Buffering and Flow Control in Optical Switches for ... - Optics InfoBase

sierra.ece.ucdavis.edu

Buffering and Flow Control in Optical Switches for ... - Optics InfoBase

A60 J. OPT. COMMUN. NETW./VOL. 3, NO. 8/AUGUST 2011 Ye et al.2. We propose a two-level flow control scheme to preventbuffer overflow, resolve the data rate mismatching causedby output queuing, and prevent the congestion occurring atcertain ports from influencing other ports.3. We evaluate the performance of the hybrid optical switchunder both uniform random traffic and hotspot traffic viasimulation.The simulation shows that both the DLB and the MLB canprovide lower latency than can the SLB proposed in [2], whilethe MLB occupies fewer AWGR ports and uses fewer tunabletransmitters than does the DLB. The simulation results alsoconfirm that the proposed two-level flow control not onlyprevents packet drop, but also ensures that the congestionoccurring on the hotspot port does not impact the performanceon non-congested ports.The remainder of this paper is organized as follows.Section II reviews the related work. Section III presentsan overview of the architecture of the hybrid opticalswitch. Section IV discusses in detail the different loopbackbuffer architectures and associated flow control mechanisms.Section V evaluates the performance of the hybrid opticalswitch with the proposed two-level flow control. Section VIconcludes this paper.II. RELATED WORKConventional datacenter networks are built in a hierarchicalmanner, with a large number of cheap, low-speed, small-radixswitches at the bottom level to connect with the end nodes, anda few expensive, powerful, large-radix switches residing at thetop level to aggregate and distribute the traffic [3]. Recently,network architects have adopted fat tree and Clos topologies toprovide high aggregate bandwidth by constructing a switchingnetwork with small-radix switches [4,5]. Achieving low powerconsumption, low latency, and high throughput under highinput loads is the key challenge with electrical switches.Farrington et al. [6] and Wang et al. [7] suggested placingMEMS-based optical circuit switches in parallel with electricalswitches in the core network to carry slow changed inter-podtraffic [6] or latency insensitive traffic [7], thus reducingpower consumption and cost. Nevertheless, those designs donot address the challenges of achieving low latency and highthroughput under high input loads. Dragonfly [8] uses multiplesmall-radix routers to form large-radix switch groups, so thatthere is a connection between any two groups. AlthoughDragonfly can support high input load under uniform randomtraffic, it still saturates at moderate input loads when eachnode in one group sends traffic to a randomly selected nodein another group.Several efforts have been made to use optical technologiesin the network design for computing applications. Louri et al.proposed SOCNs/SYMNET [9,10], a multi-level hierarchicalarchitecture for a large-scale optical crossbar network anda tree-based address distribution sub-network. AlthoughSOCNs/SYMNET can connect to a large number of nodes,parallel communication is not utilized and the systemthroughput is limited, since the optical token controlledaddress broadcasting scheme allows for transmitting messagesonly serially. Gemini [11] is an optical/electrical dual banyannetwork. The optical banyan network delivers long messages,while the electrical network transmits control signals andshort messages. The drawbacks of Gemini are that the banyannetwork is a blocking network and that the optimal schedulingfor a large-scale banyan network is complicated.The OSMOSIS utilizes semiconductor optical amplifiers(SOAs) to realize a synchronous optical crossbar switchingfabric with the use of a broadcast-and-select data pathcombined with both space- and wavelength-division multiplexing[12,13]. Strictly speaking, the OSMOSIS still uses thestore-and-forward mechanism and adopts input and outputqueue structures, which are commonly used in electricalswitches. Although the optical switching fabric allows theOSMOSIS to have two receivers at each output, thussustaining high input load, the power requirements of theOSMOSIS can be very high because of its broadcast-and-selectarchitecture—signals are delivered to every select unit, eventhough only one unit selects the signal. The Data Vortex is adistributed interconnection network architecture [14,15] basedon deflection routing. To prevent packet drop when contentionoccurs, the packet is deflected to another output and an accesscontrol is adopted to ensure that the network will not carrytraffic beyond its capability. In other words, Data Vortex treatsthe deflection route as temporary network storage. Due to thedeflection routing and access control, Data Vortex saturatesbefore the offered load exceeds 50% [16]. In addition, as thenumber of nodes increases, the end-to-end latency becomeslarge and non-deterministic.AWGR-based optical switches and optical routers withpacket switching capability have been investigated for anumber of years. Previous work [17–23] mainly focused on theapplication of the AWGR in access networks and in telecommunication/IPnetworks. An AWGR serves as a non-blockingswitching fabric in many switch architecture designs. However,the wavelength parallelism on AWGR outputs is not exploredin those designs. Because each AWGR output is connectedwith a fixed wavelength converter to convert the signalto a particular wavelength in order to ensure wavelengthconsistency on the input and output fibers, the occurrenceof multiple packets is not allowed. Since no practical opticalbuffer is yet available, the store-and-forwarding scheme, whichis commonly used in the electrical switch, cannot be duplicatedin the optical domain. A fiber delay line (FDL) is commonlyused to resolve the contention, provide temporary storage, andallow packets that cannot gain the resource to compete forthe resource at a later time. Use of the FDL in resolvingthe contention helps to significantly reduce the droppingprobability, but packet loss is still possible and cannot beeliminated. In addition, the FDL cannot provide arbitrarydelays, which is more critical in asynchronous switching. Theresource may be available, but the delayed packet cannotaccess it, since the packet is still traveling through the FDL.The datacenter optical switch (DOS) [2] is an optical hybridswitch that adopts the AWGR as the switching fabric andutilizes wavelength parallelism to achieve output queuing. TheSLB with N parallel transmitters and N parallel receivers cansimultaneously receive contended packets from multiple inputports and transmit packets to multiple output ports whenresources are available. However, the SLB limits the scalability

More magazines by this user
Similar magazines