NGR: Next Generation Router Initial Design Review Shahzad Ali Xia Chen Brendan Howell Yu Zhong Specifications Minimum 64 port OC-48, aggregate capacity of 160Gbps, Fair Queuing. Extra features RED MPLS, QoS Multicast Scalability (higher port density) 7/24/2016 18-757 Initial Design Review 2 Overall Design 16 Cards Switch One port on a Line card Transceiver Framer 7/24/2016 Cell buffer Network processor Scheduler + FQ Routing table 18-757 Initial Design Review 3 Components Input Port processing (IPP) Buffers Switch Scheduling Output port processing Fair Queuing 7/24/2016 18-757 Initial Design Review 4 Input Port Processing Line termination and data link processing Break variable length packet into fixed length Lookup routing information Process packets at "line speed” i.e. close to 5 million routing lookup per second. Do simple IP packet processing functions Decrease TTL, Checksum etc. MPLS or Hardware Filters 7/24/2016 18-757 Initial Design Review 5 Possible Solutions for IPP Transceiver for line termination and PHY processing HFCT-5402D by Agilent (HP) VSC8140 by Vitesse Framer for fixed size cells VSC9112 by Vitesse Network Processor IXP1200 by Intel (Level One) APP1200 and APP1400 by Agere Generic micro-processor 7/24/2016 18-757 Initial Design Review 6 Where to Queue? Input port Does not require high fabric speedup. OPP can be done at reasonable speeds Difficult to do output scheduling (FQ) Output port Easy to schedule packets on the link (FQ) Fabric and OPP have to have high speedup. 7/24/2016 18-757 Initial Design Review 7 Buffers Our initial decision: Buffer packets on the input port Handle FQ at the input port (somehow) Use standard SDRAM for buffers PC-100 10ns access latency Dimension buffer by simulation Buffers for each input port. 7/24/2016 18-757 Initial Design Review 8 Switch Scheduling We will simulate with various scheduling mechanisms including Multicast schemes. iSLIP, RR, FCFS, TARTAR (for multicast) Virtual Output Queues to avoid HOL Nominal Fabric speedup and priority classes. Implement both Drop-tail and RED buffer schemes. Local delivery of packets on the same line card. 7/24/2016 18-757 Initial Design Review 9 Switch Fabric Shoot for 1024x1024 but backup is 64x64 Vitesse VSC836 takes too long to configure!! 13cycles x 40ns = 520ns (1.9M Frames/sec) Use AMCC S2018 17x17 in a 3-stage CLOS network. Other options? 7/24/2016 18-757 Initial Design Review 10 Output Port Processing and Scheduling Basic de-framing and physical level processing. FQ discipline WFQ W2FQ How do we do this with input buffering? What level of buffering do we require at the output ports? 7/24/2016 18-757 Initial Design Review 11 Plan of Action Decide on IPP and Network Processor. Concentrate on switch fabric choice. Analyze switch scheduling schemes. Try to implement FQ with input port buffers. Implement RED and Drop Tail. Develop a simulator for the switch. Experimentally determine the buffer size. 7/24/2016 18-757 Initial Design Review 12