[ofa-general] [infiniband-diags] add --loop_ports option to perfquery
Al Chu
chu11 at llnl.gov
Wed Oct 8 15:45:18 PDT 2008
Hey Hal,
On Wed, 2008-10-08 at 09:44 -0400, Hal Rosenstock wrote:
> Hi Al,
>
> On Wed, Oct 8, 2008 at 6:26 AM, Al Chu <chu11 at llnl.gov> wrote:
> > Hey Hal,
> >
> > On Wed, 2008-10-08 at 07:03 -0400, Hal Rosenstock wrote:
> >> Al,
> >>
> >> On Tue, Oct 7, 2008 at 6:54 PM, Al Chu <chu11 at llnl.gov> wrote:
> >> > Hey Sasha,
> >> >
> >> > We have a switch here that does not report the AllPortSelect flag as a
> >> > capability. It's pretty annoying typing each port on the switch or
> >> > always having to script around this one oddball switch we have. So I
> >> > added an option --loop_ports for perfquery. If you want to do something
> >> > to all the ports on the CA/Switch, but AllPortSelect isn't available, it
> >> > loops through all the available ports instead.
> >>
> >> Why not add simulated AllPortSelect for multiple ports rather than add
> >> another perquery option for this ?
> >
> > I did try that, and it did seem to work for the switches we had. But
> > when I read the IB spec, it said something to the affect that if a
> > system doesn't support AllPortSelect, setting the PortSelect field to
> > 0xFF was undefined behavior.
>
> I was suggesting that the emulation support (when AllPortSelect is not
> supported) be enhanced for multiple ports and work on both CAs and all
> switches. The one difference is one response for AllPortSelect
> (whether emulated or not) v. many responses for port loop.
Oh. I thought you were referring the the workaround "simulation" that
was in the original code. But you're referring to aggregating the
data/output make it look like AllPortSelect was supported. I'll put
this on the TODO.
>
> >> > There was already a workaround in the tool for a CA that did not support
> >> > the AllPortSelect flag. I get the feeling the workaround may have been
> >> > for a specific hardware, so I kept the workaround in there.
> >>
> >> > Al
> >> >
> >> > --
> >> > Albert Chu
> >> > chu11 at llnl.gov
> >> > Computer Scientist
> >> > High Performance Systems Division
> >> > Lawrence Livermore National Laboratory
> >> >
> >> > _______________________________________________
> >> > general mailing list
> >> > general at lists.openfabrics.org
> >> > http:// lists.openfabrics.org/cgi-bin/mailman/listinfo/general
> >> >
> >> > To unsubscribe, please visit http:// openib.org/mailman/listinfo/openib-general
> >> >
> >>
> >> There are also 2 for loops which are not correct for some switches:
> >> for (i = 1; i <= num_ports; i++)
> >
> > I guess I've never seen a switch that doesn't go from 1 to num_ports.
> > Is there something else I need to handle?
>
> Yes, per the spec, enhanced SP0 supports PortCounters. All your
> switches likely support AllPortSelect so it's not an issue there.
Ok I see now. Wasn't aware of it. I'll get a patch together.
Thanks,
Al
> -- Hal
>
> > Al
> >
> >> -- Hal
> >>
> > --
> > Albert Chu
> > chu11 at llnl.gov
> > Computer Scientist
> > High Performance Systems Division
> > Lawrence Livermore National Laboratory
> >
> >
>
--
Albert Chu
chu11 at llnl.gov
Computer Scientist
High Performance Systems Division
Lawrence Livermore National Laboratory
More information about the general
mailing list