|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] FW: ipfc discussion of draft-otis-fc-sctp-ip-00.txtFor those not on the ipfc list, here is the discussion of draft-otis-fc-sctp-ip.txt on the use of SCTP and draft-xie-stewart-usctp-00.txt on a proposed unreliable delivery mechanism for SCTP. Anyone replying to this message who quotes or forwards its entire contents will be summarily shot :-). --David ----------------------------------------------------- From: Randall R. Stewart [randall@stewart.chicago.il.us] Sent: Thursday, August 24, 2000 9:47 AM To: ipfc@standards.gadzoox.com Subject: Re:draft-otis-fc-sctp-ip-00.txt Greetings: (I apologize for the earlier subscribe mis-send... I should know better than to do anything before a second cup of coffee in the A.M. :->) I have just finished reading both the above mentioned draft and the draft - draft-ietf-ipfc-fcoverip-02.txt I think that Douglas did a nice job of summarizing some of the benefits that FC over SCTP/IP can gain in this draft i.e.: - Headers contained within one frame. - Objects aligned at 32-bit boundaries. - Out of sequence frame processing. - Standard authentication. - Independent streams under common control. - Session restart. - Improved error detection. - Prevention of blind spoofing and denial of service attacks. - Standard Heartbeat and multi-homing. (optional) One other things not mentioned was: - Congestion control Now after reviewing the ipfc-fcoverip draft particularly the section that specifies why TCP should not be used, there may be objections to the Otis draft... One of problems mentioned in ipfc-fcoverip is the retransmission by TCP being in conflict to retransmissions that would occur at the FC level. I would like to call your attention to a draft that I think will solve this problem... http://search.ietf.org/internet-drafts/draft-xie-stewart-usctp-00.txt This draft defines a new extension that we have proposed for SCTP that allows a particular stream to be specified has "no retransmission". This way you can mix both completly reliable and unreliable transmissions over the same association AND still maintain correct congestion control policy's in the network. We still are debating the addition of a couple of things to this draft: - Should we have the ability to turn off checksums on selected data portions traveling on a stream. This way one could disable the alder-32 checksum on a unreliable stream packet but not the header, control, or any reliable data. - How do you handle larger than the P-MTU sized messages? Currently the draft dis-allows this since when sending to an unreliable stream you can never re-assemble the larger packet if a piece of one is lost. We may want to think about adding the ability to still fragment and hook in a mechanism to the reliable re-assembly queues to discard in-complete reassembly's where some datagrams are dropped due to the unreliable transmission. This would be desireable to keep from having to do a double P-MTU discover (one at the SCTP layer and the other at the IPFC layer). We would appreciate any comments on the draft and I would be more than glad to present either here on the mailing list or in San Diego a short overview of both SCTP or SCTP-U if this is helpful... Thanks for your time and I would welcome any comments or questions.. R -- Randall R. Stewart randall@stewart.chicago.il.us or rrs@cisco.com 815-342-5222 From: Wayland Jeong [wayland@troikanetworks.com] Sent: Thursday, August 24, 2000 12:42 PM To: 'Randall R. Stewart'; ipfc@standards.gadzoox.com Subject: RE: draft-otis-fc-sctp-ip-00.txt [ stuff deleted ] > Now after reviewing the ipfc-fcoverip draft particularly the > section that specifies why TCP should not be used, there may > be objections to the Otis draft... > > One of problems mentioned in ipfc-fcoverip is the retransmission > by TCP being in conflict to retransmissions that would occur > at the FC level. I would like to call your attention to a > draft that I think will solve this problem... > > http://search.ietf.org/internet-drafts/draft-xie-stewart-usctp-00.txt > > This draft defines a new extension that we have proposed for SCTP > that allows a particular stream to be specified has "no retransmission". > This way you can mix both completly reliable and unreliable transmissions > over the same association AND still maintain correct congestion control > policy's in the network. We still are debating the addition of > a couple of things to this draft: > The problem with this is that you don't want to rely on an I/O level retry to solve congestion problems in the network. If frames are being dropped due to congestion issues in the network (i.e. WRED) then having SCSI suffer a timeout and retry an entire I/O will help exacerbate congestion issues and bring performance to its knees. Depending on the traffic type, an I/O level retry may result in a 64K block retransmission after a timeout which is on the order of seconds. Unless you can absolutely rely on diffserv and/or RSVP or some other QoS mechanism to guarantee reliable delivery, the overall SCSI performance will suffer. I guess you can solve the problem with