Tunneling Compressed Multiplexed Traffic Flows (TCM-TF) Reference Model

This document describes a way to combine existing protocols for header compression, multiplexing and tunneling to save bandwidth for applications that generate long-term flows of small packets.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

The interactivity demands of some real-time services (VoIP, videoconferencing, telemedicine, video vigilance, online gaming, etc.) require a traffic profile consisting of high rates of small packets, which are necessary in order to transmit frequent updates between the two extremes of the communication. These services also demand low network delays. In addition, some other services also use small packets, although they are not delay-sensitive (e.g., instant messaging, M2M packets sending collected data in sensor networks using wireless or satellite scenarios). For both the delay-sensitive and delay-insensitive applications, their small data payloads incur significant overhead. When a number of flows based on small packets (small-packet flows) share the same path, bandwidth can be saved by multiplexing packets belonging to different flows. If a transmission queue has not already been formed but multiplexing is desired, it is necessary to add a multiplexing delay, which has to be maintained under some threshold if the service presents tight delay requirements.

The first design of the Internet did not include any mechanism capable of guaranteeing an upper bound for delivery delay, taking into account that the first deployed services were e-mail, file transfer, etc., in which delay is not critical. RTP was first defined in 1996 in order to permit the delivery of real-time contents. Nowadays, although there are a variety of protocols used for signaling real-time flows (SIP , H.323, etc.), RTP has become the standard par excellence for the delivery of real-time content. RTP was designed to work over UDP datagrams. This implies that an IPv4 packet carrying real-time information has to include 40 bytes of headers: 20 for IPv4 header, 8 for UDP, and 12 for RTP. This overhead is significant, taking into account that many real-time services send very small payloads. It becomes even more significant with IPv6 packets, as the basic IPv6 header is twice the size of the IPv4 header. illustrates the overhead problem of VoIP for two different codecs. IPv4 IPv6 IPv4+UDP+RTP: 40 bytes header IPv6+UDP+RTP: 60 bytes header G.711 at 20 ms packetization: 25% header overhead G.711 at 20 ms packetization: 37.5% header overhead G.729 at 20 ms packetization: 200% header overhead G.729 at 20 ms packetization: 300% header overhead

At the same time, there are many real-time applications that do not use RTP. Some of them send UDP (but not RTP) packets, e.g., First Person Shooter (FPS) online games , for which latency is very critical. The quickness and the movements of the players are important, and can decide if they win or lose a fight.

Other applications without delay constraints are also becoming popular (e.g., instant messaging, M2M packets sending collected data in sensor networks using wireless or satellite scenarios). The number of wireless M2M (machine-to-machine) connections is steady growing since a few years, and a share of these is being used for delay-intolerant applications, e.g., industrial SCADA (Supervisory Control And Data Acquisition), power plant monitoring, smart grids, asset tracking.

In the moments or places where network capacity gets scarce, allocating more bandwidth is a possible solution, but it implies a recurring cost. However, including optimization techniques between a pair of network nodes (reducing bandwidth and packets per second) when/where required is a one-time investment. Thus, in scenarios including a bottleneck with a single Layer-3 hop, header compression standard algorithms , , , can be used for reducing the overhead of each flow, at the cost of additional processing. However, if header compression is to be deployed in a network path including several Layer-3 hops, tunneling can be used in order to allow the header-compressed packets to travel end-to-end, thus avoiding the need to compress and decompress at each intermediate node. In these cases, compressed packets belonging to different flows can be multiplexed together, in order to share the tunnel overhead. In this case, a small multiplexing delay will be necessary as a counterpart, in order to join a number of packets to be sent together. This delay has to be maintained under a threshold in order to grant the delay requirements. A demultiplexer is necessary at the end of the common path, so as to rebuild the packets as they were originally sent, making multiplexing a transparent process for the extremes of the flow. If only one stream is tunneled and compressed, then little bandwidth savings will be obtained. In contrast, multiplexing is helpful to amortize the overhead of the tunnel header over many payloads. The obtained savings grow with the number of flows optimized together , .

As an additional benefit, the reduction of the sent information, and especially the reduction of the amount of packets per second to be managed by the intermediate routers, can be translated into a reduction of the overall energy consumption of network equipment. According to internal packet processing engines and switching fabric require 60% and 18% of the power consumption of high-end routers respectively. Thus, reducing the number of packets to be managed and switched will reduce the overall energy consumption. The measurements deployed in on commercial routers corroborate this: a study using different packet sizes was presented, and the tests with big packets made the energy consumption get reduced, since a certain amount of energy is associated to header processing tasks, and not only to the sending of the packet itself. All in all, a tradeoff appears: on the one hand, energy consumption is increased in the two extremes due to header compression processing; on the other hand, energy consumption is reduced in the intermediate nodes because of the reduction in the number of packets transmitted. Thi tradeoff should be explored more deeply.

