Nadia Derbey wrote:
On Wed, 2009-05-27 at 14:25 -0400, Jeff Squyres wrote:
Excellent points; Ralph and I chatted about this on the phone today --
we concur with George.
Bull -- would peruse work for you? I think you mentioned before that
it didn't seem attractive to you.
Well, it didn't because from what I understood, the MPI program need to
be changed (register a callback routine for the event, activate the
event, etc), and this is something we wanted to avoid.
Now, if we are allowed to
1. define new "internal" PERUSE events,
2. internally set the associated callback routines
why not using peruse? This combined with the orte notifier framework,
could do the job I think.
FWIW, I did a prototype of some dtrace probes piggybacking on the PERUSE
macros and letting those changes be enabled/disabled at configure
time. One word of caution, if you start adding if statements to all
the PERUSE macros you will more than likely end up significantly slowing
down the performance. So be careful as keep an eye on the overhead as
you add stuff to the macros.
--td
Regards,
Nadia
I think George's point is that we
already have lots of hooks in place in the PML -- and they're called
peruse. So if we could use those hooks, then a) they're run-time
selectable already, and b) there's no additional cost in performance
critical/not-critical code paths (for the case where these stats are
not being collected) because PERUSE has been in the code base for a
long time.
I think the idea is that your callbacks could be invoked by the peruse
hooks and then they can do whatever they want -- increment counters,
conditionally invoke the ORTE notifier system, etc.
On May 27, 2009, at 11:34 AM, George Bosilca wrote:
What is a generic threshold? And what is a counter? We have a policy
against such coding standards, and to be honest I would like to stick
to it. The reason is that the PML is a very complex piece of code, and
I would like to keep it as easy to understand as possible. If people
start adding #if/#endif all over the code, we diverging from this
goal.
The only way to make this work is to call the notifier or some other
framework in this "slow path" and let this other framework do it's own
logic to determine what and when to print. Of course the cost of this
is a function call plus an atomic operation (which is already not
cheap). It's starting to get expensive, even for a "slow path", which
in this particular context is just one insertion in an atomic FIFO.
If instead of counting in number of times we try to send the fragment,
and switch to a time base approach, this can be solved with the PERUSE
calls. There is a callback when the request is created, and another
callback when the first fragment is pushed successfully into the
network. Computing the time between these two, allow a tool to figure
out how much time the request was waiting in some internal queues, and
therefore how much delay this added to the execution time.
george.
On May 27, 2009, at 06:59 , Ralph Castain wrote:
ORTE_NOTIFIER_VERBOSE(api, counter, threshold,...)
#if WANT_NOTIFIER_VERBOSE
opal_atomic_increment(counter);
if (counter > threshold) {
orte_notifier.api(.....)
}
#endif
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
