Bob Woodruff wrote:
Doug Ledford wrote,
The suggestion from Bob Woodruff was to use the verbs-cm support since
kDAPL isn't included in the kernel.
Yes, for now I would suggest to use the socket CM version of uDAPL.
Arlin is starting to debug the CMA version, but for now the socket
CM version of uDAPL seems stable. We ran it on a 128 node cluster in
Dupont for SC'05 with no problems.
For example, to build and install the socket CM
version of uDAPL on a Red Hat EL 4.0 kernel,
from the userspace directory,
cd dapl/dapl/udapl
make VERBS=openib_scm
cp -f ./Target/libdapl.a /usr/local/lib/libdapl-openib.a
cp -f ./Target/libdapl.so /usr/local/lib/libdapl-openib.so
cd ../../dat/udat
make OS_VENDOR=REDHAT_EL4
cp -f ./Target/x86_64/libdat.a /usr/lib64
cp -f ./Target/x86_64/libdat.so /usr/lib64
Not sure why we needed the REDHAT_EL4 option on building libdat.so,
but Arlin Davis can provide details if needed.
I'll keep that in mind, but to be honest, I've had my best luck so far
allowing rpmbuild to do the right thing (using things like the
%configure macro instead of trying to hand configure the installation,
etc). In addition, my kernel's include support that the REDHAT_EL4 flag
may be assuming is missing. My goal is to get everything compiling as
though this were an upstream kernel and up to date user space
components, so I'm hoping the REDHAT_EL4 flag won't be necessary by the
time I'm done. But, user space deadlines are after kernel space
deadlines, so I'm off working on kernel things right now.
--
Doug Ledford <[EMAIL PROTECTED]>
http://people.redhat.com/dledford
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
