Ethan and I talked about this in Paris. I think it's Whacky. :-) But hey, whacky can be good. :-)
I'm not a big fan of the --run-engine idea; my counter-proposal was to make a --get-value option where you could query the value of any field in any section (subject to full expansion rules). Hence, you could have a shell script that does something like:
mtt --get-value "mpi_details: test-matrix" exec > stuff_to_run Then iterate over the lines in stuff_to_run and execute each of them. On Sep 29, 2007, at 12:04 AM, Ethan Mallove wrote:
Folks, I had this wacky idea the other day. I have a collection of scripts that I've accumulated to run MTT with parameter combinations that MTT does not iterate on (see #245). E.g., I have a dozen or so scripts that do variations on something like this: Do a 32-bit mtt run with compiler X, on 2 nodes Do a 32-bit mtt run with compiler Y, on 2 nodes Do a 64-bit mtt run with compiler X, on 2 nodes Do a 64-bit mtt run with compiler Y, on 2 nodes Do a 32-bit mtt run with compiler X, on 1 node Do a 32-bit mtt run with compiler Y, on 1 node Do a 64-bit mtt run with compiler X, on 1 node Do a 64-bit mtt run with compiler Y, on 1 node ... So I thought, "can I somehow harness xargs to do this instead of creating another shell script every time I have another testing combination I want to run?" xargs doesn't seem powerful enough. Then I realized that the MTT run engine *itself* does exactly what I want. Xargs on steroids. So I created the following "MPI" details section to get MTT to run itself: [MPI Details: test-matrix] exec = <<EOT client/mtt --debug --trial --file @INI_NAME@--section "mpi.get.*ompi-nightly-trunk mtt mpi.details.*open.mpi reporter.*iu.database mpi.install.*autotools ..." --scratch /opt/mtt/@bitness@/@compiler@/&mpi_details_name()/ my-scratch-areaprepend_path="PATH /path/to/my/compiler/@compiler@" bitness=@bitness@ run_on_single_node=@run_on_single_node@ EOT bitness = &enumerate(32,64) run_on_single_node = &enumerate(0, 1) compiler = &enumerate("A", "B") The above expands to 8 MTT runs (2 compilers x 2 bitnesses x 2 hostlists). Then, I had to fake MTT out in order to *use* the run engine. So I created some Noop.pm modules that do nothing but scoot MTT along to the RunEngine phase: [MPI get: noop] # Fill this in from the command line mpi_details = module = Noop [MPI install: noop] mpi_get = noop module = Noop [Test get: noop] module = Noop [Test build: noop] test_get = noop module = Noop [Test run: noop] test_build = noop module = Noop timeout = -1 Finally, I can kick off a series of MTT runs like this: $ client/mtt --section "noop test-matrix" mpi_details=test-matrix What do you think? I'd like to make this use case even easier. E.g., create a --run-engine (?) option for the specific purpose of skipping all the phases in order to simply execute commands. Or maybe there is another tool to do what I am trying to do? In which case, this has been an exercise :-) -Ethan _______________________________________________ mtt-devel mailing list mtt-de...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/mtt-devel
-- Jeff Squyres Cisco Systems
