I find some of these arguments somewhat strange. We have all entered into this effort to bring optimized processing with Remote Direct Data Placement, yet we continue to talk about, and give serious weight to, the argument concerning an all software implementation. It seems that we might think about a complete software implementation, in passing, but this is NOT the goal of RDMA or this RDDP workgroup (in my opinion). Therefore, I do not think it is important to do things that will make the protocol more complicated for this small segment of the market, assuming that it exist at all other then for a testing platform.
Most of the applications that will be using the RDMA protocols already have existing TCP/IP implementations, and they are not enhanced or accelerated by a software implementation of RDMA. If systems only have a standard NIC, then the applications should continue to use TCP/IP.
Expanding on this point further, the work being done in various places on Socket Direct Protocols, will permit the application to continue to use the Socket APIs, and the OS will convert that to RDMA where possible and reasonable. I submit that will be when there is an RDMA chip set available.
There are other applications that I am aware of that are "adding" the RDMA capability, not taking away the TCP/IP implementation. They check to see which one they can use and act according.
So again, being overly concerned about a software implementation of RDMA is not something on which we should spend much time.
John L. Hufferd
Senior Technical Staff Member (STSM)
IBM/System Group, San Jose CA
Main Office: (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688
Home Office: (408) 997-6136, Cell: (408) 499-9702
Internet Address: hufferd <at> us.ibm.com
||"Williams, Jim" <Jim.Williams <at> emulex.com>
Sent by: rddp-admin <at> ietf.org
04/08/2003 05:17 PM
To: "'rddp <at> ietf.org'" <rddp <at> ietf.org>
Subject: RE: [rddp] Option or mandatory CRC in MPA
The argument for making the CRC optional
in the RDDP over TCP mapping is the following.
A. A large percentage of usage will be on
highly reliable private networks where
the TCP checksum is more that sufficient
the achieve the desired level or reliability.
B. Optimizing the performance of software
implementations is important, and computing
an unneeded CRC is too high a price to pay
for avoiding the need to negotiate an option.
The input from NFS is important because it takes
into account years of field experience. If they
determine that for most all usage, even on private
networks, the TCP checksum is NOT adequate, then
I would regard that as a refutation of point A.
If they determine that for a large enough percentage
of implementations the TCP checksum IS adequate,
then I would regard that as support of point A.
Regarding point B, if one envisions that all
implementations of interest are hardware customized
for RDDP, then the point does not carry much weight.
However, I would argue that this may be too narrow
a view for the IETF to take.
rddp mailing list
rddp <at> ietf.org