2015 OpenFabrics Software User Group Workshop
I will take part at the 2015 OpenFabrics Software User Group Workshop,
it will be held March 19-20 on Monterey, CA.
I'll talk there for a whole 30 minutes about RDMA technologies and about blog.
Here is the URL to the presentation that I had in that workshop:
RDMAmojo: Insights and Thoughts of a Blogger
Dotan
Comments
Tell us what do you think.
Great to hear that.
Please share your talk weblink on the blog.
Thanks
Santosh
Thanks
:)
I'll try...
Dotan
Great job at the conference! Love the blog.
Thanks
:)
Hi Dotan,
Could you share the RDMA session web link which you presented in OpenFabric Group.
Thanks
Santosh
Hi.
I added to the post a URL to the presentation that I had.
Thanks
Dotan
Hi Dotan,
Could you introduce some other opensource projects using libibverbs, such as accelio or gluster.
Thanks
Felix
Hi Felix.
Yes, this is planned - sometime in the future:
I'm maintaining RDMAmojo in my free time, and there isn't much from this resource...
Thanks
Dotan
Hi Dotan,
We plan to develop a RDMA middleware which has more friendly interface and better performance than accelio. Accelio has multiple-level concepts such as session -> connection -> nexus , I think it's too complex. What's your opinion?
Thanks
Felix
Hi Felix.
This isn't a trivial question;
what do you want to achieve?
which API do you want to expose?
What is the reason that rdmacm/rsocket interfaces aren't good?
There are abstraction over libibverbs: Accelio/OpenShmem.
In general, I don't find libibverbs API too complex; but this is only my opinion.
Thanks
Dotan
Hi Dotan,
Thanks for your opinion. I find that most projects's implementation of rdma communication use send/receive mode, I don't know the exact reason. If we implement a abstraction over libibverbs which use RDMA READ/WRITE mode, do you think it will get more benefit such as performance or others?
Thanks
Felix
Hi.
Most project's implementation are using send/receive mode in order to expose an API similar to sockets.
But the rsocket API exposes rread and rwrite as well.
RDMA Write and Read will have better performance then Send/Receive
(since remote side doesn't post Receive Requests nor poll for the Work Completion).
There is a lot of management for the buffers in RDMA Read/Write.
Before starting a new project, which I don't know who will use - I suggest checking the current solution and maybe improve/expand them to support RDMA Write/Read.
But if you want to write a new library for your fun - this is good too
:)
My 2 cents ..
Thanks
Dotan
Thank you very much for the detailed explanation , it took you much time and I learned a lot.
By the way, I have connected you in LinkedIn , that's me :)
:)