PDP43

Demonstration of 2.5 Gbps Optical Burst Switched WDM Rings Network Jaedon Kim, Jinwoo Cho, Mayank Jain David Gutierrez and Leonid G. Kazovsky

Ching-Fong Su, Richard Rabbat and Takeo Hamada

Stanford University 350 Serra Mall 058, Stanford, California, 93095

Fujitsu Labs of America 1240 East Argues Ave., M/S345 Sunnyvale, California, 94085

Abstract: We present a 2.5 Gbps OBS network testbed named Optical Burst Transport (OBT). OBT combines a reliable control channel with a tightly controlled high speed burst mode transmission. The result is verified by burst mode BER test. © 2006 Optical Society of America OCIS codes: (060.4250) Networks

1. Motivation Wavelength Division Multiplexing brings tremendous transmission capacity to the optical network. However, current optical data links in Metro Area Network (MAN) are mostly configured with static point to point connection, where a whole wavelength is dedicated to a pair of source and destination nodes. Therefore, despite the high potential that transmission capability offers, the current optical network has not fully utilized the potential to increase overall system throughput due to the rigid partition of wavelength resources. In particular, since traffic volume generated by applications varies over time and location, the provisioning of fixed-bandwidth circuits cannot accommodate an unpredictable traffic surge while also maintaining high bandwidth utilization. In this regard, the future metro optical network architecture should not only have a high speed transmission capability, but also an appropriate media access control (MAC) associated with the high speed link to enable efficient sharing of the transmission resources. In order to avoid expensive O-E-O conversion at the intermediate nodes and deal with increasing data traffic, several approaches have been introduced. Among those solutions, optical burst switching (OBS) is attractive in that it doesn’t require the use of optical buffers. Moreover, it transmits largesize frames at a time so that the synchronization requirement among the nodes is less stringent than that in the optical packet switched network. Hence, implementation complexity and cost can be lower than optical packet switched network while cost of operation is lower than static networks with the same transmission and switching capacity. Much research regarding the benefit of OBS has been conducted. However, little verification on the real testbed has been shown so far. We designed an OBS architecture on WDM ring, termed Optical Burst Transport, to support bursty data traffic transmission at 2.5 Gbps as well as sub-lambda bandwidth sharing. In this paper, a prototype testbed is constructed to demonstrate the performance of our design. 2.

Optical Burst Transport architecture

Figure 1. OBT node architecture

Figure 2. Testbed configuration

PDP43

Figure 1 shows the node configuration of OBT. OBT uses separate WDM channels for the data and control messages. A node can access the data channel by grabbing a token on the control channel. Once the node grabs a token, it sends a control header followed by the data frame for the destination. The source can decide on the number of destinations to which it will transmit data, but the maximum size of data frame should be bounded to keep to a maximum media access delay. While each data channel has its own controlling token, the token, like other control message, passes through the control channel. All control messages are added or dropped at every node through mux and demux. Data is added or dropped by the optical switch. The optical switch has 2 states. First, during the bar state, it separates the transceiver of the node from the ring so that the data frame destined to other nodes can transparently pass through the intermediate nodes. Second, when the node initiates transmission or receives data, the switch goes into the cross state for a specific amount of time. The duration of the cross state is decided as follows. For the transmission, the source decides the duration which is the same as the data transmission time; in addition the source sends a control header that includes the duration information. For the reception, the destination gets the duration information from the control header. Since a data frame for each destination is terminated at each destination, which is known as the destination stripping, the destination can initiate another sub transmission during which it receives the data frame from the source. Therefore, we can achieve spatial reuse to increase data channel efficiency while still having guaranteed maximum media access delay. We have reported the details of the spatial reuse property in another paper [1]. For the maximum transmission capacity, a node may have a dedicated transceiver and switch per data channel. However, using an Arrayed Wave Guide (AWG) and a coupler, a tunable transceiver can also be used to mitigate cost and bandwidth requirement. 3. Prototype Implementation A prototype testbed is constructed for demonstration and investigation as shown in figure 2. The control channel transceiver is not described in figure 2 to simplify the description. The testbed consists of Field Programmable Gate Arrays (FPGAs) and evaluation boards. The control channel signal is added or dropped through thin film filters at every node. A control header contains such information as source, destination and data burst length. In order to represent more than 1 megabyte data length information, the control header should have more than 23 bits for its data information sector. Hence, we assign 40 bits for the control message to configure the control message. The transmission speed of the control channel is set to 1.25 Gbps. and continuous mode transmission, while the data channel has 2.5 Gbps of speed and burst mode. In the data channel, the time spent on the reconfiguration of the data path would decide a guard time between control header and data burst, or between consecutive data bursts. As the data rate increase, the amount of wasted time slots for the guard time becomes severe. Therefore, to reduce the guard time, it is desirable to use a device with a very fast switching time. Using today’s technology, it is possible to get switches with a switching time is order of nanoseconds. Therefore, we can reconfigure the optical path very quickly to reduce the overhead. In the prototype, the guard time is the sum of control header processing time and switching time of the switch. In the prototype, the control message processing time is less than 5 clock cycles and the switching time of optical switch including rise / fall time of control signal is 165 nsec. Since the master reference clock of the system is 62.5 MHz, the signal processing time is 16×5 = 80 nsec. Therefore, we set 256 nsec (16 clocks) for the guard time. At the node, packets waiting for transmission are stored in two virtual output queues (VOQs) associated with their destination nodes on the ring. A traffic scheduler monitors VOQ lengths and allocates appropriate transmission windows for packets in different VOQs. Before sending out a burst targeting a destination node, the node sends a control header to inform the destination node to change its switch to receive that burst.

