Re: MPA startup sequence issues (1) and (2)
"Caitlin Bestler" <[email protected]>
| Newsgroups | gmane.ietf.rddp |
|---|---|
| Message-ID | <[email protected]> |
[email protected] said: > This is the first of several emails whose goal is to achieve > mailing list consensus on issues discussed in Seoul. This > email covers issues (1) and (2), which concern permissible > startup sequences for MPA. The Seoul minutes say: > > (1) Fast FPDU return (responder need not wait for first pdu). > Tentative proposal: leave as-is (responder must wait). > No dissension in room. This is not needed for client/server, > as the server expects to wait for the client. Need to take > proposal to the list. The additional discussion that clarified the problem being finding DDP Segments should probably be incorprated. The justification giving initially was incorrect because it assume the active side had not chosen an endpoint, when in fact it would typically have done so. The problem is not in processing these DDP Segments, its in finding them. In any event, the class of application that I was concerned with are "quick fetchers", such as a possible RDMA-HTTP server. It has since occurred to me that the ULP for such servers can solve the problem using the current specification by simply having the side that is passive for TCP be the initiator for MPA. This class of application will not typically need the Private Data to select an Endpoint (care in selecting an Endpoint is a characteristic of a connection that will last a long time, not a get-in/get-out service). > > (2) Active/Active startup. Tentative proposal: leave as-is (no > support for active/active). No dissension in room. Note, no > known applications require it. Need to take proposal to the list. > I'll make the change in the SCTP draft sometime in the next week or so. I remain ambivalent on the issue. Technically it is not a justifiable restriction. On the other hand it is not a needed capability. The consensus seems to favor reducing the test cases, which strikes me as a valid rationale. In any event, when the one application that actually comes along that needs it, we can at least say that we gave them a chance to object.