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

Bringing UCX down to commodity hardware - support socket-based client/server connection #3663

Open
mrocklin opened this issue Jun 4, 2019 · 4 comments

Comments

@mrocklin
Copy link

commented Jun 4, 2019

I've recently been using UCX along with Dask, a Python library for parallel computing, to speed up communications on workstations with many GPUs with NVLink. UCX has been great. Thank you all for such a useful library.

I would now like to explore using UCX on a personal laptop without GPUs for TCP and shared memory communication. I'm curious about issues that might arise when using UCX in a more commodity system. A few things come to mind.

  1. I won't have InfiniBand or any RDMA libraries. I understand that currently this can be troublesome because UCX may be using RMDA libraries for addressing.
  2. I would like to use this both on Linux machines, and also OS-X machines. I suspect that UCX is less frequently compiled on OSX and was curiuos if I should expect any issues here.
  3. I will likely want to use shared memory transports on OSX, and I suspect that this has not been written.
  4. Eventually, I can imagine Windows users also being interested in this. Is this possible with modest effort?

I thought I would ask both about the particular issues above, and also about any other issues that maintainers here might anticipate.

Thank you all for reading and for your efforts on UCX. Also, @Akshay-Venkatesh originally recommended that I put this on the mailing list. I would be happy to do so if that's the right place for this kind of conversation. I just default to GitHub first.

@mike-dubman

This comment has been minimized.

Copy link
Contributor

commented Jun 4, 2019

  • UCX does not require rdma stack and will work on laptop with tcp and shared memory. (you can use it in with win10 WSL as well)
  • the OSX support is ongoing with big help of @hiroyuki-sato, you are welcome to contribute as well.
    @yosefe
@hiroyuki-sato

This comment has been minimized.

Copy link
Contributor

commented Jun 4, 2019

Hello, @mrocklin

I and @keisukefukuda working OpenUCX porting to macOS.
It is still in progress. It still exists many steps yet.
I'm using this branch for changing for macOS.
https://github.com/hiroyuki-sato/ucx/tree/osx_fixes

All of the issues in this URL is macOS specific.
https://github.com/hiroyuki-sato/ucx/issues

This is the summary page.
hiroyuki-sato#1

Thanks.

@Akshay-Venkatesh

This comment has been minimized.

Copy link
Contributor

commented Jun 4, 2019

@miked-mellanox Isn't rdmacm required for client server connection establishment in UCX?

@yosefe

This comment has been minimized.

Copy link
Contributor

commented Jun 4, 2019

@Akshay-Venkatesh yes, your sockcm PR is required to support client/server mode without rdmacm

@yosefe yosefe added the Question label Jun 4, 2019

@yosefe yosefe changed the title Bringing UCX down to commodity hardware Bringing UCX down to commodity hardware - support socket-based client/server connection Aug 8, 2019

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
5 participants
You can’t perform that action at this time.