Hi, On 2014-1-13, at 4:40, Xuxiaohu <xuxiaohu@xxxxxxxxxx> wrote: >> I don't think it's right to try to solve this in MPLS, because MPLS is not a >> forwarding protocol - it's a connectivity protocol. right, MPLS is the wrong place to address is. The UDP encaps/decaps function needs to have this functionality. >> In any use of UDP, congestion >> control is either left to something above UDP or ignored (left to queue >> management). There are several cases, see Section 3.1.3 of RFC5405. MPLS-over-UDP can fall into any of the three cases, depending on what traffic is inside the LSP being encapsulated. You'll notice that RFC5405 for the first case - encapsulation of IP-based congestion controlled "normal" Internet traffic - even says that the tunnel SHOULD NOT employ any congestion control scheme of its own. Having layered control loops fighting is not productive. The issue with MPLS-in-UDP (and GRE-in-UDP, and any other encaps scheme that can carry non-IP traffic) are with cases two and three. When the workload that is being encapsulated isn't known to be congestion controlled by its endpoints, it is the obligation of the tunnel to detect congestion and react to it by reducing the traffic volume. Because for the rest of the network, that tunnel is the UDP sender, and we have IETF consensus that we don't want UDP senders that don't react to congestion on the net. (That's one of the main reasons for the existence of the RMCAT WG - we don't want non-congestion-controlled RTP media traffic on the net.) The key difference between putting MPLS e.g. into IP compared to putting it into UDP is that once it's in UDP, it can go pretty much anywhere on the net, because UDP traverses NATs and firewalls much more easily than IP traffic with a rare protocol number does. >> Similarly, you want the client of MPLS to be responsible for >> managing its traffic. MPLS gives you paths, it doesn't push packets over them. Right. However, once you slap a UDP header on a packet during encapsulation, you now subjected yourself to the rules for Internet UDP senders. Those are documented in RFC5405, and require the tunnel to implement some sort of congestion detection and control. I'd personally consider a circuit breaker mechanism sufficient, like RTP and I think PWE are using. > Fully agree. The congestion control should be performed either by the UDP tunnel itself or the client of MPLS. In the former case, it'd better to specify the practical congestion control mechanisms (if there were any) in a generic draft (e.g., RFC5405bis) and then any use of the UDP tunnel could refer to that generic draft with regard to congestion control. The general concept of a circuit breaker is easy enough that it doesn't really need to be written down. And it wouldn't be possible to describe it in a generic fashion, because congestion detection is typically specific to the protocol being encapsulated (e.g., RTP uses RTCP feedback to derive loss information, etc.) And the reaction to congestion is also dependent on the protocol being encapsulated (does it support multiple rates or only on/off, what timescales are OK for reaction, etc.) > In the latter case, if the client of MPLS is TCP-friendly, that is great. Otherwise (e.g., circuit emulation service), it shouldn't be deployed on the Internet at all, just as has been pointed out in RFC3985, therefore there is no need for any specific congestion control mechanism on the client. > > "... In essence, this requirement states that it is > not acceptable to deploy an application (using PWE3 or any other > transport protocol) on the best-effort Internet, which consumes > bandwidth arbitrarily and does not compete fairly with TCP within an > order of magnitude." (quoted from Section 6.5 of RFC3985) > > The above choice seems no conflict with the following congestion control guidelines as quoted from Section 3.1.1 of RFC5405, as those non-TCP-friendly traffic would be transported over a provisioned path, rather than on the Internet. > > "...Finally, some bulk transfer applications may choose not to implement > any congestion control mechanism and instead rely on transmitting > across reserved path capacity. This might be an acceptable choice > for a subset of restricted networking environments, but is by no > means a safe practice for operation in the Internet." How is that in conflict? Both quotes say that Internet traffic needs congestion control, which is a restatement of RFC2914. Lars
Attachment:
signature.asc
Description: Message signed with OpenPGP using GPGMail