[GH-ISSUE #106] Attempting to deploy on AWS #62

Open
opened 2026-10-26 09:44:23 +00:00 by bsakaguchi · 4 comments
Owner

Originally created by @rvencu on GitHub (Mar 5, 2025).
Original GitHub issue: https://github.com/deepseek-ai/3FS/issues/106

AWS has its own optimized networking with EFA and storage optimized instances usually comes with arm CPUs.

I attempted to compile 3FS on this instance, it went quite far, up to this error:

[ 26%] Completed 'apache-arrow-cpp'
[ 26%] Built target apache-arrow-cpp
gmake: *** [Makefile:146: all] Error 2

It would be great to check if supporting AWS would be easy to obtain.

Originally created by @rvencu on GitHub (Mar 5, 2025). Original GitHub issue: https://github.com/deepseek-ai/3FS/issues/106 AWS has its own optimized networking with EFA and storage optimized instances usually comes with arm CPUs. I attempted to compile 3FS on this instance, it went quite far, up to this error: ``` [ 26%] Completed 'apache-arrow-cpp' [ 26%] Built target apache-arrow-cpp gmake: *** [Makefile:146: all] Error 2 ``` It would be great to check if supporting AWS would be easy to obtain.
Author
Owner

@rvencu commented on GitHub (Mar 5, 2025):

there are amd64 instances too, however same EFA

<!-- gh-comment-id:2701329138 --> @rvencu commented on GitHub (Mar 5, 2025): there are amd64 instances too, however same EFA
Author
Owner

@Riatre commented on GitHub (Mar 5, 2025):

See https://github.com/deepseek-ai/3FS/issues/54 for issues on building Arrow.

It is unlikely to work on AWS though, 3FS wants to make RC QPs via libibverbs, which is not something EFA supports.

<!-- gh-comment-id:2701615334 --> @Riatre commented on GitHub (Mar 5, 2025): See https://github.com/deepseek-ai/3FS/issues/54 for issues on building Arrow. It is unlikely to work on AWS though, 3FS wants to make [RC QPs](https://github.com/deepseek-ai/3FS/blob/49e0ad50ae541763323a88aefb71891ffcf04105/src/common/net/ib/IBConnect.cc#L593) via libibverbs, which is not something EFA supports.
Author
Owner

@zartbot commented on GitHub (Mar 6, 2025):

AWS EFA does not support standard RDMA Reliable Connection(RC) QPs, it may takes some effort change the comm module.
Meanwhile, Allibaba cloud is the only Cloud provider which provide standard RDMA RC verbs on all 8th generation generic computing instance, available on all region and all AZ, @rvencu you can have a try.

<!-- gh-comment-id:2702473044 --> @zartbot commented on GitHub (Mar 6, 2025): AWS EFA does not support standard RDMA Reliable Connection(RC) QPs, it may takes some effort change the comm module. Meanwhile, Allibaba cloud is the only Cloud provider which provide standard RDMA RC verbs on all 8th generation generic computing instance, available on all region and all AZ, @rvencu you can have a try.
Author
Owner

@vyalamar commented on GitHub (Feb 1, 2026):

ucx 1.20.x seems to have support for EFA . I tried for a different project and it worked.

<!-- gh-comment-id:3832133425 --> @vyalamar commented on GitHub (Feb 1, 2026): ucx 1.20.x seems to have support for EFA . I tried for a different project and it worked.
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/3FS#62
No description provided.