A group of us have been working on an RDMA system called WARP. The RDMA layer aims to provide exact data placement for iSCSI, NFS, CIFS, as well as the VI...
6
csapuntz@...
Jan 24, 2001 12:30 am
Mail should be sent to "rdma@..." and not "rdma@..." due to the vaguaries of the Cisco mail system. My apologies for any confusion this...
7
Douglas Otis
dotis@...
Jan 24, 2001 3:41 am
Welcome to the exciting and dynamic world of RDMA. Doug...
8
Douglas Otis
dotis@...
Jan 24, 2001 7:19 pm
http://www.stanford.edu/~csapuntz/draft-jpink-warp-summary-00.txt Use SCTP structure exactly rather than something similar. If there is expectation of RDMA or...
9
David Robinson
David.Robinson@...
Jan 25, 2001 3:49 am
... Point of order. Is it not a stated requirement of WARP to run over TCP? Yes WARP over TCP is harder than WARP over SCTP but you don't get to solve the...
10
Douglas Otis
dotis@...
Jan 25, 2001 6:40 am
David, That was not my point. If you are to use SCTP like structures for TCP, use the exact structure where possible and not a structure similar to SCTP. In ...
11
Constantine Sapuntzakis
csapuntz@...
Jan 25, 2001 4:23 pm
Hi Doug, Thanks for the feedback. What I get from your feedback is the following. SCTP chunks currently have a: - stream identifier - stream sequence...
12
Douglas Otis
dotis@...
Jan 25, 2001 7:24 pm
Costa, If you look to applications interfaces like BEEP or the like, the concept of establishing parallel processing on a single connection is not something ...
13
Douglas Otis
dotis@...
Jan 25, 2001 8:07 pm
Costa, As a follow-on, should you place SCTP headers as fixed intervals within the TCP stream, then you could replace the port designators with a "STCP frame ...
14
Arkady Kanevsky
arkady@...
Feb 13, 2001 2:51 pm
Will there be a discussion meeting on WARP proposal at the next IETF conference in March in Minneapolis? Arkady -- Arkady Kanevsky...
15
Minturn, Dave B
dave.b.minturn@...
Feb 13, 2001 6:12 pm
What are the current thoughts on how VI semantics map to WARP. In reviewing the WARP specification, I noticed that both the RDMA and Send packets do not...
16
Jim Williams
jimw@...
Feb 13, 2001 10:01 pm
... From: "Minturn, Dave B" <dave.b.minturn@...> To: <rdma@...> Sent: Tuesday, February 13, 2001 1:06 PM Subject: Question on VI over WARP...
17
csapuntz@...
Feb 14, 2001 12:39 am
... Arkady, I have requested (as of today) a Birds-of-a-Feather session (BOF) to discuss RDMA in general at the next IETF in Minneapolis. Whether we make it...
18
Y P Cheng
Ycheng@...
Feb 14, 2001 7:14 pm
I am totally confused on Jim's response to Dave's question. May be the difference lies between my familiarity with the FC-VI Rev. 1.6 and Giganet's ...
19
Jim Williams
jimw@...
Feb 14, 2001 8:47 pm
... Let me back up a little, then. WARP provides a mechanism for transfering data over IP in which packets are sufficiently self describing that suitably...
20
Y P Cheng
Ycheng@...
Feb 14, 2001 10:20 pm
... Thank you for a quick clarification. A VI-Send by its own semantics must have waiting anonymous buffers on the other side. It is no different from a ...
21
Minturn, Dave B
dave.b.minturn@...
Feb 14, 2001 11:14 pm
Thanks for the answers (and more questions). The following is my take on what WARP is missing in the specified packet formats to support a simple VI ULP. ...
22
Michael Krause
krause@...
Feb 15, 2001 12:00 am
... I do not understand why there is any requirement for VI API support. The standard is a wire protocol with defined semantics. While VI may be used to...
23
Y P Cheng
ycheng@...
Feb 15, 2001 12:29 am
Dave, Please read the FC-VI by downloading from the T11 site. It would answer most of your questions. In fact, WARP is similar to FC-VI. Reliable VI ...
24
Y P Cheng
ycheng@...
Feb 15, 2001 12:43 am
... While I am not very familiar with the exact packet format of InfiniBand, how do we map VI to Ethernet using WARP and still allowing WARP to support TCP and...
25
Douglas Otis
dotis@...
Feb 15, 2001 4:07 am
Y.P. What are your thoughts in providing a general encapsulation of SCTP within TCP. As example, here is the general SCTP header modified to include a ...
26
Jim Pinkerton
jpink@...
Feb 15, 2001 6:59 am
WARP is substantially simpler than Infiniband. It is specifically focused on the data steering issue to enable NICs to implement zero-copy models in the face...
27
Michael Krause
krause@...
Feb 15, 2001 4:13 pm
... WARP should define an acknowledgement scheme above what TCP provides since the checksum / CRC used within the chunks is independent of the TCP checksum....
28
Michael Krause
krause@...
Feb 15, 2001 4:39 pm
... The people in the InfiniBand link workgroup that I co-chair spent a great deal of time creating these headers and going through all of the RDMA ...
29
Michael Krause
krause@...
Feb 15, 2001 4:43 pm
... While InfiniBand is certainly more complex / robust than WARP, I do not see the need to have different headers / semantics being used for the subset of ...
30
Jim Pinkerton
jpink@...
Feb 15, 2001 6:57 pm
... We had lengthy discussions about this, and felt that creating a reliable protocol on top of a reliable protocol (TCP) went beyond the keep-it-simple-stupid...
31
Y P Cheng
ycheng@...
Feb 15, 2001 7:02 pm
... Doug, Before I reply, I must give you my disclaimer. In my years of designing of hard disk, SCSI, and Fibre Channel adapters, my jobs were always taking a...
32
Michael Krause
krause@...
Feb 15, 2001 7:26 pm
... WARP is in essence a session layer concept similar to what is being pursued for iSCSI in some respects. As such, the value of having it determine ...
33
Minturn, Dave B
dave.b.minturn@...
Feb 15, 2001 7:30 pm
... goals ... The question about the explicit ACK was concerning the effect between the TCP acknowledgement (i.e. the receiving side's delayed ACK timer) and...
34
Douglas Otis
dotis@...
Feb 15, 2001 9:12 pm
Jim, Here is an example of using SCTP within TCP using a fixed interval alignment of a pseudo-frame allowing frame recovery where padding is inserted should ...