This document uses a number of terms to refer to the roles played by the entities using TCM-TF. native packet A packet sent by an application, belonging to a flow that can be optimized by means of TCM-TF. native flow A flow of native packets. It can be considered a "small-packet flow" when the vast majority of the generated packets present a low payload-to-header ratio. TCM packet A packet including a number of multiplexed and header-compressed native ones, and also a tunneling header. TCM flow A flow of TCM packets, including a number of optimized native flows. TCM optimizer The host where TCM optimization is deployed. It corresponds to both the ingress and the egress of the tunnel transporting the compressed and multiplexed packets. If the optimizer compresses headers, multiplexes packets and creates the tunnel, it behaves as a "TCM-ingress optimizer", or "TCM-IO". It takes native packets or flows and "optimizes" them. If it extracts packets from the tunnel, demultiplexes packets and decompresses headers, it behaves as a "TCM-egress optimizer", or "TCM-EO". The TCM-egress optimizer takes a TCM flow and "rebuilds" the native packets as they were originally sent. TCM-TF session The relationship between a pair of TCM optimizers exchanging TCM packets. policy manager A network entity which makes the decisions about TCM-TF parameters: multiplexing period to be used, flows to be optimized together, depending on their IP addresses, ports, etc. It is connected with a number of TCM-TF optimizers, and orchestrates the optimization that takes place between them.

Different scenarios of application can be considered for the tunneling, compressing and multiplexing solution. They can be classified according to the domains involved in the optimization:

In this scenario, the TCMT-TF tunnel goes all the way from one network edge (the place where users are attached to the ISP) to another, and therefore it can cross several domains. As shown in , the optimization is performed before the packets leave the domain of an ISP; the traffic crosses the Internet tunnelized, and the packets are rebuilt in the second domain.

(_ -|TCM-IO|--- _) ---> ( ) `) ----->(_-|TCM-EO|--_)--> ( +------+ _) (_ (_ . _) _) ( +------+ _) (_ _ _ _) (_ _ ( _) _) ISP 1 Internet ISP 2 <------------------TCM-TF--------------------> ]]> Note that this is not from border to border (where ISPs connect to the Internet, which could be covered with specialized links) but from an ISP to another (e.g. managing all traffic from individual users arriving at a Game Provider, regardless users' location). Some examples of this could be: An ISP could place a TCM optimizer in its aggregation network, in order to tunnel all the packets of a service, sending them to the application provider, who would rebuild the packets before forwarding them to the application server. This would result in savings for both actors. A service provider (e.g., an online gaming company) could be allowed to place a TCM optimizer in the aggregation network of an ISP, being able to optimize all the flows of a game or service. Another TCM optimizer would rebuild these packets once they arrive to the network of the provider.

TCM-TF is only activated inside an ISP, from the edge to border, inside the network operator. The geographical scope and network depth of TCM-TF activation could be on demand, according to traffic conditions. If we consider the residential users of a real-time interactive application (e.g., VoIP, an online game generating small packets) in a town or a district, a TCM optimizing module can be included in network devices, in order to group packets with the same destination. As shown in , depending on the number of users of the application, the packets could be grouped at different levels in DSL fixed network scenarios, at gateway level in LTE mobile network scenarios or even in other ISP edge routers. TCM-TF may also be applied for fiber residential accesses, and in 2G/3G mobile networks. This would reduce bandwidth requirements in the provider aggregation network

( ` )_ +------+ ( ` )_ M users -|TCM-IO|------> ( ) `) --|TCM-EO|--> ( ) `) +------+ / ->(_ (_ . _) _) +------+ (_ (_ . _) _) / +------+ / ISP Internet P users -|TCM-IO|/ +------+ <------------TCM-TF--------------> ]]> At the same time, the ISP would implement TCM-TF capabilities within its own MPLS network in order to optimize internal network resources: optimizing modules could be embedded in the Label Edge Routers of the network. In that scenario MPLS would be the “tunneling” layer, being the tunnels the paths defined by the MPLS labels and avoiding the use of other tunneling protocols. Finally, some networks use cRTP in order to obtain bandwidth savings on the access link, but as a counterpart it consumes considerable CPU resources on the aggregation router. In these cases, by means of TCM, instead of only saving bandwidth on the access link, it could also be saved across the ISP network, without the CPU impact on the aggregation router.

