The explosion in the volume of electronic communication has resulted in an acute need for high-performance networks. Some promising all-optical switching networks (ASNs), such as optical burst-switched (OBS)1 and optical packet-switched networks, have been proposed to meet this need. ASNs use optical switches rather than optical-electrical-optical conversion to achieve transparent transmission. Since wavelength conversion (WC) and optical random access memory (RAM) are not readily available so far, contention will occur when two or more switching entities (e.g., packets, bursts) from different input ports need to be forwarded simultaneously on the same wavelength and output port.
Many methods for resolving such conflicts have been studied. The literature2 explores three dimensions of contention-resolution schemes (wavelength, time, and space) using wavelength conversion, fiber delay lines (FDLs), and deflection routing, respectively. Although wavelength conversion is a good method, the technology is not mature at present. FDLs and deflection routing increase data latency. Thus, under the current and foreseeable limitations of optical technology, ASN performance is mainly hampered at the network node by conflict over resources. To address this problem, we have developed a novel node architecture called a packet calking switch.
Figure 1. In the packet calking switch architecture, the gaps between switching entities are filled by appropriately inserting IP packets. IP: Internet protocol. RAM: Random access memory. WDM: Wavelength-division multiplexing. Tx: Transmitter. Rx: Receiver.
Calking is the process of making seams tight. Packet calking means filling the gaps between switching entities by inserting Internet protocol (IP) packets. The basic function of a packet calking switch is as follows. First, the node categorizes the traffic into two types, single-hop traffic and multihop traffic, based on whether it uses one or more physical links. Second, the node supports prioritized packet transmission, and the pass-by multihop traffic is always forwarded first. Third, newly added traffic can be aggregated if needed. Finally, single-hop traffic is calked into gaps in the multihop traffic for transmission.
An example of packet calking switch architecture using OBS is shown in Figure 1. It has two important functions: edge aggregating and core switching. IP packets from the access interface are routed by the traffic sorter according to their destinations. If they are going to an adjacent node, the packet is queued in the single-hop queue. If not, it goes to a multihop queue. IP packets in multihop queues are assembled into bursts according to assembly algorithms. Bursts and single-hop IP packets are stored in the electronic RAM. If the node detects that a wavelength is idle for a specific time period, single-hop IP packets are aligned as a calking entity with a duration to fit the period. Figure 2 traces the calking process for four pass-by bursts, labeled 1 to 4, and two calking entities marked C1 and C2, respectively. Before the node transmits a calking entity, the node sends a calking control packet to inform the adjacent node.
Figure 2. The calking process of an OBS node, showing four pass-by bursts (1–4) and two calking entities (C1 and C2). WC: Wavelength conversion. E/O: Electrical to optical.
Figure 3. This plot of drop probability versus offered load shows that packet calking decreases the probability of dropped packets. PWC: Partial wavelength conversion. FDL: Fiber delay line.
Figure 4. This plot of link utilization versus offered load demonstrates the 5∼25% improvement with packet calking. BL: Bottleneck link. OL: Ordinary link.
We have developed and simulated five schemes in a four-node line network: calking, no calking, FDL, partial WC (PWC), and calking combined with PWC. We selected an assembly algorithm3 and made the following assumptions for the simulation: IP packet arrival is a Poisson process, in which length follows an exponential distribution, with a mean length of 1250 bytes. Offset time is 0.5ms, and switching and processing time is 5μs. All nodes employ just enough time1 signaling and a first-in first-out scheduling policy. Figure 3 shows the drop probability versus the offered load. Simulation results show that packet calking decreases the probability of dropped packets by about 50% compared to the no-calking scheme. Figure 4 demonstrates the link utilization of the bottleneck link (BL) and ordinary links (OL). Link utilization is improved 5∼25%, which varies with the offered load in this simple network.
This node architecture is suitable for prioritized packet transmission. It is more cost-effective than the existing node architectures because it requires many fewer optical switches and less WC to achieve nearly the same packet-dropping probability as the node configured with PWC. In addition, link utilization is improved. In the future, we will investigate the feasibility of packet calking and the technologies suitable for its application. At the same time, we plan to focus on burst mode optical transmission at high bit rates, which is still a subject of intensive research.
Yuan Chi, Zhengbin Li, Anshi Xu
State Key Laboratory of Advanced Optical Communication
Systems and Networks