Re: [Anima] GRASP maximum message size considerations
On 23-Aug-21 12:26, Michael Richardson wrote: > > Brian E Carpenter wrote: > >> Brian E Carpenter wrote: > >> > (1) Flooding (M_FLOOD) messages. These are UDP multicasts, so in > effect > >> > all nodes must agree on the same maximum size. To send messages above > >> > the present limit, the maximum flood message size would have to be > >> > increased everywhere in the autonomic network. That is trivial if > > we > >> > allow operator configuration, but since an AN should be > self-creating, > >> > we want to avoid operator configuration. Therefore, we need GRASP > > to be > >> > able to self-configure this. > >> > >> For the flooded messages over UDP, it seems unwise to ever assume we > can > >> reliably get more than 1280 through. In production, this goes over > IPsec ESP tunnels. > > > Why would that break IPv6 fragmentation? We can assume a well-defined > > MTU within an autonomic network, I think. These are all link-local > > addressed packets, so there is no PMTUD problem. > > But, it's not over link-local. > It's over a mesh of point to point IPsec ESP tunnels. Sure, but the VRF has to emulate LL multicast to ff02::13, and there must be a well-defined MTU for that emulation, regardless of the mesh. > Choices are: > 1) fragment before encrypt. Reassemble after decrypt. Please. > 2) encrypt and then fragment the ESP. > Depends upon ESP assembly buffer being big enough. Please not. Brian > > Both tend to work, up to some ill-defined limit which is not always 64K. > (1) is likely easier to fix if it's broken, since re-assembly happens in the > control plane CPU, rather than, possibly, in some IPsec hardware. > > -- > Michael Richardson. o O ( IPv6 IøT consulting ) >Sandelman Software Works Inc, Ottawa and Worldwide > ___ Anima mailing list Anima@ietf.org https://www.ietf.org/mailman/listinfo/anima
Re: [Anima] GRASP maximum message size considerations
Brian E Carpenter wrote: >> Brian E Carpenter wrote: >> > (1) Flooding (M_FLOOD) messages. These are UDP multicasts, so in effect >> > all nodes must agree on the same maximum size. To send messages above >> > the present limit, the maximum flood message size would have to be >> > increased everywhere in the autonomic network. That is trivial if > we >> > allow operator configuration, but since an AN should be self-creating, >> > we want to avoid operator configuration. Therefore, we need GRASP > to be >> > able to self-configure this. >> >> For the flooded messages over UDP, it seems unwise to ever assume we can >> reliably get more than 1280 through. In production, this goes over IPsec ESP tunnels. > Why would that break IPv6 fragmentation? We can assume a well-defined > MTU within an autonomic network, I think. These are all link-local > addressed packets, so there is no PMTUD problem. But, it's not over link-local. It's over a mesh of point to point IPsec ESP tunnels. Choices are: 1) fragment before encrypt. Reassemble after decrypt. 2) encrypt and then fragment the ESP. Depends upon ESP assembly buffer being big enough. Both tend to work, up to some ill-defined limit which is not always 64K. (1) is likely easier to fix if it's broken, since re-assembly happens in the control plane CPU, rather than, possibly, in some IPsec hardware. -- Michael Richardson. o O ( IPv6 IøT consulting ) Sandelman Software Works Inc, Ottawa and Worldwide signature.asc Description: PGP signature ___ Anima mailing list Anima@ietf.org https://www.ietf.org/mailman/listinfo/anima
Re: [Anima] GRASP maximum message size considerations
