How about DTM, Frank?
A New Medium for IP?
netinsight.se
Over the past year, the networking industry has become increasingly disillusioned with ATM protagonists who promised to deliver a panacea for all problems. Together with the stupendous growth of IP traffic, this is fuelling the debate over alternate networking technologies such as IP over SDH. This month, ETSI, the European standards body, is considering an entirely new approach to the problem of moving vast amounts of IP packets. Known as dynamic synchronous transfer mode (DTM), it promises to be the network architecture that will support both current services as well as future IP-based multimedia applications.
--------------------------------------------------------------------------------
The next generation of networks will need to support traditional telecommunication services as well as quickly growing volumes of Internet-based services. The networks will be technically integrated and with a purpose to support different kinds of applications: non-real-time applications, file transfer and e-mail, as well as delay-sensitive real-time applications, such as voice, hi-fi-audio and video.
Data networks have traditionally provided asynchronous communication, which is typically packet-switched and uses store-and-forward techniques, such as the Internet. Multiplexed telephone networks, on the other hand, have provided real-time communication by means of circuit-switched, time-division techniques.
ATM has been introduced as a unifying transfer mode for all kinds of services, with an emphasis on network utilisation. However, problems with the traffic management of ATM have become more and more apparent, and it seems unlikely that ATM can cope with those services that need high Quality of Service (QoS) guarantees, such as voice, hi-fi-audio and video in a sufficiently economical manner. IP over ATM has been found to be a less elegant combination of protocols. New solutions for IP are being sought within the IP domain, in particular different methods to establish connections are being proposed.
In the telecoms domain, new solutions are being sought to integrate IP traffic with the existing telecommunications network for voice. The ETSI project TIPHON is an example of these efforts. The phenomenon that IP traffic is becoming more and more real-time increases the need to find a transfer mode that can accommodate real-time as well as non real-time traffic, without the need for very complex traffic policing methods.
The DTM approach is, therefore, to go back to circuit-switching, since the capacity of fibres can overcome the previous problem that bandwidth was a scarce and expensive resource that had to be utilised fully by means of sophisticated time-sharing mechanisms. Circuit-switched networks have many attractive properties. Circuits are isolated from each other in the sense that traffic on one circuit is unaffected by activities on the others. This makes it possible to provide guaranteed transfer quality with constant delay, which is suitable for applications with timing requirements.
Furthermore, data and control are separated in circuit-switched networks. Processing of control information only takes place at establishment and tear-down of circuits, and the actual data transfer can be done without processing of the data stream, congestion control, and so on. This allows large volumes of data to be transferred efficiently. QoS management will be even more important in the future, as developments in photonics will dramatically reduce the cost of transmission. The cost of routers and switches for high bandwidth communication under strict QoS guarantees will become a strategic problem for telecoms operators.
However, the static nature of ordinary circuit-switched networks makes them less useful for small volume, bursty types of traffic. Traditionally, circuits have fixed capacity, long set-up delay and poor support for multicast. These shortcomings make it difficult to efficiently support, for example, bursty data communications in a circuit-switched network. DTM, however, is designed to overcome the shortcomings of traditional circuit-switching while retaining the desired properties of ATM: bandwidth-on-demand and ability to serve applications with varying quality requirements.
The basic principle is to avoid traffic policing while enabling bandwidth-on-demand via a simple mechanism for dynamic time slot allocation. Fibre bandwidth can be increased at a lower price than the increase in computer capacity for processing and buffering. Therefore, DTM being a system based on fibre capacity, can scale much more economically than a system that relies heavily on processing and buffering, such as traditional TCP/IP and ATM.
What is DTM?
DTM is an effort to combine the advantages of asynchronous and synchronous media access schemes. It is fundamentally a time division multiplexing scheme, and as with other such schemes, it guarantees each host a certain bandwidth, and uses a large fraction of the available bandwidth for effective data transfer. In addition, the DTM scheme has in common with asynchronous schemes, such as ATM, support for dynamic reallocation of bandwidth between hosts. This means that the network can adapt to variations in the traffic, and divide its bandwidth between hosts according to their demand.
Hosts connected to a DTM network communicate with each other on channels. A DTM channel is a dynamic resource that can be set up with a bandwidth ranging from 512 kbps in quantum steps of 512 kbps up to the bandwidth of the fibre. To each channel a set of network time slots are initially associated. These time slots can then be changed dynamically during the lifetime of the channel. Since the time slots can be guaranteed during the lifetime of the channel, it is possible to give guarantees from the DTM system to higher levels about real-time properties.
DTM channels are multicast in nature. This means that any channel at a given time can occupy one sender and any number of receivers (Figure 1). Over the network, any number of multicasting groups can be active simultaneously. The extreme case of multicasting is broadcasting, where one node is sending and all other nodes are receiving.
