opal/cma: improve Linux CMA detection #1306
Conversation
|
:bot:milestone:v2.0.1 This commit enables Linux CMA support by default. This will help shared memory performance. |
|
Test FAILed. |
1 similar comment
|
Test FAILed. |
|
Build Failed with XL compiler! Please review the log, and get in touch if you have questions. Gist: https://gist.github.com/1e504c62707ca698974c73d9950275fd |
|
Build Failed with GNU compiler! Please review the log, and get in touch if you have questions. Gist: https://gist.github.com/8d5398ba793aa2bc4494abc3b6f7da49 |
|
Opps, error in three-way merge. Fixing. |
|
Test FAILed. |
|
Test FAILed. |
This commit improves the CMA detection when the installed glibc doesn't have support for CMA. In this case we need to verify that the syscall numbers in opal/include/opal/sys/cma.h are valid for the architecture. This verification is done by attempting to use CMA while including the internal header. This was back-ported to v2.0.x from master and does not include support for ARMv8. Signed-off-by: Nathan Hjelm <hjelmn@me.com> (cherry picked from commit open-mpi/ompi@4a2bd83) Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
|
Threw away the merged version and just updated to master - the summary stuff. |
|
Test PASSed. |
|
looks good. @jsquyres you may want to double check the configury change. |
|
postponing this to 2.1 |
|
|
||
| # If the user specifically requests CMA go ahead and enable it even | ||
| # if the glibc version does not support process_vm_readv | ||
| if test "x$with_cma" = "xyes" || test "$ompi_check_cma_happy" = "yes" ; then |
There was a problem hiding this comment.
I don't quite grok this: why would we enable CMA if $ompi_check_cma_happy isn't yes?
There was a problem hiding this comment.
Probably can be improved but it doesn't hurt. This only happens if the user tries to force CMA using --with-cma=yes. The user will get a compile-time error if they try to force CMA to be enabled and either 1) we do not have the syscall numbers, or 2) the syscall numbers are wrong.
There was a problem hiding this comment.
That's not really the way we do this in the rest of OMPI. If the user specifies --with-FOO[=yes] and we can't compile/link with FOO support, then configure should fail (vs. failing much later in the make).
|
I realize this is already on master, but it might still be worth strengthening the master version, too. |
|
Please see open-mpi/ompi#1997. |
|
@hjelmn I have merged open-mpi/ompi#1997 to master -- do you want to pull that commit in here? |
|
The |
|
This repo is now closed. |
This commit improves the CMA detection when the installed glibc doesn't
have support for CMA. In this case we need to verify that the syscall
numbers in opal/include/opal/sys/cma.h are valid for the architecture.
This verification is done by attempting to use CMA while including the
internal header.
This was back-ported to v2.0.x from master and does not include
support for ARMv8.
Signed-off-by: Nathan Hjelm hjelmn@me.com
(cherry picked from commit open-mpi/ompi@4a2bd83)
Signed-off-by: Nathan Hjelm hjelmn@lanl.gov