(a)

(b)

Figure 3. Testbed configuration. (a) Control signals shown on the logic analyzer. Signal receiving, sending, and bypassing are represented by “r”, “s”, and “b” respectively. (b) Data monitored at the 1st destination.

PDP43

Then, after a destination-dependent offset time, the node transmits the burst to the destination node during the allocated transmission window. During reception of the data, the destination node can also initiate the secondary transmission on the same data channel, since traffic from the source will not pass through the destination. Figure 3 shows one operational example. The operation of the control channel is monitored by the logic analyzer, and the data channel is monitored by the oscilloscope. In figure 3-(a), the source transmits 48 usec data burst to destinations 1 and 2. While destination 1 receives data for 32 usec, destination 2 receives data for 16 usec after corresponding transmission and propagation delay. Propagation delay from source to destination 1 is 125 usec and 200 usec from source to destination 2. One of the nodes in the ring network is responsible for initializing tokens upon system power-up. In addition, a timer in this node estimates the token round-trip time plus holding time by other nodes and triggers a regeneration process if it does not receive the token back within that specific period of time. For the data channel, we use a different size of frames because the guard time is too small to be shown in the same scale of control channel. In order to verify the data link performance, we did a burst mode bit error rate (BER) test as in the previous paper [2]. As described in figure 4-(a), the transmitter is directly modulated by a pattern generator which generates a continuous pseudo random bit sequence 2^23-1 at 2.5 Gbps. Whenever a token arrives at the node 1, it sends the control header and changes the switch to connect between transmitter and ring. When node 2 receives the control header, it changes the switch to receive data for a specific duration, and then switches back. During the time node 2 receives data, node2 also enables a gating signal for the bit error detector so that the error detector counts bit errors. Burst Continuous (Base line)

-5

Bit Error Rate (log)

-6 -7 -8 -9 -10 -11 -12 -19

-18

-17

-16

-15

-14

-13

-12

Received Optical Power (dBm)

(a)

(b) Figure 4. BER test (a) experimental setup (b) result

In the figure 4-(b), the circle indicates transmission without switching, which becomes continuous mode transmission. For the burst mode transmission, we send data for 100 usec to destination 1 and 50 usec to destination 2. BER is measured at destination 1. Compared to continuous mode transmission, burst mode transmission shows about 4.5 dB power gain at the same bit error rate. However, care should be taken in the interpretation. Noting the power meter takes the time average to calculate the received power, despite the fact that burst and continuous modes use the same power at the transmitter, the power meter shows a smaller value for the burst mode transmission, because the power meter receives nothing during the idle time. In our experiment, the round trip time for a token is 475 usec. The time during which data is transmitted to destination 1 is 100 usec. Therefore, the power meter would receive the real signal only for 100 usec every 475 usec and during the other 375 usec receive nothing. As a result, the power in the burst mode transmission is measured almost 100 / 475 = 0.315 times less than that of the continuous mode, which corresponds to about 5 dB. The interpretation shows that, in fact, two transmissions show similar result less than 1 dB power difference. The difference may come from the timing mismatch between the gating signal and real received data. However, the result confirms the fact that we can save transmission energy using the burst mode transmission. 4. Summary We demonstrate 2.5 Gbps OBS network testbed. Combined with a reliable control channel, a high speed data channel successfully operates in the burst mode. Also, using a fast optical switch, we can reduce the guard time overhead. The designed protocol and accompanying physical architecture allows much flexibility to shift between circuit-oriented and packet-switched traffic scenarios. The OBT architecture is thus a promising candidate for future WDM-based ring networks. Reference [1] Jaedon Kim, et al., “Spatial reuse on the optical burst transport network,” OFC, Anaheim, CA, OthF2, March 2006. [2] K. Shrikhande, et al., "Performance Demonstration of a Fast-Tunable Transmitter and Burst-Mode Packet Receiver for HORNET," OFC, Anaheim, CA, ThG2-1, March 2001.