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

[julian_satran@il.ibm.com]

Message boundaries are only part of the proposal - and they don't imply
additional work at the client. Doing zero copy based only on the
information in current headers is certainly possible at low speed. Over 1
Gb/s it requires some innovation and lots of silicon. The RDMA option makes
it possible at a far lower price. And the zero copy it enables might go
deep into the application space as it is only an annotation on packets.
It certainly makes sense on all the new applications (NFS4, SCSI, etc) and
the retrofit into existing ones is not that difficult either.
And placing it over TCP puts it on a safer ground that having to use at
higher speed completely new and unproven protocols (VIA, NGIO etc.).

Julo
Julian Satran - IBM Research

David Robinson <David.Robinson@EBay.Sun.COM> on 25/02/2000 03:42:05

Please respond to David Robinson <David.Robinson@EBay.Sun.COM>

To:   ips@ece.cmu.edu
cc:   tcp-impl@grc.nasa.gov (bcc: Julian Satran/Haifa/IBM)
Subject:  Re: TCP RDMA option to accelerate NFS, CIFS, SCSI, etc.




To efficently determine boundaries within a packet stream work
must be done somewhere.  In the RDMA proposal it is up to the
clients to do the work to make the server's job easier.  In traditional
intelligent NIC cards the server does the work by parsing the headers.

It seems that the design of RDMA is backwards as it relies on changes
to the many clients to enable efficiency on the server. A traditional
intelligent NIC card with a modest amount of hardware/firmware
can handle 99+% of requests from unmodified clients.  The existence
proof is checksumming NICs and NFS accelerator boards.

For an efficient IP storage device it will have to deal with legacy IP
client stacks (no RDMA) and a competitive IP storage vendor will
implement the smart NIC described above. Why is RDMA more compelling?

     -David