1. Designing Networks with Little or No Buffers or Can Gulliver Survive in Lilliput? Yashar Ganjali High Performance Networking Group Stanford University yganjali@stanford.edu http://www.stanford.edu/~yganjali Joint work with Prof. Ashish Goel, Prof. Nick McKeown, Prof. Tim Roughgarden October 21, 2004 - SNRC

2. 2 Motivation Networks with Little or No Buffers  Problem  Internet traffic is doubled every year  Disparity between traffic and router growth (space, power, cost)  Possible Solution  All-Optical Networking  Consequences  Large capacity  large traffic  Little or no buffers  Gulliver wakes up in Lilliput!

3. October 21, 2004 - SNRC 3  Why do we need buffers?  How large are buffers today?  How small can buffers be?  Summary and conclusions Outline

4. October 21, 2004 - SNRC 4  Why do we need buffers?  Pathological example  Contention  Congestion  How large are buffers today?  How small can buffers be?  Summary and conclusions Outline

5. October 21, 2004 - SNRC 5 Pathological Example  If S 1 sends a packet at time t, S 2 cannot send any packets at time t, and t+1.  To achieve 100% throughput we need at least one unit of buffering. 12 S1S1 S2S2 D 34  Flow 1: S 1  D; Load = 50%  Flow 2: S 2  D; Load = 50%

6. October 21, 2004 - SNRC 6 Why do we need buffers?  Internet traffic is bursty in nature  Starting time  Flow length variations  Rate changes over time  We need buffers to compensate momentary over-subscriptions  Short-term: Contention  Long-term: Congestion

7. October 21, 2004 - SNRC 7 Contention  Even when links are under-subscribed contention occurs due to stochastic collision of packets.

8. October 21, 2004 - SNRC 8 Congestion  Internet uses distributed congestion control.  Links are temporarily over-subscribed.  TCP needs buffers to work.

9. October 21, 2004 - SNRC 9 Rule for adjusting W  If an ACK is received:W ← W+1/W  If a packet is lost:W ← W/2 Congestion Only W packets may be outstanding

10. October 21, 2004 - SNRC 10 Congestion Rule for adjusting W  If an ACK is received:W ← W+1/W  If a packet is lost:W ← W/2 SourceDest t Window size Only W packets may be outstanding

11. October 21, 2004 - SNRC 11  Why do we need buffers?  How large are buffers today?  Congestion control  Contention resolution  How small can buffers be?  Summary and conclusions Outline

12. October 21, 2004 - SNRC 12  Universally applied rule-of-thumb:  A router needs a buffer size: 2T is the two-way propagation delay C is capacity of bottleneck link  The buffer needed for contention resolution is dominated by this. Router Buffer Needed for Congestion Control C Router Source Destination 2T

13. October 21, 2004 - SNRC 13 Example  Link capacity 40Gbps  Two way propagation delay 100ms  Required buffer size 500MB  2,000,000 packets  Hard to implement even in electronics

14. October 21, 2004 - SNRC 14  Why do we need buffers?  How large are buffers today?  How small can buffers be?  Congestion control Aggregate flows No buffering  Contention resolution Scheduling Load balancing in time and space  Summary and conclusions Outline

15. October 21, 2004 - SNRC 15 Buffer Size in the Core Probability Distribution B 0 Buffer Size

16. October 21, 2004 - SNRC 16  The rule of thumb is true for one flow.  Thousands of flows in the core of the Internet.  Required buffer is instead of  n is the number of flows  [Appenzeller, Keslassy, McKeown 2004]  Required buffer: 20,000 packets  Much less than 2,000,000 packets  Still not feasible in optics Required Buffer Size

17. October 21, 2004 - SNRC 17 TCP with ALMOST No Buffers Utilization of bottleneck link = 75%

18. October 21, 2004 - SNRC 18 TCP with Almost No Buffers  With almost no buffering and just a single flow we lose about 25% of the capacity.  Capacity is not a bottleneck anymore.  Small buffers not a major problem for congestion control.

19. October 21, 2004 - SNRC 19 Throughput with Little Buffers Capacity: 100Mbps, Buffer size: 10 packets

20. October 21, 2004 - SNRC 20  Why do we need buffers?  How large are buffers today?  How small can buffers be?  Congestion control Aggregate flows No buffering  Contention resolution Scheduling Load balancing in time and space  Summary and conclusions Outline

21. October 21, 2004 - SNRC 21 Contention Resolution  Two extreme approaches  No scheduling Use large buffers  Tight scheduling Inject packets according to a schedule, so that contention is minimized. No schedulingTight scheduling

22. October 21, 2004 - SNRC 22 Exact Schedule  Hard to find in general  Must be distributed  Exact measurements of flows needed  Extreme synchronization required

23. October 21, 2004 - SNRC 23 Randomization  Randomization  Randomly delay packet at injection time Imitate a Poisson process  Reshape traffic at the core Keep Poisson No schedulingTight scheduling ??? Randomization

24. October 21, 2004 - SNRC 24 Randomization (cont’d) 2134567 21345 Time Input #1 Input #2 6 Before randomization 4123567 21634 Time Input #1 Input #2 5 After randomization

25. October 21, 2004 - SNRC 25 Randomization (cont’d)  Mimic an M/M/1 queue  Under low load, queue size is small with high probability  Loss can be bounded  1 1 M/M/1 X b P(Q > b) Buffer B Packet Loss

26. October 21, 2004 - SNRC 26 Example 16 Node Network, Tree Shaped Topology, Load = 70%, Loss = 1%

27. October 21, 2004 - SNRC 27 Implications of Randomization  Alleviates short-term contention  Simple to implement  Guaranteed performance

28. October 21, 2004 - SNRC 28 Preliminary Results Theorem. We can achieve constant factor throughput (roughly 70-80%) with very small amount of loss using buffers of size O(log L), where L is the length of the longest path in the network.  Assumption: No reactive flow control  Typical value of L is 3 to 5

29. October 21, 2004 - SNRC 29  Why do we need buffers?  How large are buffers today?  How small can buffers be?  Congestion control Aggregate flows No buffering  Contention resolution Scheduling Load balancing in time and space  Summary and conclusions Outline

30. October 21, 2004 - SNRC 30 Preliminary Results Conjecture. Routers in the core of the current Internet only need buffers of size O(log L) instead of the 2TxC.  Justification  Currently maximum window size is 64 and 12 packets for Linux and Windows XP  Access links are the bottleneck and limit the flow  Randomization takes care of momentary collisions

31. October 21, 2004 - SNRC 31 Implications & Considerations  The required buffer size does not depend on the link capacity.  All we need is 10-20 packets worth of buffering (instead of thousands)  Need to study  The interaction between randomization and congestion control  Impact of co-existing flows  Emerging applications (non-TCP or modified TCP) which allow much large windows per flow

32. October 21, 2004 - SNRC 32 Summary and conclusions  We need buffers.  To gain 100% we need relatively LARGE buffers  Not as large as the widely used rule of thumb  At the price of losing some capacity, TCP performs well even with small buffers.  We can address contention by adding randomization.  Gulliver might not have the best life in Lilliput, but he seems to be able to survive! Many thanks to Guido Appenzeller for providing the flash animations.