End users can also optimize traffic end-to-end from network borders. TCM-TF is used to connect private networks geographically apart (e.g. corporation headquarters and subsidiaries), without the ISP being aware (or having to manage) those flows, as shown in , where two different locations are connected through a tunnel traversing the Internet or another network.

( ) `) --|TCM-EO|-->( ) `) (_ (_ . _) _) +------+ (_ (_ . _) _) +------+ (_ (_ . _)_) Location 1 ISP/Internet Location 2 <-----------TCM-TF----------> ]]> Some examples of these scenarios: The case of an enterprise with a number of distributed central offices, in which an appliance could be placed next to the access router, being able to optimize traffic flows with a shared origin and destination. Thus, a number of remote desktop sessions to the same server could be optimized, or a number of VoIP calls between two offices could also require less bandwidth and fewer packets per second. In some cases the tunnel is already included for security reasons, so the additional overhead of TCM-TF is lower. An Internet café, which is suitable of having many users of the same application (e.g., VoIP, a game) sharing the same access link. Internet cafés are very popular in countries with relatively low access speeds in households, where home computer penetration is usually low as well. In many of these countries, bandwidth can become a serious limitation for this kind of business, so TCM-TF savings may become interesting for their viability. Community networks (typically deployed in rural areas or in developing countries), in which a number of people in the same geographical place share their connections in a cooperative way, and a number of wireless hops are required in order to reach a router connected to the Internet. Satellite communication links that often manage the bandwidth by limiting the transmission rate, measured in packets per second (pps), to and from the satellite. Applications like VoIP that generate a large number of small packets can easily fill the maximum number of pps slots, limiting the throughput across such links. As an example, a G.729a voice call generates 50 pps at 20 ms packetization time. If the satellite transmission allows 1,500 pps, the number of simultaneous voice calls is limited to 30. This results in poor utilization of the satellite link's bandwidth as well as places a low bound on the number of voice calls that can utilize the link simultaneously. TCM optimization of small packets into one packet for transmission would improve the efficiency. In a M2M/SCADA (Supervisory Control And Data Acquisition) context, TCM optimization can be applied when a satellite link is used for collecting the data of a number of sensors. M2M terminals are normally equipped with sensing devices which can interface to proximity sensor networks through wireless connections. The terminal can send the collected sensing data using a satellite link connecting to a satellite gateway, which in turn will forward the M2M/SCADA data to the to the processing and control center through Internet. The size of typical M2M application transaction depends on the specific service and it may vary from a minimum of 20 bytes (e.g., tracking and metering in private security) to about 1,000 bytes (e.g., video-surveillance). In this context, TCM-TF concepts can be also applied to allow a more efficient use of the available satellite link capacity, matching the requirements demanded by some M2M services. If the case of large sensor deployments is considered, where proximity sensor networks transmit data through different satellite terminals, the use of compression algorithms already available in current satellite systems to reduce the overhead introduced by UDP and IPv6 protocols is certainly desirable. In addition to this, tunneling and multiplexing functions available from TCM-TF allows extending compression functionality throughout the rest the network, to eventually reach the processing and control centers. Desktop or application sharing where the traffic from the server to the client typically consists of the delta of screen updates. Also, the standard for remote desktop sharing emerging for WebRTC in the RTCWEB Working Group is: {something}/SCTP/UDP (Stream Control Transmission Protocol ). In this scenario, SCTP/UDP could be used in other cases: chatting, file sharing and applications related to WebRTC peers. There could be hundreds of clients at a site talking to a server located at a datacenter over a WAN. Compressing, multiplexing and tunneling this traffic could save WAN bandwidth and potentially improve latency.

Different combinations of the previous scenarios can be considered. Agreements between different companies can be established in order to save bandwidth and to reduce packets per second. As an example, shows a game provider that wants to TCM-optimize its connections by establishing associations between different TCM-IO/EOs placed in the game server and several TCM-IO/EOs placed in the networks of different ISPs (agreements between the game provider and each ISP would be necessary). In every ISP, the TCM-IO/EO would be placed in the most adequate point (actually several TCM-IO/EOs could exist per ISP) in order to aggregate enough number of users.

