ERCIM News No.37 - April 1999
Wormhole IP over ATM
by Manolis Katevenis
Speeding up the Internet is of capital importance; this will need some hardware assistance. Hardware switches operate best on fixed-size quanta, while IP packets, on the other hand, have a variable size. It is not the first time, though, that hardware is called upon to route variable-size packets at high speed: Wormhole Routers did precisely that, in the eighties, for multiprocessor inter-connection networks. The same techniques are applicable today to the Internet: we introduced a novel, wormhole-like way to route IP over ATM, we simulated its performance, we built a first prototype, and we are now proceeding to testing it in real systems.
IP is the uncontested protocol for data communications. At the same time, ATM technology finds widespread use, owing to its fixed-size cells that allow high-speed hardware switching, owing to the small size of these cells that allows fast preemption and hence low latency, and owing to its quality-of-service (QoS) architecture. IP can fruitfully run on top of ATM, thus getting the best of both worlds. This has been done in software, on general-purpose computers; we do it in hardware, at a lower cost and with a lower latency.
ATM has some similarity with wormhole routing, the most popular multiprocessor interconnection network technique of the eighties. Just like virtual channels in wormhole routing carry packets segmented into flits, a number of hardware-managed VCs in ATM can carry IP packets segmented into cells according to AAL-5; each VC is dedicated to one packet for the duration of that packet, and is afterwards reassigned to another packet, in hardware. This was proposed by Barnett in 1997 and was named connectionless ATM. We, at ICS-FORTH in Heraklion, Crete, Greece, modified the Barnett proposal to make it applicable to existing ATM equipment: we proposed a novel single-input, single-output Wormhole IP Router, that functions as a VP/VC translation filter between ATM subnetworks.
Our wormhole IP routing filter has a number of advantages: (i) it works together with standard, existing ATM equipment; (ii) it allows the co-existence and integration of both IP and native ATM traffic in the same networks; (iii) the quality of service of native ATM traffic can stay unaffected by the added IP traffic, while IP can benefit from ATMs QoS capabilities; (iv) for IP traffic, the system operates equivalently to a network of low-latency gigabit IP routers, while being a lot less expensive; (v) packet routing delay is minimized owing to virtual-cut-through routing - segmentation and reassembly delays at intermediate routers are eliminated; (vi) packet routing delay is minimized for all packets - not just for the rest of the packets after a flow has been recognized, as in IP switching; (vii) the number of pre-established connections (labels) is small and fixed, and does not grow with the size of the network (as in tag switching), yet all packets are routed through pre-established connections. Based on actual internet traces, we showed by simulation that a few tens of hardware-managed VCs per outgoing VP suffice for all but 0.01% or less of the packets.
We have built a first prototype of a bi-directional wormhole IP routing filter with two OC-3 ports (155 Mbps in each direction), which is shown in the photograph. Two 16-MByte DRAM SIMMs plug into the sockets on the right; they hold the two-level IP routing table and the ATM connection table. The datapath, the control FSM, and the VCout free list (bit map) are contained in one FPGA. The routing delay is fixed, equal to about one cell time (a couple of micro-seconds), for all IP packets.
We are currently testing this prototype by connecting it to our ATM and IP network. Further development is needed, especially on the software side, and we are seeking partners for that and for commercially exploiting the technology.
For more information, see http://archvlsi.ics.forth.gr/wormholeIP/
Manolis Katevenis - ICS-FORTH
Tel: +30 81 391664