1. The Crosspoint Queued Switch Yossi Kanizo (Technion, Israel) Joint work with Isaac Keslassy (Technion, Israel) and David Hay (Politecnico di Torino, Italy)

2. Typical Switch Architectures IQ – Input Queued Linecards Switch Fabric CICQ – Combined Input and Crosspoint Queued Linecards Assumes Instantaneous Closed Loop

3. Single-Rack Router  Instantaneous closed loop → works in a single rack  Problem: multi-rack routers Linecards Switch Fabric

4. Current Router Architectures [Source: N. McKeown] Is the closed loop still instantaneous?

5. Time Trends ns

6. Hiding Propagation Delays  Traditional solutions:  Increase time-slot  poor switch performance  Hide propagation delays using buffers  impractical amount of buffering  Proposed solution: closed loop → open loop  Performance degradation vs. instantaneous closed loop

7. Outline  CQ: Open-loop switch architecture  Performance Evaluation  Analytical results  Simulations  CQ performance degradation is not significant

8. Proposed Architecture: The Crosspoint-Queued (CQ) Switch  No queues in the linecards  Buffering only inside the fabric  Independent output schedulers  Drops with full buffers Switch Core Linecards 10s of meters

9. CQ Properties  Open loop  No communication overhead  No linecard queues  No linecard queue management  “Router on a chip”  Buffering and switch fabric on same chip

10. Why not 10 years ago?  No need: single rack  No technology: SRAM density  Moore’s law: density doubling every 2.5 years  Aggressive 128x128 CQ switch: 4 cells of 64 bytes per crosspoint → 64 cells today  Conservative buffer requirements  TCP Stanford model with smaller buffer needs [Appenzeller, Keslassy and McKeown ’04]

11. Outline  CQ: Our open-loop switch architecture  Performance Evaluation  Analytical results  Simulations

12. 100% Throughput as B→  Throughput bounds: OQ(2B-1) ≤ CQ(B)≤ OQ(NB) Buffer size B, LQF scheduling algorithm 100% Throughput ∞

13. Uniform Traffic, B=1  Uniform traffic model:  At each time-slot, at each of the N inputs: Bernoulli IID packet arrivals with probability   Each packet is destined for one of the N outputs uniformly at random  Theorem: Under uniform traffic and B=1, the performance of the switch is independent of the specific work-conserving scheduling algorithm  Intuition: Symmetry

14. Uniform Traffic, B=1  Theorem: The throughput and waiting time of a CQ switch, B=1 is:  Proof: Based on Z-transform q=1-  /N Goes to 100% as N goes to infinity

15. Models for larger buffers  Approximate Performance Analysis  Model for exhaustive round-robin scheduling  Based on modifications to polling system with zero switch-over times  Model for random scheduling algorithm  Show 100% throughput as N→∞

16. Trace-Driven Simulation Buffers of size 64 suffice to ensure 99% throughput for N=32. 32x32 CQ switch with different buffer sizes (in units of 64-byte packets)

17. Conclusions  CQ is open loop → allows multi-rack configuration  CQ provides easy scheduling  CQ is feasible to implement in a single chip  CQ shows good performance in simulations

18. Thank You