One point in line: On 22-Aug-21 10:43, Michael Richardson wrote: > > Brian E Carpenter wrote: > > (1) Flooding (M_FLOOD) messages. These are UDP multicasts, so in effect > > all nodes must agree on the same maximum size. To send messages above > > the present limit, the maximum flood message size would have to be > > increased everywhere in the autonomic network. That is trivial if we > > allow operator configuration, but since an AN should be self-creating, > > we want to avoid operator configuration. Therefore, we need GRASP to be > > able to self-configure this. > > For the flooded messages over UDP, it seems unwise to ever assume we can > reliably get more than 1280 through. In production, this goes over IPsec ESP > tunnels. Why would that break IPv6 fragmentation? We can assume a well-defined MTU within an autonomic network, I think. These are all link-local addressed packets, so there is no PMTUD problem. Brian > > At some point, we probably should specify use of > IP-TFS: Aggregation and Fragmentation Mode for ESP and its Use for IP > Traffic Flow Security > draft-ietf-ipsecme-iptfs-09 > > This would provide a better fragmentation regime than what we do now, and > there maybe some traffic analysis and predictability advantages here. > (If ACP traffic *always* takes 1% of bandwidth, it becomes more predictable, > but also there may be congestion advantages to having "reserved" that > bandwidth) > > > (2) Unicast messages (discovery responses, synchronization, > > negotiation). On investigation, it is very easy to allow such messages > > over TCP to be of any reasonable length, with no protocol change and > > without all nodes necessarily agreeing on what a "reasonable length" > > is. I discovered that my demo implementation of GRASP would already > > handle messages up to 4096 bytes, and tweaked the code to remove that > > limitation. I tested it with a GRASP objective exceeding 5000 bytes. > > For the messages over TCP, which in response to a UDP flooded or discovery > message, the discovering node can announce it's ability to receive bigger > messages (and the size). > > In the Discovery Response Message, the discovered node could initiate it's > maximum too. > > How to fit this into the messages, I don't know yet. > > It's not clear me that we need to do this network-wide, but we could do that. > > -- > Michael Richardson. o O ( IPv6 IøT consulting ) >Sandelman Software Works Inc, Ottawa and Worldwide > ___ Anima mailing list Anima@ietf.org https://www.ietf.org/mailman/listinfo/anima
Re: [Anima] GRASP maximum message size considerations
Brian E Carpenter wrote: > (1) Flooding (M_FLOOD) messages. These are UDP multicasts, so in effect > all nodes must agree on the same maximum size. To send messages above > the present limit, the maximum flood message size would have to be > increased everywhere in the autonomic network. That is trivial if we > allow operator configuration, but since an AN should be self-creating, > we want to avoid operator configuration. Therefore, we need GRASP to be > able to self-configure this. For the flooded messages over UDP, it seems unwise to ever assume we can reliably get more than 1280 through. In production, this goes over IPsec ESP tunnels. At some point, we probably should specify use of IP-TFS: Aggregation and Fragmentation Mode for ESP and its Use for IP Traffic Flow Security draft-ietf-ipsecme-iptfs-09 This would provide a better fragmentation regime than what we do now, and there maybe some traffic analysis and predictability advantages here. (If ACP traffic *always* takes 1% of bandwidth, it becomes more predictable, but also there may be congestion advantages to having "reserved" that bandwidth) > (2) Unicast messages (discovery responses, synchronization, > negotiation). On investigation, it is very easy to allow such messages > over TCP to be of any reasonable length, with no protocol change and > without all nodes necessarily agreeing on what a "reasonable length" > is. I discovered that my demo implementation of GRASP would already > handle messages up to 4096 bytes, and tweaked the code to remove that > limitation. I tested it with a GRASP objective exceeding 5000 bytes. For the messages over TCP, which in response to a UDP flooded or discovery message, the discovering node can announce it's ability to receive bigger messages (and the size). In the Discovery Response Message, the discovered node could initiate it's maximum too. How to fit this into the messages, I don't know yet. It's not clear me that we need to do this network-wide, but we could do that. -- Michael Richardson. o O ( IPv6 IøT consulting ) Sandelman Software Works Inc, Ottawa and Worldwide signature.asc Description: PGP signature ___ Anima mailing list Anima@ietf.org https://www.ietf.org/mailman/listinfo/anima