DTM is a switching as well as a transmission technique that can serve as a substitute for ATM since it offers stronger bandwidth management than ATM, and scales more easily up to higher bandwidths. Functionally, it is similar to SDH/SONET. It may work stand-alone or run on top of a SDH/SONET pipe, although the overlaying of DTM over SDH/SONET does not offer any advantages over pure DTM.
DTM Characteristics
DTM is designed for a unidirectional medium with multiple access -- a medium with capacity shared by all connected nodes. It can be built on several different local topologies, such as ring, folded bus or dual bus. The service provided is based on channels. A channel is a set of time slots with a sender and an arbitrary number of receivers; it is guaranteed that the data will reach the receivers at the rate given by the acknowledged capacity of the channel. The channels on the physically shared medium are realised by a time division multiplexing (TDM) scheme.
There is, however, a trade-off between throughput and access delay in DTM. The time delay to establish channels depends on the availability of slots and the number of hops in the network. For small channels, the needed slots are more likely to be immediately available, while for large channels, the node may have to ask other nodes for more slots.
It is reasonable to assume that a session that has very high demands on throughput and service quality, such as two-way video, can tolerate an access delay in the order of 10-20 milliseconds. On the other hand, for small IP-packet streams that arrive frequently enough, it can be justified to have the network permanently fully connected with a guaranteed minimum of 512 kbps.
Signalling delay associated with creation and deletion of communication channels determines much of the efficiency of fast circuit-switching. DTM is therefore designed to create channels fast, within a fraction of a millisecond.
Recent developments in fibre-optics, including dense WDM, makes it possible to transfer large amounts of data on a single fibre, but today there is no commercially available technique that can fully switch the data sent on such a fibre and connect the fibres to build a flexible network. The growth of demands for processors and buffers tend to grow exponentially, thus making such solutions impossible in economical terms.
DTM, that can make use of the full capacity of fibre-optics to build a cost-efficient high-capacity network with low latency, almost no jitter, completely separated traffic flows and flexible resource reservation. It is possible to carry IP over DTM in such a way that IP traffic takes full advantage of the high capacity and QoS support in DTM, while still providing very little overhead to conventional low bandwidth best-effort traffic.
IP Traffic Over DTM
For IP over DTM to be efficient, the statistical characteristics of IP-traffic has to be taken into consideration. One important issue is flow management. Flows are important because they should be mapped to channels in DTM. The problem is that in IP there is no information available on how much traffic will be sent on a specific flow or when a flow actually ends. One method of getting around this problem is to estimate traffic. Studies have indicated that in a wide area environment, 50-60 per cent of flows are less than 200 bytes, and 70-80 per cent of the flows consist of less than 10 packets. Recent research has also shown that different protocols have very different characteristics, some protocols create many short-lived flows, whereas some create very few but long-lived flows. In addition, it is hard to identify flow characteristics based on information about different protocols.
It is obvious that a network that can transport large amounts of IP traffic without utilising expensive and complicated high-capacity IP routers, will need to treat the two different classes of IP flows, long-lived bandwidth and short-lived low bandwidth, separately. The large flows are better served if they can be switched at the DTM level of the network, whereas it would be inefficient to establish channels for all the small flows that constitute a large part of the total number of IP flows.
In addition to transporting large amounts of IP traffic, the network should also be capable of delivering QoS-flows with strict requirements for high bandwidth, low latency and low jitter. There are two DTM-based approaches available: IP over DTM (IPOD) and DTM LAN emulation (DLE).
IPOD is a technique that fully utilises the DTM networks in transmitting IP traffic both hop-by-hop and via shortcuts. DLE is used to establish virtual LANs across the DTM network and makes it possible to attach Ethernet-nodes to the DTM network.
IPOD specifies how to run IP directly on top of DTM. It also leaves a possibility to add support for other protocols in the same framework if the need arises. To combine the advantages of IP?s connection-less service with DTM?s capability of high-bandwidth data transfer, IPOD supports both hop-by-hop routing through the IPOD network and establishment of direct channels between sender and receiver. This gives IPOD its unique capabilities to efficiently carry both best-effort and real-time traffic streams.
In fact, IPOD bears a lot of resemblance with MPOA (multiprotocol over ATM), but while MPOA builds on the concept of emulated LANs, IPOD does not need to partition the network into LANs. IPOD and MPOA also differs in a lot of details due to differences between ATM and DTM.
The IPOD system builds a logical routing structure on top of the DTM network. This structure does not necessarily match the physical links in the network. The logical routing structure only describes how packets are forwarded hop-by-hop through the network whereas the DTM layer, independently of the logical routing structure, routes the shortcuts. It is, for example, possible to connect the DTM network like a mesh, but make the IP routing structures hierarchical. It is also possible to change the logical structure of the IPOD network by setting up direct channels between routers and thus making them adjacent.