a well-tuned private network or a bandwidth matched WAN link, but this can be expensive and takes away from the business case for this technology. I think that any protocol which cannot tolerate lost data must have a reliable transport underneath it in order to achieve performance over gigabit networks. I'm not all that familiar with SCTP, but I thought that one of the features that would be useful is the capability of using SACK and message numbers to implement selective retransmission. It seems like, for storage, if we use SCTP as the transport, we want retransmission. -Wayland From: Randall R. Stewart [randall@stewart.chicago.il.us] Sent: Thursday, August 24, 2000 1:17 PM To: Wayland Jeong Cc: ipfc@standards.gadzoox.com Subject: Re: draft-otis-fc-sctp-ip-00.txt Wayland Jeong wrote: > > [ stuff deleted ] > > > Now after reviewing the ipfc-fcoverip draft particularly the > > section that specifies why TCP should not be used, there may > > be objections to the Otis draft... > > > > One of problems mentioned in ipfc-fcoverip is the retransmission > > by TCP being in conflict to retransmissions that would occur > > at the FC level. I would like to call your attention to a > > draft that I think will solve this problem... > > > > http://search.ietf.org/internet-drafts/draft-xie-stewart-usctp-00.txt > > > > This draft defines a new extension that we have proposed for SCTP > > that allows a particular stream to be specified has "no retransmission". > > This way you can mix both completly reliable and unreliable transmissions > > over the same association AND still maintain correct congestion control > > policy's in the network. We still are debating the addition of > > a couple of things to this draft: > > > > The problem with this is that you don't want to rely on an I/O level retry > to solve congestion problems in the network. If frames are being dropped due > to congestion issues in the network (i.e. WRED) then having SCSI suffer > a timeout and retry an entire I/O will help exacerbate congestion issues > and bring performance to its knees. Depending on the traffic type, an I/O > level retry may result in a 64K block retransmission after a timeout which > is on the order of seconds. Unless you can absolutely rely on diffserv > and/or RSVP or some other QoS mechanism to guarantee reliable delivery, > the overall SCSI performance will suffer. > The point of the u-sctp draft is that you DO NOT retransmit the data. You let the upper layer do it (if it so desires). Instead you drop your cwnd (as normal) and move forward the cumulative TSN point. There is NO retransmission on a u-sctp stream. Now we also did not specify in this draft if the upper layer should be informed of data not making it to the peer.. this may be a good thing to add ... > I guess you can solve the problem with a well-tuned private network or a > bandwidth matched WAN link, but this can be expensive and takes away > from the business case for this technology. > > I think that any protocol which cannot tolerate lost data must have a reliable > transport underneath it in order to achieve performance over gigabit networks. > I'm not all that familiar with SCTP, but I thought that one of the features > that would be useful is the capability of using SACK and message numbers > to implement selective retransmission. It seems like, for storage, if we use > SCTP as the transport, we want retransmission. Hmm. thats an option as well.. but if you look at: http://www.ietf.org/internet-drafts/draft-ietf-ipfc-fcoverip-02.txt it specifically states that you DO NOT want retransmission... In either case a SCTP endpoint that supported u-sctp and sctp would be able to have retransmission turned on OR turned off, on any particular stream... and it would be able to have one association doing both at the same time :-> You basically can have it any way you want it :-) R > > -Wayland -- Randall R. Stewart randall@stewart.chicago.il.us or rrs@cisco.com 815-342-5222 From: Wayland Jeong [wayland@troikanetworks.com] Sent: Thursday, August 24, 2000 10:37 PM To: 'Marjorie Krueger'; Randall R. Stewart Cc: Wayland Jeong; ipfc@standards.gadzoox.com Subject: RE: draft-otis-fc-sctp-ip-00.txt > It seems like people are missing the intended application of FCoverIP - namely, a > switching device, NOT an endpoint. Switching devices are not aware of the > protocol running over FC and must treat every FC packet similarly. SCSI is only > one application that runs over FC. An FCoverIP device can't selectively decide to > retransmit a packet without implementing some pretty complicated policies, and > looking into the headers of each packet, which would degrade performance so much > that it's questionable whether this would be a feasable application at all. This > has been discussed somewhat on the IPS reflector. > Layer 4 through layer 7 LAN switches exist today (Foundry, Alteon, Extreme, etc.) that inspect deep into each packet and make wire-speed decisions about how to service the data. There are routers and firewalls which terminate TCP and thus run entire stacks in the box. I'm not saying that it is trivial to design a port