(_ (_ . _) +---+ ISP 1 \ _ _ \ _ _ _ _ _ M users ( ` )_ \ ( ` ) ( ` ) ( ` ) +---+ ( ) `) \ ( ) `) ( ) `) +---+ ( ) `) |TCM|->(_ (_ ._)---- (_ (_ . _) ->(_ (_ . _)->|TCM|->(_ (_ . _) +---+ ISP 2 / Internet ISP 4 +---+ Game Provider _ _ / ^ O users ( ` )_ / | +---+ ( ) `) / +---+ |TCM|->(_ (_ ._) P users->|TCM| +---+ ISP 3 +---+ ]]>

In conclusion, a standard able to compress headers, multiplex a number of packets and send them together using a tunnel, can benefit various stakeholders: network operators can compress traffic flows sharing a common network segment; ISPs; developers of VoIP systems can include this option in their solutions; service providers, who can achieve bandwidth savings in their supporting infrastructures; users of Community Networks, who may be able to save significant bandwidth amounts, and to reduce the number of packets per second in their networks. Other fact that has to be taken into account is that the technique not only saves bandwidth but also reduces the number of packets per second, which sometimes can be a bottleneck for a satellite link or even for a network router.

The current standard defines a way to reduce bandwidth and pps of RTP traffic, by combining three different standard protocols: Regarding compression, is the selected option. Multiplexing is accomplished using PPP Multiplexing Tunneling is accomplished by using L2TP (Layer 2 Tunneling Protocol ). The three layers are combined as shown in the :

In contrast to the current standard , TCM-TF allows other header compression protocols in addition to RTP/UDP, since services based on small packets also use by bare UDP, as shown in :

Each of the three layers is considered as independent of the other two, i.e. different combinations of protocols can be implemented according to the new proposal: Regarding compression, a number of options can be considered: as different standards are able to compress different headers (, , , ). The one to be used could be selected depending on the traffic to compress and the concrete scenario (packet loss percentage, delay, etc.). It also exists the possibility of having a null header compression, in the case of wanting to avoid traffic compression, taking into account the need of storing a context for every flow and the problems of context desynchronization in certain scenarios. Although non shown in , ESP (Encapsulating Security Payload ) headers can also be compressed. Multiplexing is also accomplished using PPP Multiplexing . Nevertheless, other multiplexing protocols can also be considered. Tunneling is accomplished by using L2TP (Layer 2 Tunneling Protocol ) over IP, GRE (Generic Routing Encapsulation ) over IP, or MPLS (Multiprotocol Label Switching Architecture ). It can be observed that TCRTP is included as an option in TCM-TF, combining , and . If a single link is being optimized a tunnel is unnecessary. In that case, both optimizers can perform header compression between both of them. Multiplexing may still be useful, since it reduces packets per second which is interesting in some environments (e.g., satellite). Another reason for that is the desire of reducing energy consumption. Although no tunnel is employed, this can still be considered as TCM-TF optimization, so TCM-TF signaling protocols can be employed here in order to negotiate the compression and multiplexing parameters to be employed. Payload compression schemes could also be used, but they are not the aim of this document.

This section describes how to combine protocols belonging to trhee layers (compressing, multiplexing, and tunneling), in order to save bandwidth for the considered flows.

TCM-TF can be implemented in different ways. The most straightforward is to implement it in the devices terminating the flows (these devices can be e.g., voice gateways, or proxies grouping a number of flows):

Another way TCM-TF can be implemented is with an external optimizer. This device could be placed at strategic places in the network and could dynamically create and destroy TCM-TF sessions without the participation of the endpoints that generate the flows ().

A number of already compressed flows can also be merged in a tunnel using an optimizer in order to increase the number of flows in a tunnel ():

