Fix typos in OSC RDMA BTL allowlist #7823
Conversation
|
@bwbarrett @awlauria @rhc54 This is the cause of so many of the Cisco MTT failures over the past month or two. |
|
Either someone is going to put in effort to fix the PT2PT component or fix the RDMA component. Why would we put effort into the PT2PT component? |
|
I believe nobody disputes that statement - the issue is: who is going to put in the effort? Simply removing p2p isn't the answer. To date, nobody has been willing to make the effort. Perhaps providing a clean mechanism by which BTLs can reject OSC operations with an appropriate error would solve it - but somebody would have to make the effort to create that too. 🤷♂️ |
|
I will take a look at btl/tcp. |
|
Even if BTL tcp is fixed, do users running on a single node (e.g., a laptop) have to run Same question for usNIC: if I'm using usNIC (which can't be used for loopback communication), do I have to run |
"openib" no longer exists. "tcp" had a typo. Signed-off-by: Jeff Squyres <[email protected]>
jsquyres
force-pushed
the
pr/put-osc-pt2pt-back
branch
from
fd65517 to
18cfcc8
Compare
|
Per discussion on the weekly OMPI call today, I changed this PR to solely remove |
| OBJ_RELEASE(new_enum); | ||
|
|
||
| ompi_osc_rdma_btl_names = "openib,ugni,uct,ucp"; | ||
| ompi_osc_rdma_btl_names = "ugni,uct,tcp"; |
There was a problem hiding this comment.
Good question. I defer to @hjelmn to answer that...
bwbarrett
left a comment
bwbarrett
left a comment
There was a problem hiding this comment.
forgot to approve earlier...
|
bot:ompi:retest |
1 similar comment
|
bot:ompi:retest |
|
One more time.. |
|
bot:ompi:retest |
5 participants
OSC rdma had a reference to
openib, which no longer exists on master. It also had a typo for thetcpBTL (but even after fixing that typo, OSC rdma does not activate itself when the TCP BTL is used).