ECE 526 – Network Processing Systems Design

1. ECE 526 – Network Processing Systems Design Switching Fabric D. E. Comer: chapter 10

2. Overview • Computation • Single CPU is not fast enough for processing packets • Multiple optimized processors (e.g. Network Processors) • Multiple ports with each port powered by NPs • Storage • Memory is bottleneck (memory wall) • On-chip, heterogeneous memories • Multi-threading to hide long memory access • Communication • Bus: data, control and address • Memory: can also used for communication ECE 526

3. Port Network Processor e Processing Processing c a Engine Engine f r e t n Interconnect I k Processing r o Processing Engine w Engine t e N I / O Router Data Path • How can ports be connected? • “System backplane” • Switching Fabric: • Hardware mechanism serving as backplane by providing path for data flow among the ports ECE 526

4. Switching Fabric Requirement • Connectivity: • interconnection between ports and control processor • High throughput • Low cost • Scalability • Scale with data rate on input/output (one port) • Scale with packet rate on input/output (one port) • Scale with number of input/output ports • Feature • Support transfer of unicast, multicast, and broadcast packets Conflicting requirements: tradeoffs necessary ECE 526

5. Switching Fabric Types • Synchronous vs. asynchronous • Sync: regular traffic (e.g., fixed cells); async: variable size (e.g., packets) • In practice synchronous fabrics are used • Packets are divided into fixed-sized cells • Multiplexing: • Time Division: single or few paths shared among many ports • Space Division: many paths used for less delay • Stage: • Single or multiple ECE 526

6. Switching Fabric Design Spectrum • Design spectrum • Evaluation criteria • Low cost • Performance • Scalability • Design option • Data transferred granularity • Queuing for contention ECE 526

7. Shared Bus Fabric • Disadvantage: lower aggregate data rate–Hardware needs to operate at N times faster than ports • What granularity is best for access: packet, cells? ECE 526

8. Granularity for Bus • Packet • Pro.: Simple hardware • Con.: delay- small packet suffers waiting for large pkt • Fixed size block • Pro.: multiple transfer concurrently • Con.: overhead from switching senders ECE 526

9. Shared Memory Fabric • Input ports deposit packets in memory • Problem: cost of memory interfaces ECE 526

10. Fully Interconnected Fabric • Separate physical data path between each port pair • Advantage and disadvantage? ECE 526

11. Fully Interconnected Fabric • Advantage • possible N concurrently transfers • Disadvantage • Port contention: two input port to the same output simultaneously • worst case: only one transfer • Expensive: • N*M wires • Output port needs contention circuitry • Question: is it possible to N concurrently transfers but not N*M wires • Observation: for each output port, only one of N wires is used at anytime ECE 526

12. Crossbar Fabric • Switched interconnect • Allows multiple Simultaneous connections • Controller decides on assignment of active connections • Pros and cons? ECE 526

13. Crossbar Fabric • Pros: • Using N + M wires to achieve maximum simultaneous transfers up to Min(N,M) • High aggregate data rate • Cons: N * M switches ECE 526

14. Contention • What happens if multiple input ports send to the same output port? • Causes “port contention” • Is this a problem of the fabric architecture? • How can contention be addressed? • Queues that buffer packets ECE 526

15. Input Queueing • Incoming packet traffic is burst • Head of line blocking • The first packet in the queue to output port s, which is busy • The second packet in the queue to output port t, which is idle • Solution for head of line blocking • Multiple queues for each port • Change queue from FCFS to randomly access • Virtual queueing: • based on estimate of processing time to reserve packet position in the queue • requires complex coordination • What is right size of input queue • Upper bound: port rate ECE 526

16. Output Queueing • Bursts from inputs to one output are buffered on output side • Queue size • Too small -> packet dropped • Too large -> keeping packet are longer valid • Voice packet arrived to destination too late • TCP packet already retransmitted ECE 526

17. Queuing Summary • Variety of queuing designs possible • Input queuing: • Packets are buffered on input side until output is available • Problem: head of line blocking • Output queuing: • Bursts from inputs to one output are buffered on output side • Problem: queue size, switch fabric speedup • Virtual queuing: • Partial output queues are maintained on input size • Each port gets rate or virtual time assigned for sending • Problem: requires complex coordination ECE 526

18. Multistage Fabrics • None of the fabrics so far really scales • Full interconnect: squared cost with port number • Crossbar: limitations in controller • Bus: interconnect needs to be N times faster than port • Shared Memory: memory interfaces are costly and not scalable • Multistage fabrics: • Multiple “steps” between input and output • Provides multiple data paths, but not as many as fully connected • Packets can be queued or not on each stage • Nice properties: • Scalability • Self routing ECE 526

19. Switching Elements • Switching element is basic component: • Two inputs, two outputs • Two different settings • Straight through • Crossover • Setting is determined by “label” • Multiple switching elements make up larger fabric • Various topologies possible: • Banyan • Benes • Delta • Etc. ECE 526

20. Banyan Switch • From 2-port switch to 4-port switch: • Can be applied recursively ECE 526

21. Banyan Switch • 8-port: • Recursive extension • Self-routing: • Based on output port “label” each stage can make local decision • Recursive structure will lead to correct destination ECE 526

22. Nice Feature of Banyan • Scalability • Can be operated at lower rate than a shared bus • Applying to arbitrary number of input ports • Modular design: cheap 2-input switch • Self-routing • Each output port including internal port is identified by a unique binary label • Sender prefixes the appropriate label on the packet • At each stage, switch extract the next bit and use it to determines the path • Internal packet queues allowed but not required • With queues will increase throughput, packet will move forward as long as it allowed • Without queues is cheaper, path established first, then send it ECE 526

23. Banyan Switch • Scalability due to recursive structure • Contention can occur • How? • Usually due to overloading of one output • Other multistage switch fabrics: • Different recursive rules • Additional stages: distribution to avoid blocking • Practical note: • Standards define Switch Fabric Interfaces ECE 526

24. Exercises 1 • Draw a banyan network for 8 input with full details, and label intermediate connection with label prefix to reach to the connection. ECE 526

25. Exercise 2 • Derive the number of stage required in a Banyan switch with 2n inputs. How many of 2-input switch required? How many queues required if queues allowed? ECE 526

26. Exercises 3 • What is the maximum number of transfers that can occur in a 8-input Banyan switch in the best case? Does the answer changes with or without buffers? (transfers defines as a connection from input of switch to its output) • What is the maximum number of transfers that can occur in a 8-input Banyan switch in the worst case? Does the answer changes with or without buffers? ECE 526

27. Summary • Switching fabric provides connections inside single network system • Two basic approaches – Time-division has lowest cost – Space-division has highest performance • Multistage designs compromise between two • Banyan fabric is example of multistage

28. ECE 526