Characterization of A Shared Buffer Optoelectronic Packet Router †Shunyuan Ye, ‡Marina Thottan, ‡Jesse E. Simsarian, †Shivendra Panwar †Department of ECE, Polytechnic Institute of NYU ‡Bell Laboratories, Alcatel-Lucent e-mail: sye02@students.poly.edu, {marina.thottan, jesse.simsarian}@alcatel-lucent.com, panwar@catt.poly.edu Abstract—The rapid increase in Internet traffic is forcing packet routers to grow in capacity to meet the demand. Optical packet routers with less buffering and a greater degree of optical transparency are actively being researched as a way to improve energy efficiency and capacity scaling over traditional electronic routers. Since it is difficult to buffer packets in the optical domain, in this paper we analyze the performance of a hybrid optoelectronic packet router. The router architecture has multiple optical switch planes and a shared electronic buffer to resolve output-port contention. By using multiple ports on the switch planes for each input and output fiber, and by using some switch- plane ports to inter-connect the planes, we can achieve a relatively low packet loss ratio in a router with no buffer. In this case, most traffic can be switched using only the through optical paths of the router without entering the shared buffer. The shared electronic buffer is primarily used to reduce the packet drop ratio under periods of heavy loads and occasionally for optical regeneration of a packet. We run extensive simulations to evaluate the performance of the router with varying number of switch plane ports, number of connections to the electronic buffer, and number of interconnections between the switch planes. We show that the router can provide good throughput, with realistic on-off bursty traffic and asynchronous packet arrivals. I. I NTRODUCTION Internet traffic has been rapidly increasing thereby driving research in energy-efficient routing technologies to meet band- width demands without large increases in the power consump- tion of networking equipment. While there has been substantial research in all-optical routers, they remain impractical due to the difficulty of buffering packets optically. Devices like fiber delay lines (FDLs) have been used to buffer optical packets by delaying optical signals for a specific amount of time that is proportional to the length of the FDL. However, FDLs are bulky and inflexible in their buffering capabilities. Moreover, it is inefficient to use FDLs with asynchronous packet arivals. Therefore, researchers have pursued hybrid optoelectronic architectures [1]–[7] that exploit advantages of both electronic and photonic technologies: packets are switched over an optical fabric, and electronic buffers are used to resolve output port contentions when needed. The OSMOSIS (Optical Shared Memory Supercomputer Interconnect System) project [2], [3] at IBM adopted an input-queued architecture, in which every packet arriving at the switch is buffered by virtual output queues (VOQs). To eliminate the head-of-line blocking problem, each input has to maintain N VOQs, where N is the number of ports. Packets are then switched over an optical fabric. Optical-to-electrical !"#$%& ()*+$,*-. /%01*$ 2 3-"4+5 !4+"4+5 2 67$7*871 91%-5:*;71 2 2 Fig. 1: Model of an optoelectronic shared-buffer router archi- tecture and electrical-to-optical conversions are required at the inputs and outputs of the switch, respectively, and known scheduling algorithms for input-queued switches, i.e. maximum weight matching (MWM) [8], [9] and iSLIP [10], can be directly applied. This architecture requires every packet to be buffered, even when there is no packet contention. To reduce electronic buffering, a shared-buffer architecture, as shown in Fig. 1, has been proposed by many researchers [1], [4]–[7]. Different from an input-queued (IQ) or an output- queued (OQ) switch, an electronic buffer is placed in the loopback path to resolve output-port contentions. This archi- tecture has been used in all-optical switch designs [11], [12], where the buffer is implemented by FDLs. By introducing the loopback path, a packet can either go to the output directly, or first be sent to the buffer and then back to the fabric to be switched to the output. For hybrid shared-buffer architectures, where the buffer is electronic, packets are only sent to the buffer either when there is output-port contention or for packet regeneration, thus reducing power consumption and packet queuing delay as compared to switches that electronically buffer every packet. This work uses the hybrid shared-buffer architecture of Ref. [1] that has some of the features shown in Fig. 1 but with some modifications: As with typical optical transmission systems, each incoming fiber to the router uses wavelength- division multiplexing (WDM) for increased capacity. Also, instead of using one single large switch fabric to route the packets, we first de-multiplex the incoming wavelengths and