On Dec 17, 2006, at 2:36 PM, Scott Atchley wrote:
Hi all,
I am ready to start testing MX support.
What do I need to do to include bmi_mx in the configure and make
process? Would it be easier to work with a cvs checkout or a
release tarball?
As for testing, I do not have disks fast enough to stress
Myrinet-2000 (2Gb/s) NICs let alone Myri-10G NICs. In other emails,
these options have been mentioned. Which makes most sense (least
amount of coding, compiling issues, etc.)? Do they require disk IO?
You shouldnt need disks IO to test whether or not your port works.
Most of the critical problems that you will run into will be on the
BMI layer, which doesnt really manage the disks. The only thing
we've ever 'needed' disks to test for in the past is fine tuning for
these new 10Gb+ hardwares. If someone wants to send me the MX
hardware to test on fast disks I'm sure I could convince someone here
to let us test it, but I dont think that'll be an option. Good luck.
Netpipe proved to be very useful in testing/reproducing some of our
problems with the IB code.
I can get you an early release once you get the MX stuff to a point
where you can do pvfs2-cp on it if you're interested.
+=Kyle
We have a flow method called 'flowproto-dump-offsets', which will
dump
offset-length pairs. the problem is that the method isn't used very
much, and it has not kept up with some of the API changes and no
longer compiles. If that sort of thing sounds useful to you, we
might
be able to get it back into shape.
There are some BMI tests in the test/io/bmi directory which will
stress just the BMI layer. Those tests might also need some manual
coaxing to exercise bmi_mx (the gm versions hard-code 'bmi_gm' into
the call to BMI_Initialize, for example).
There's also Julian's Trove 'TAS' implementation, which he's got
sitting in a branch ready to be merged. I had him update it so it
works with the new trove-method stuff I committed, so maybe it
makes sense to merge that to trunk before a next release? Its
probably going to be easier to do that than try to figure out
exactly why dump-offsets isn't working.
If anyone wants to look at a tarball of what I have so far, let me
know. It is only about 2300 LOC. I still need to add more error
checking, convert to PVFS-style error reporting, and cleanup of
peer state when a peer has restarted.
Scott
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
!DSPAM:4585ab91217681542719951!
_______________________________________________
Pvfs2-developers mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
