I'm inclined to say 'let's just do it right the first time' meaning
let's implement the 3D table we designed a few weeks ago. This would
allow for arbitrary storage of performance data. But would take a
considerable amount of effort.
However, it would probably be quicker to implement a NBC performance
table and do the right plumbing to make it work. The problem is that
next time we want to add a new performance data table will have to do
all this all over again. :/
I'll have to see the layout of the data when Torsten gets back to
determine how complex the database table will need to be. I suspect
it won't be too bad, but can't say for sure at the moment.
Unfortunately I'm swamped at the moment, so I probably won't be able
to work on this for a few months out, but I can help with planning
for it.
Cheers,
-- Josh
On Oct 8, 2007, at 12:49 PM, Jeff Squyres wrote:
Ethan/Josh --
Torsten's libnbc benchmark program outputs a whole bunch of numbers
that we are not currently recording. Essentially, it outputs a fixed
number of double precision numbers for each message size (each of
which has a different meaning). We obviously currently cannot record
all of these numbers in the database.
Torsten will come find Josh when he returns to IU to discuss if it
would be easy/hard to store this stuff in the database -- that would
be the first step. Next we will need to update submit.php and the
client to submit the data properly and then store it in the DB.
Finally, we will need to update the reporter to visualize this
information in a new way (it's more than just standard latency/
bandwidth information).
There is no immediate need to get this all done, but it should be
added to the "to do" list to eventually get done. I'll file a ticket
about it.
I think that this is different than the SKaMPI storage problem
because the number of values that are output for each message size is
fixed (the extended SKaMPI output has a number for each MPI
process). So I *think* it should be fairly easy to add a new table
for libnbc data in the database...?
--
Jeff Squyres
Cisco Systems
_______________________________________________
mtt-devel mailing list
mtt-de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/mtt-devel