RDMA Bridging? ("You don't know that you don't know")

RDMA Bridging? ("You don't know that you don't know")

[Joshua West]

Hey Folks,
 - For background, I am working with a small (home-lab) ceph cluster
on a proxmox cluster.
 - Proxmox uses a shared cluster storage to pass configuration files
around, including ceph.conf
 - All nodes are connected with Mellanox connectx-3 (mlx4_core) 56GbE
cards connected via qsfp switch.
 - I am experimenting with Mellanox connectx-4 (mlx5_core) 100GbE
cards between two nodes, primarily for fun and to learn, as I can't
afford a 100GbE switch for a hobby.
    -- (The two 100GbE cards are direct attached to one another)
 - Lastly, ceph is using the (unsupported?) rdma config
(`ms_async_transport_type = rdma`, `ms_bind_ipv4 = true`,
`ms_cluster_type = async+rdma`)

My question:
As each ceph osd is communicating with each other by specifically
naming the rdma device it should be reached at, for example:
```[osd.0]
ms_async_rdma_device_name = rocep8s0```

I am having trouble rationalizing how to allow the two nodes with
multiple RDMA devices to communicate with one another, without
disrupting the ability for the rest of the cluster to still
communicate with the lower throughput rdam device.

For example:
rdam dev
`0: rocep68s0: node_type ca fw 12.16.1020 node_guid
7cfe:9003:0026:88ac sys_image_guid 7cfe:9003:0026:88ac`
`1: rocep65s0: node_type ca fw 2.42.5000 node_guid 0002:c903:0042:dff0
sys_image_guid 0002:c903:0042:dff3`

If ms_async_rdma_device_name is set to either device, then
communication only partial connectivity can occur.

What is the correct method for "routing" rdma / RoCE ?



Josh

Re: RDMA Bridging? ("You don't know that you don't know")

[Leon Romanovsky]

On Wed, Jan 12, 2022 at 08:32:17AM -0700, Joshua West wrote:
> Hey Folks,

<...>

> What is the correct method for "routing" rdma / RoCE ?

I don't know anything about CEPH, so my answer comes from HW/kernel
perspective.

In RoCE, we use GID table entries to perform routing.

Thanks

> 
> 
> 
> Josh