11/20/02 RDDP WG David Black, Chair Notes by Allyn Romanow -- Agenda bashing -- Work and goals draft-black-rdma-concerns-00.txt will be submitted as an RDDP draft to make it easier to find it, but the authors are not planning to submit for RFC status. It's purpose it to cause people to consider issues. -- WORKSHOP announcement (Allyn Romanow) Network-I/O Convergence: Experience, Lessons, Implications (NICELI) - SIGCOMM 2003 Workshop August 25-29, Karlsruhe Germany - Call for Papers - technical paper or position paper - Will be at www.acm.org/sigcomm/sigcomm2003/workshop/niceli - Or contact allyn@cisco.com -- Review of Problem Statement and Architecture drafts (Tom Talpey) draft-romanow-rdma-over-ip-problem-statement-01.txt draft-bailey-roi-ddp-rdma-arch-01.txt Two new versions Problem statement done Architecture issues ordering of placement and delivery steering tag details security considerations and threat analysis See David Black if you want to work on security section of the doc These two drafts will become RDDP WG drafts. ------ -- SCTP Mapping draft-stewart-rddp-sctp-00.txt (Randall Stewart) Short doc, provides SCTP mechanism for knowing what each side expects to do Details of ways to use SCTP features not filled in yet how use SCTP most effectively how optimize transfer with un-ordered packets other details Soliciting help, work with Randy. Get in touch with David Black or Randy ----------- -- Overview of DDP and RDMA (Renato Recio) draft-shah-iwarp-ddp-00.txt draft-recio-iwarp-rdma-00.txt See overview slides - these explain high level functionality and relationship of these two drafts. ------------------ -- DDP draft (Hemal Shah) draft-shah-iwarp-ddp-00.txt See slides for details. Overview of the DDP mechanism Distinction between placement and delivery is extremely important Placement - self describing segments, can place in any order Delivery - in-order delivery, DDP notifies ULP only when all messages received Only interested in LLP that provides reliable in-order delivery STag Validation semantics - 4 models After discussion, the rough consensus of the room was to carry two of the four models, STag bound to a single stream and access group (protection domain model forward. A functional reference model will be needed to specify the required access control behavior; implementations must exhibit the externally visible behavior of the model (this is similar to the specification of the IPsec Security Policy Database and Security Association Database - see RFC 2401). A performance concern was raised about the consequences of access control. Combining the above functional model with a functional model of a NIC (hardware) should help illuminate the issues here. Clarifications on DDP Segmentation - Send segments in increasing order, for both untagged and tagged transfers - No overlapping of payload is allowed among DDP segments in the same message. - Interleaving DDP segments of different DDP messages not allowed at the data source. Use multiple streams for concurrency, not multiplex on the same stream Requirements for Unreliable Transport Only considering reliable transports currently. Anyone interested in unreliable transports should "Send Draft" .. says WG chair. Local interface requirement for buffers Need to specify? yes, need to specify access control and protection, but not from the perspective of specifying a full API Reserved ULP field in DDP Header Remote invalidate will need to be taken up, can be dealt with online Q&A: DDP should cleanly layer on unordered SCTP. Should layer on ordered SCTP with a little care and attention. SCTP can support immediate placement and in-order delivery, provide that access to cumulative TSN is available to DDP. This will become and official WG draft (next version containing access control text) ------------------- -- RDMAP (Renato Recio) draft-recio-iwarp-rdma-00.txt There have been 3 issues on mailing list (1) RDMAP Initialization - How to transition from non-RDMA to RDMA TCP mapping needs to look at SCTP, understand what it did (DDP and RDMA support announced during connection setup) and either do something similar, or explain why not. (2) Security Have begun working with Catherine Meadows, security expert, on the Security Directorate, lots more work needed on security. Security Concerns -STag -what layer is the association made at? -how can IPsec be used to protect an DDP RDMA stream? -Enumerate the threat model (e.g. DOS attack ) (3) Multi-homed hosts for SCTP Proposal - this issue doesn't belong here, it belongs in SCTP WG (TSVWG)-a problem with multi-homed hosts. WG Chair - SCTP mapping draft should deal with this and say how it should be handled or avoided. This will become an official WG draft (next version including access control text) ------------------ David Black (WG Chair)- What might be the draft mapping RDMA to TCP look like? Reviews what happened in TSVWG meeting the previous day. It will take a while to work out any proposed changes to TCP. It is premature to adopt the Williams draft as WG doc, although the WG chair thinks it has good stuff in it, and the draft suggests how to do the mapping without changes to TCP Scott Bradner (Area Director) - reports on a discussion that took place after the TSVWG meeting. ADs (tsvwg chairs) are very reluctant to approve changes to TCP. It would be necessary to put together an I-D that proves that no possible combination of tweaked and non-tweaked TCPs could possibly get confused, meaning that it is necessary to prove that there is no way that different TCPS can get confused between each other. Agrees that it is premature to accept the Williams draft as a WG item. Q: Can we use the rddp WG as where the discussion takes place? Bradner - the initial discussion was in TSVWG for the broader community. Detailed work can be done here, in RDDP, then go to TSVWG, for review --------------------- -- Mapping DDP/RDMA to TCP, IFT (Jim Williams) draft-williams-iwarp-ift-00.txt See slides Open issues CRC - slide proposed two CRCs if assume non-aligned PDUs as in IFT/TCP One CRC should be sufficient for assume aligned PDUs as in SCTP or MPA/TCP - A long inconclusive discussion ensued - the issue was redirected to the mailing list. Issue - markers Can't really discuss till figured out in TSVWG WG Chair (David Black): MPA uses an aggressive form of markers. Comment: Reason for markers is that they allow the ULP to progress to find the next PDU, after an error is found as opposed to just dropping the connection. Other IFT details - not worth dealing with now. Q: Does TCP need to be "FIXED"? does TCP placement as in SCTP cover the needs of RDDP? David Black - MPA and IFT appear to be at opposite ends of the design space, thinks there might be a design strategy in between. Will need to follow Scott's suggestion for any TCP changes. Comment: Wants support for bufferless NICS as well as buffered NICs. David Black - there's no such thing as a bufferless NIC, as dealing with a DMA controller requires buffering - should talk in terms of the amount of buffering. -- Process going forward (David Black) Take discussion to the list Focus on specific features rather than one approach or another Drafts are extremes in the design space, there should be a way in between Too early and too much unknown to charter a design team now Question: Should RDDP write a document on what a ULP needs to do to use DDP/RDMA? A reference model was mentioned- will some of this come out of a reference model? David Black - ref model in two places, NIC functionality and security. - Another issue is the memory model. Joe Touch in TSVWG, asked, who owns what memory, when? It needs a state machine. A related issue is what happens when there are concurrent operations on different streams to the same memory? That needs to be specified minimally in order to avoid running afoul of multiprocessor memory coherence issues. Application designer wants description of what NOT to do in both cases.