that can keep-up with the wire and run a session protocol in hardware, but it is not out of the question. If it was, iSCSI would be doomed as a high-performance gigabit solution (and that is an entirely different argument). I guess, what I'm trying to say is that someone should clearly state the requirements for FCoverIP. If the requirements are that it simply run on a private tuned network which can virtually guarantee a loss-less medium, then I would argue that we don't need congestion control and hence a retry mechanism. I would also argue that FCoverIP is a niche, proprietary solution which probably does not belong in IETF since it is not intended for the internet. If the requirement is that FCoverIP must co-exist with other LAN/WAN traffic, then I would argue that at a minimum we need a congestion management scheme since the networks depend on it. If we have a need for congestion management that implies that things are getting dropped (unless QoS ensures that does not happen.). And if things are getting dropped, then retries at the SCSI I/O level could severely limit performance. That was the convoluted reason for my comments about needing to have a retry mechanism. I believe that you need a reliable medium for transporting any SCSI-based protocol. > Marjorie Krueger > Vixel Corporation > -Wayland From: Randall R. Stewart [randall@stewart.chicago.il.us] Sent: Friday, August 25, 2000 4:45 AM To: Wayland Jeong Cc: 'Marjorie Krueger'; ipfc@standards.gadzoox.com Subject: Re: draft-otis-fc-sctp-ip-00.txt Wayland: My sole reason for pointing out the unreliable/non-retransmit option of SCTP is because of a working group document. In the document it is very very specific and states that the WG does not like TCP because of the retransmit storms that can occur if TCP and the FC layer both retransmit. I simply offered a solution, and since SCTP could offer both non-retransmit and retransmit it takes this argument off the table... Other comments in-line below: Wayland Jeong wrote: > > > It seems like people are missing the intended application of FCoverIP - namely, a > > switching device, NOT an endpoint. Switching devices are not aware of the > > protocol running over FC and must treat every FC packet similarly. SCSI is only > > one application that runs over FC. An FCoverIP device can't selectively decide to > > retransmit a packet without implementing some pretty complicated policies, and > > looking into the headers of each packet, which would degrade performance so much > > that it's questionable whether this would be a feasable application at all. This > > has been discussed somewhat on the IPS reflector. > > > Layer 4 through layer 7 LAN switches exist today (Foundry, Alteon, Extreme, etc.) that > inspect deep into each packet and make wire-speed decisions about how to service the > data. There are routers and firewalls which terminate TCP and thus run entire stacks in > the box. I'm not saying that it is trivial to design a port that can keep-up with the wire and > run a session protocol in hardware, but it is not out of the question. If it was, iSCSI would > be doomed as a high-performance gigabit solution (and that is an entirely > different argument). > > I guess, what I'm trying to say is that someone should clearly state the requirements for > FCoverIP. If the requirements are that it simply run on a private tuned network which can > virtually guarantee a loss-less medium, then I would argue that we don't need congestion > control and hence a retry mechanism. I would also argue that FCoverIP is a niche, proprietary > solution which probably does not belong in IETF since it is not intended for the internet. > If you ARE on a private net, even if Congestion Control is in place it SHOULD never be engaged. Here a SCTP with non-retransmit will help you. You never have the danger of a retransmit storm, and the fact that your network IS so reliable means you never get the CC algorithms working against you . Also if someone makes a mistake, and puts a private configuration across the big I-Internet it provides a protection.. > If the requirement is that FCoverIP must co-exist with other LAN/WAN traffic, then I would > argue that at a minimum we need a congestion management scheme since the networks > depend on it. If we have a need for congestion management that implies that things are > getting dropped (unless QoS ensures that does not happen.). And if things are getting > dropped, then retries at the SCSI I/O level could severely limit performance. And here, no matter what the performance level, one could turn on the reliable side transport of SCTP since it IS on a WAN and the retransmit on the SCTP level would save you from hitting SCSI I/O retransmits.. > > That was the convoluted reason for my comments about needing to have a retry mechanism. > I believe that you need a reliable medium for transporting any SCSI-based protocol. > > > Marjorie Krueger > > Vixel Corporation > > > > -Wayland R -- Randall R. Stewart randall@stewart.chicago.il.us or rrs@cisco.com 815-342-5222
Home Last updated: Tue Sep 04 01:07:43 2001 6315 messages in chronological order |