applied research laboratory 1 Scaling Internet Routers Using Optics Isaac Keslassy, et al. Proceedings of SIGCOMM Slides:
applied research laboratory 2 Do we need faster routers? Traffic still growing 2x every year Router capacity growing 2x every 18 months By 2015, there will be a 16x disparity –16 times the number of routers –16 times the space –256 times the power –100 times the cost => Necessity for faster, cost effective, space and power efficient routers. Source: Dr. Nick McKeown’s SIGCOMM talk
applied research laboratory 3 Current router : Juniper T640 T640: Half-rack –37.45 x x 31 in (H x W x D) –95.12 x x cms (area ≈ 3 m 2 ) –32 interface card slots –640 Gbps front side switching capacity –6500 W power dissipation –Black body radiation = T 4 W/m 2 –at 350 F, Power radiated = 2325 W/m 2 –Operating temp. = 32 to 104 F = 0 to 40 C = Stefan Boltzmann constant = * W / m 2 K 4 References: – 670_RSP.jhtmlhttp:// 670_RSP.jhtml – –
applied research laboratory 4 Multi-rack routers Switch fabric and linecards on separate racks Problem: Switch fabric power density is limiting –Limit = 2.5 Tbps (scheduler, opto-electronic conversion, other electronics) Switch fabric can be single stage or multi stage –Single stage: complexity of arbitration algorithms –Multi-stage: unpredictable performance (unknown throughput guarantees) Switch fabric Linecards
applied research laboratory 5 Optical switch fabric Pluses –huge capacity –bit rate independent –low power Minuses –slow to configure (MEMS ≈ 10 ms) –fast switching fabrics based on tunable lasers are expensive Reference: –
applied research laboratory 6 Goals Identify architectures with predictable throughput and scalable capacity –Use the load balanced switch described by C-S. Chang –Find practical solutions to the problems with the switch when used in a realistic setting Use optics with negligible power consumption to build higher capacity single rack switch fabrics (100 Tbps) Design a practical 100 Tbps switch with 640 linecards each supporting 160 Gbps
applied research laboratory 7 Load balanced switch 100 % throughput for a broad class of traffic No scheduler => scalable VOQ
applied research laboratory 8 Problems with load-balanced switch Packets can be mis-sequenced Pathological traffic patterns can make throughput arbitrarily small Does not work when some of the linecards are not present or are have failed Requires two crossbars that are difficult or expensive to implement using optical switches
applied research laboratory 9 Linecard block diagram Both input and output blocks in one linecard Intermediate input block for the second stage in the load balanced switch
applied research laboratory 10 Switch reconfigurations The crossbars in the load balanced switch can be replaced with a fixed mesh of N 2 links each of rate R/N The two meshes can be replaced with a single mesh carrying twice the capacity (with packets traversing the fabric twice) RR/N R R2R/N R
applied research laboratory 11 Optical switch fabric with AWGRs AWGR: data-rate independent passive optical device that consumes no power Each wavelength operates at rate 2R/N Reduces the amount of fiber required in the mesh (N 2 ) N = 64 is feasible but N = 640 is not AWGR = Arrayed Wavelength Grating Router
applied research laboratory 12 Decomposing the mesh 2R/ Source: Dr. Nick McKeown’s SIGCOMM slides
applied research laboratory 13 Decomposing the mesh 2R/4 2R/ TDM WDM Source: Dr. Nick McKeown’s SIGCOMM slides
applied research laboratory 14 Full Ordered Frames First (FOFF) Every N time slots –Select a queue to serve in round robin order that holds more than N packets –If no queue has N packets, pick a non-empty queue in round robin order –Serve this queue for the next N time slots N FIFO queues (one per output) input To intermediate input block
applied research laboratory 15 FOFF properties No Mis-sequencing –Bounds the amount of mis-sequencing inside the switch –Resequencing buffer at most N packets FOFF guarantees 100 % throughput for any traffic pattern Practical to implement –Each stage has N queues, first and last stages hold N 2 +1 packets/linecard –Decentralized and does not need complex scheduling Priorities are easy to implement using kN queues at each linecard to support k priority levels
applied research laboratory 16 Flexible linecard placement When second linecard fails, links between first and second linecards have to support a rate of 2R/2 Switch fabric must be able to interconnect linecards over a range of rates from 2R/N to R => Not practical 2R/3
applied research laboratory 17 Partitioned switch M input/output channels for each linecard Theorems: 1)M = L+G-1, each path supporting a rate of 2R 2)Polynomial time reconfiguration when new linecards are added or removed.
applied research laboratory 18 M = L + G -1 illustration Total traffic going out or coming in at Group 1 = LR Total number of linecards = L + G -1 Number of extra paths needed to/from first group = L -1 LC 1 LC 2 LC L Group 1 LC 1 Group 2 Group G LC 1 LC 2 LC L Group 1 LC 1 Group 2 Group G
applied research laboratory 19 Hybrid electro-optical switch
applied research laboratory 20 Optical Switch
applied research laboratory Tb/s Load-Balanced Router L = Gb/s linecards Linecard Rack G = 40 L = Gb/s linecards Linecard Rack 1 L = Gb/s linecards x 40 MEMS Switch Rack < 100W Source: Dr. Nick McKeown’s SIGCOMM slides