Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Issues regarding ucc_coll_args::tag #893

Open
nirandaperera opened this issue Dec 19, 2023 · 6 comments
Open

Issues regarding ucc_coll_args::tag #893

nirandaperera opened this issue Dec 19, 2023 · 6 comments

Comments

@nirandaperera
Copy link

nirandaperera commented Dec 19, 2023

The docs mention that a collective tag is a 64-bit integer.

to provide the “tag”, which is an unsigned 64-bit integer.

Butucc_coll_args.tag is a typedef uint16_t ucc_coll_id_t type. Do you know if this is a bug or intentional?

Also, should the tag be unique to the collective operation individually, or should it be unique to all collective operations globally?

@Sergei-Lebedev
Copy link
Contributor

Hi @nirandaperera, I think it's bug in documentation. Currently "tag" exists only in API, no TLs in UCC support it and it would be difficult to implement 64 bit tags. @manjugv can you please take a look.

@nirandaperera
Copy link
Author

@Sergei-Lebedev so does it mean that none of the tls currently support unordered collectives?

@Sergei-Lebedev
Copy link
Contributor

@Sergei-Lebedev so does it mean that none of the tls currently support unordered collectives?

yes, it was added to API for future use.

@nirandaperera
Copy link
Author

It seems like the tag value is used in the internal impl.
https://github.com/openucx/ucc/blob/master/src/components/tl/ucp/tl_ucp_coll.h#L339-L340

It seems to be copied to ucc_tl_ucp_task.tagged.tag value. And in the UCP impl, that seems to be used in creating the ucp tag. So, it seems to me that we can still use tagged/ unordered collectives, isn't it?

@nirandaperera
Copy link
Author

https://github.com/openucx/ucc/blob/master/src/components/tl/ucp/tl_ucp_tag.h#L16
This tells that a ucp message id is 15 bits long. So, I guess that uint16 type seems to be correct in the underneath impl

@Sergei-Lebedev
Copy link
Contributor

It seems like the tag value is used in the internal impl. https://github.com/openucx/ucc/blob/master/src/components/tl/ucp/tl_ucp_coll.h#L339-L340

It seems to be copied to ucc_tl_ucp_task.tagged.tag value. And in the UCP impl, that seems to be used in creating the ucp tag. So, it seems to me that we can still use tagged/ unordered collectives, isn't it?

Yes, this message tag is used to correctly schedule send/recv operations along with source rank and team id. TL UCP generates tag when user doesn't provide it. The idea is if user gives us a tag we don't generate but use whatever user gives us. Apparently it should work only with TL UCP.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants