briefly checked into the MPICH2 tests for this reason). The version that is in our CVS tree has many fixes to the original tests. Bill William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign On Sep 29, 2012, at 1:20 PM, Pavan Balaji wrote:
I don't understand many of the below patches. Why is sync vs. sync/barrier/sync important?
Can you correct the committed test based on this patch? Note that you'll not be able to apply the patch directly as the test has several changes to fit into the MPICH test suite.
-- Pavan
On 09/29/2012 10:56 AM, Rajeev Thakur wrote:
fyi...
Begin forwarded message:
From: Kevin Harms <[email protected]> Date: March 27, 2012 5:09:15 PM CDT To: Rob Latham <[email protected]> Cc: Rajeev Thakur <[email protected]>, William Gropp <[email protected]> Subject: Re: LLNL IO Test?
My comments for testmpio would not be kind. I ran it in early 2009. It flags problems about unexpected error classes which after talking with RobL it wasn't clear that those were defined in the standard or could be implementation specific. There was also a question of the sync versus sync/barrier/sync. There was also a question of checking MPI_Get_count and if the results were defined for certain test cases in the code.
I've attached the original source I used, the modified source I ran and the diff of the two. I ran this at 32 racks of BG/P on GPFS and got it to pass to my satisfaction.
kevin
On Mar 27, 2012, at 4:04 PM, Rob Latham wrote:
On Tue, Mar 27, 2012 at 03:54:50PM -0500, Rajeev Thakur wrote:
Bill, I don't know. I am CCing Rob Latham.
Yes, none of us run testmpio with any regularity.
I am cc'ing Kevin Harms, who bravely squinted at each and every failure during BGP acceptance to determine if it was a defect in the test or defect in the implementation.
Kevin: Bill Gropp wants to compare notes with someone who has run testmpio recently (bgp counts as recent i think).
==rob
Rajeev
On Mar 27, 2012, at 3:48 PM, William Gropp wrote:
Rajeev,
Has anyone been running this test on recent MPICH2? I'm seeing failures both on my Mac using MPICH2, and on the Cray, using the MPICH2 that they ship. In addition to the error class I reported, I'm seeing problems with MPI_File_get_position_shared and MPI_File_get_size. For example, does it work on BG? I'm going to try it on octopus as soon as I have a recent MPICH2 built.
Bill
William Gropp Director, Parallel Computing Institute Deputy Director for Research Institute for Advanced Computing Applications and Technologies Paul and Cynthia Saylor Professor of Computer Science University of Illinois Urbana-Champaign
-- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA
-- Pavan Balaji http://www.mcs.anl.gov/~balaji