RE: MPA alignment
"John Hufferd" <[email protected]>
| Newsgroups | gmane.ietf.ips |
|---|---|
| Message-ID | <[email protected]> |
Well, the initiator side must still field the RDMA messages sent by the target (RDMA Read & RDMA write). And that means that the RDMA Read response must be handled by initiator side, and of course the RDMA write must be handled and data moved into the appropriate buffers when the RDMA write message is received from the target. The software RDMA stack on the Initiator is really quite busy. . . . John L Hufferd Sr. Executive Director of Technology Brocade Communications Systems, Inc [email protected] Office Phone: (408) 333-5244; eFAX: (408) 904-4688 Alt Office Phone: (408) 997-6136; Cell: (408) 627-9606 _____ From: Dan Bar Dov [mailto:[email protected]] Sent: Sunday, April 23, 2006 1:47 AM To: John Hufferd; Eddy Quicksall; Mike Ko Cc: [email protected] Subject: RE: [Ips] MPA alignment I just wanted to note in addition to what John said, that ISER is not a symmetric protocol. Only the target side performs the RDMA operations. The initiator does only the setup. Dan _____ From: John Hufferd [mailto:[email protected]] Sent: Tuesday, April 18, 2006 9:56 PM To: Eddy Quicksall; Mike Ko Cc: [email protected] Subject: RE: [Ips] MPA alignment Eddy, I agree with your statement about software RDMA being a bit silly, but only when the thought is about implementations on servers. There is at least one type of environment where using RDMA with a software iWARP stack may not be too silly, and that is when the object of the interconnect is to optimize the over all costs and save Server Cycles. In that case is may make sense to operate with a software RDMA stack on the desktop systems or workstations. Those environments often have a plentiful amount of available CPU cycles on the desktop/workstation. So, the approach is to burn CPU cycles at the desktop/workstation and save CPU cycles on the servers, which have RNICs. This type of a configuration actually seems to be a useful configuration. The same argument can be used with iSCSI and iSER, but only in the case where there are RNICs in the target. It is a little bit strange, as you point out, to have the storage controller implement a software version of iSER, except perhaps for some type of compatibility or migration reasons. . . . John L Hufferd Sr. Executive Director of Technology Brocade Communications Systems, Inc [email protected] Office Phone: (408) 333-5244; eFAX: (408) 904-4688 Alt Office Phone: (408) 997-6136; Cell: (408) 627-9606 _____ From: Eddy Quicksall [mailto:[email protected]] Sent: Tuesday, April 18, 2006 5:48 AM To: Mike Ko Cc: [email protected] Subject: Re: [Ips] MPA alignment Thanks, Here is my feeling on this. If one were to use a standard TCP/IP stack (that does copying) then it is a bit silly to use iSER. If one were to use a software implementation and write/modify the stack to get the desired effects of DDP then he will probably heed to the alignment notes in the spec. So if one were to support misaligned segments using a slower method then it would be OK. Eddy ----- Original Message ----- From: Mike Ko <mailto:[email protected]> To: Eddy Quicksall <mailto:[email protected]> Cc: [email protected] Sent: Tuesday, April 18, 2006 2:16 AM Subject: Re: [Ips] MPA alignment Eddy, If your question is about MPA behavior in general, then it should be directed to the rddp forum. Assuming your question is related to iSER, then if iWARP is implemented in software and uses the existing TCP/IP stack, then it cannot guarantee that FPDU's will be aligned with TCP segments. For the general case where iWARP is implemented in an RNIC with an integrated TCP/IP stack, then it is to the RNIC vendor's advantage to optimize his product's performance and therefore FPDU's should be aligned with TCP segments. However, the receiver can still receive FPDU's which are not aligned with TCP segments if there are resegmenting middleboxes somewhere in the connection path. Mike To: <[email protected]> cc: Subject: [Ips] MPA alignment It is preferable to have TCP sender to align MPA header where a TCP segment begins and to keep the MPA FPDU within a TCP segment, so that the receiver hardware doesn't need to temporarily buffer the FPDU for the CRC checking. Does anyone know of any software or hardware that is not an FPDU alignment sender? Does anyone have a feel for the likelihood of software or hardware to not align FPDU's on a TCP segment? Eddy_______________________________________________ Ips mailing list [email protected] https://www1.ietf.org/mailman/listinfo/ips _____ _______________________________________________ Ips mailing list [email protected] https://www1.ietf.org/mailman/listinfo/ips _______________________________________________ Ips mailing list [email protected] https://www1.ietf.org/mailman/listinfo/ips