There are different protocols that can be used for compressing IP flows: IPHC (IP Header Compression ) permits the compression of UDP/IP and ESP/IP headers. It has a low implementation complexity. On the other hand, the resynchronization of the context can be slow over long RTT links. It should be used in scenarios presenting very low packet loss percentage. cRTP (compressed RTP ) works the same way as IPHC, but is also able to compress RTP headers. The link layer transport is not specified, but typically PPP is used. For cRTP to compress headers, it must be implemented on each PPP link. A lot of context is required to successfully run cRTP, and memory and processing requirements are high, especially if multiple hops must implement cRTP to save bandwidth on each of the hops. At higher line rates, cRTP's processor consumption becomes prohibitively expensive. cRTP is not suitable over long-delay WAN links commonly used when tunneling, as proposed by this document. To avoid the per-hop expense of cRTP, a simplistic solution is to use cRTP with L2TP to achieve end-to-end cRTP. However, cRTP is only suitable for links with low delay and low loss. Thus, if multiple router hops are involved, cRTP's expectation of low delay and low loss can no longer be met. Furthermore, packets can arrive out of order. ECRTP (Enhanced Compressed RTP ) is an extension of cRTP that provides tolerance to packet loss and packet reordering between compressor and decompressor. Thus, ECRTP should be used instead of cRTP when possible (e.g., the two TCM optimizers implementing ECRTP). ROHC (RObust Header Compression ) is able to compress UDP/IP, ESP/IP and RTP/UDP/IP headers. It is a robust scheme developed for header compression over links with high bit error rate, such as wireless ones. It incorporates mechanisms for quick resynchronization of the context. It includes an improved encoding scheme for compressing the header fields that change dynamically. Its main drawback is that it requires significantly more processing and memory resources than the ones necessary for IPHC or ECRTP. This standard does not determine which of the existing protocols has to be used for the compressing layer. The decision will depend on the scenario, and will mainly be determined by the packet loss probability, RTT, and the availability of memory and processing resources. The standard is also suitable to include other compressing schemes that may be further developed.

When the compressor receives an RTP packet that has an unpredicted change in the RTP header, the compressor should send a COMPRESSED_UDP packet (described in ) to synchronize the ECRTP decompressor state. The COMPRESSED_UDP packet updates the RTP context in the decompressor. To ensure delivery of updates of context variables, COMPRESSED_UDP packets should be delivered using the robust operation described in . Because the "twice" algorithm described in relies on UDP checksums, the IP stack on the RTP transmitter should transmit UDP checksums. If UDP checksums are not used, the ECRTP compressor should use the cRTP Header checksum described in .

ROHC includes a more complex mechanism in order to maintain context synchronization. It has different operation modes and defines compressor states which change depending on link behavior.

Header compressing algorithms require a layer two protocol that allows identifying different protocols. PPP is suited for this, although other multiplexing protocols can also be used for this layer of TCM-TF. When header compression is used inside a tunnel, it reduces the size of the headers of the IP packets carried in the tunnel. However, the tunnel itself has overhead due to its IP header and the tunnel header (the information necessary to identify the tunneled payload). By multiplexing multiple small payloads in a single tunneled packet, reasonable bandwidth efficiency can be achieved, since the tunnel overhead is shared by multiple packets belonging to the flows active between the source and destination of an L2TP tunnel. The packet size of the flows has to be small in order to permit good bandwidth savings. If the source and destination of the tunnel are the same as the source and destination of the compressing protocol sessions, then the source and destination must have multiple active small-packet flows to get any benefit from multiplexing. Because of this, TCM-TF is mostly useful for applications where many small-packet flows run between a pair of hosts. The number of simultaneous sessions required to reduce the header overhead to the desired level depends on the average payload size, and also on the size of the tunnel header. A smaller tunnel header will result in fewer simultaneous sessions being required to produce adequate bandwidth efficiencies.

Different tunneling schemes can be used for sending end to end the compressed payloads.

L2TP tunnels should be used to tunnel the compressed payloads end to end. L2TP includes methods for tunneling messages used in PPP session establishment, such as NCP (Network Control Protocol). This allows to negotiate ECRTP compression/decompression parameters. Other tunneling schemes, such as GRE may also be used to implement the tunneling layer of TCM-TF.

In some scenarios, mainly in operator´s core networks, the use of MPLS is widely deployed as data transport method. The adoption of MPLS as tunneling layer in this proposal intends to natively adapt TCM-TF to those transport networks. In the same way that layer 3 tunnels, MPLS paths, identified by MPLS labels, established between Label Edge Routers (LSRs), could be used to transport the compressed payloads within an MPLS network. This way, multiplexing layer must be placed over MPLS layer. Note that, in this case, layer 3 tunnel headers do not have to be used, with the consequent data efficiency improvement.

The packet format for a packet compressed is:

The packet format of a multiplexed PPP packet as defined by is:

The combined format used for TCM-TF with a single payload is all of the above packets concatenated. Here is an example with one payload, using L2TP or GRE tunneling:

| |<-------- Mux payload --------->| ]]> If the tunneling technology is MPLS, then the scheme would be:

| |<-------- Mux payload --------->| ]]> If the tunnel contains multiplexed traffic, multiple "PPPMux payload"s are transmitted in one IP packet.