Re: TCP RDMA option to accelerate NFS, CIFS, SCSI, etc.

[Zachary Amsden <zamsden@cthulhu.engr.sgi.com>]

> I must have missed something.
> 
> If we don't have this, you can take the destination port, convert to a
> table address, use the sequence number,
> do some calculations and come up with a buffer address and an offset. If
> you want to mess up the layering
> of your stack, they are all things you can do now.

Yes, you can do that, and it isn't that tricky for simple protocols.  For any 
bulk data protocol higher than TCP with fixed headers, determining the payload 
offset is pretty straigtforward for a single transfer per connection.

Once you start with multiple transfers per connection, variable headers, or 
many different protocols, it becomes harder to do all this work in silicon.  
The advantage I see in RDMA is giving a generic payload pointer for the NIC to 
separate protocol data and payload.

I wonder whether it would be better to do this at the IP layer to enable it to 
be used for UDP and other protocols as well.

Of course RDMA is only going to help in cases where you have receive bandwidth 
issues, and such a scenario isn't likely to be the case for web/file servers 
or desktop clients.

The scenario that it gives the most benefit for is a "middleman" server that 
needs to do lots of I/O to a network storage device while servicing requests.  
In this case, any network I/O running locally may well be over a different 
protocol like UDP or ST.  (ST however, has no need for RDMA acceleration, as 
it already has a buffer transfer design).

-- 
Zachary Amsden  zamsden@engr.sgi.com  (650) 933-6919  09U-510  Core Protocols