Figure 2 shows an example of an IPOD network. The lines symbolise DTM channels that are used for sending IP traffic between the nodes. The nodes C and D are IPOD routers. They are full-fledged IP routers running IPOD. A, B, E and F are IPOD border routers that route traffic between DTM and other network technologies. IPOD border routers are IPOD routers with additional interface cards to other network technologies. G and H are IPOD clients in computers directly attached to the DTM network with a DTM network interface card.
Each IPOD client is assigned an IPOD router as its default gateway. This means that all packets for which the IPOD client has no explicit routing information are sent to its default gateway. IP packets sent between two IPOD clients could either be forwarded hop-by-hop via the IPOD routers, or a shortcut could be established using next hop routing protocol (NHRP) address resolution (Figure 3). The address-resolution mechanism uses the same information as the normal routing procedures, thus making it easier to implement and manage.
DTM LAN Emulation
DTM LAN emulation allows DTM to be used as a bridge between different segments of an Ethernet network. Packets are forwarded through the DTM network based on the destination-address in the Ethernet header. This allows the formation of virtual LANs where a number of nodes can act as if they were connected to the same Ethernet LAN, whereas they may be on different Ethernet segments connected via DTM or even directly connected to the DTM network via a DTM NIC. The DLE is completely transparent to all connected nodes. This makes it possible to connect all standard Ethernet equipment to the Ethernet segments.
DLE is similar in functionality to the ATM Forum?s LAN Emulation (LANE), but they differ in details due to differences between ATM and DTM. Each DLE segment consists of one DLE server (DLES) and several DLE clients (DLECs). There is one DLE client for each Fast Ethernet Gateway (FEG) and one for every directly connected DTM node in the DLE segment.
Scaling the Problem
DTM was designed from the ground up with scalability in mind. Its simple forwarding mechanism with limited buffers and no possibility of buffer overflows, no priority queuing mechanisms and no data processing, could make it possible to build very high capacity switches at a reasonable cost. IPOD addresses the scalability problem of an ordinary IP network by moving as much of the traffic as possible away from the routers. This means that each kind of traffic is handled by the networking technique that is the most appropriate. The routers handle short-lived flows and single packets providing short set-up latency to the IP traffic. DTM handles the long-lived, high bandwidth flows, and flows with QoS demands, thus relieving the routers from a lot of the traffic. By adjusting the flow identification mechanisms, it is possible to make a trade-off between bandwidth usage in the network and processing power in the routers.
The trend towards more bandwidth intensive applications and thus larger flows also means that more and more traffic will be eligible for switching. The logical routing structure established by IPOD on top of DTM can easily be changed and new routers added as the need arises, without changing the physical topology of the DTM network. This gives the ability to easily upgrade the network to handle larger amounts of routed traffic. t
How DTM Works
DTM divides the total fibre capacity into frames of 125 microseconds, which are further divided, into 64-bit slots. The number of slots per frame is dependent on the bitrate. With the planned bitrate of 2.5 Gbps, the number of slots is around 4800. The choice of a frame time of 125 microseconds and 64 bits per slot will enable very simple adaptations to digital voice and ISDN.
The slots are separated into data and control slots. At any point in time (Figure 4), every slot is either a data slot or a control slot. Each node has access to at least one control slot, which is used for sending control information to the other nodes. Control messages can be sent upon request from a user, in response to control messages from other nodes or spontaneously for management purposes.
The control slots constitute a small fraction of the total capacity, while the majority of the slots are data slots carrying payload. At system start-up, the data slots are allocated to the nodes according to some predefined distribution. This means that each node ?owns? a portion of data slots. A node needs ownership of a slot to send data in it, and the ownership of slots may change dynamically among the nodes during the operation of the network.
DTM uses a distributed algorithm for slot reallocation, where the pool of free slots is distributed among the nodes. At the reception of a user request, the node first checks its own time slots to see if it has slots enough to satisfy the request and, if so, immediately sends a channel establishment message to the next hop. Otherwise, the node first has to request more slots from the other nodes on the link. Each node maintains a status table that contains information about free slots in other nodes, and when more slots are needed the node consults its status table to decide which node to ask for slots. Every node regularly sends out status messages with information about its local pool of slots.
Each node has a network controller which handles the node-to-node signalling. This signalling is done via the control slots and is used for channel management and time slots reservation.
When a DTM user wants to set up a channel, it sends a ?create primitive? to the network control element which allocates the necessary bandwidth and sends an ?announce? message to the receiving node. The network control element also sets up the channel tables in the DTM link layer and sends back an ?indication primitive?, notifying the sending user that the requested capacity, that is the requested number of slots per frame, have been allocated. The user either waits for a confirmation from the receiving side (video, voice) or sends data directly after receiving the indication (datagram).
After a complete transfer, the user sends a remove message to the receiver notifying that the channel is to be closed and to de-allocate the reserved time slots. The actual data transfer is handled by the network without requiring any additional processing elements in the switches.
Lars Kahn is director of Net Insight. |
