Revisiting router architectures with Zipf

Transcription

1 Revisiting router architectures with Zipf Steve Uhlig Deutsche Telekom Laboratories/TU Berlin Nadi Sarrar, Anja Feldmann Deutsche Telekom Laboratories/TU Berlin Rob Sherwood, Xin Huang Deutsche Telekom Labs USA Social Networks and Future Internet, 1 Annecy, France, June 27-28, 2011

2 Agenda Motivation Revisiting IP routers architecture Opportunities and challenges Evaluation Conclusion

3 Software-defined networking Beyond today s monolithic network equipment Separation of control and data plane through software modularity, e.g., Linux Communication channel Remote controller Do not change existing control plane Control plane Principles - Communication channel between forwarding engine and remote controller 3 Forwarding engine Software Hardware - Expose network equipment capabilities, e.g., TCAM, QoS

4 Agenda Motivation Revisiting IP routers architecture Opportunities and challenges Evaluation Conclusion

5 OpenFlow-based IP router System combines software router with fast switching hardware - Fast switch handles most of the traffic with a few entries, i.e., fast path - Software handles control plane and remaining traffic, i.e., slow path Communication channel Switch controller Software router Slow path fibvisor Linux FIB Our approach: - Take advantage of traffic properties Traffic 4 Flow table Fast path

6 Agenda Motivation Revisiting IP routers architecture Opportunities and challenges Evaluation Conclusion

7 Zipf in Internet traffic Data - Transcontinental link: 150Mbps backbone link (MAWI), 3.5 days - Residential ISP: 1Gbps link, 2 days - IXP: > 1Tbps, 4 days Observation: Most traffic captured by limited number of prefixes Cumulative fraction of traffic ISP MAWI IXP k 10k 100k 1M Prefixes ordered by decreasing traffic volume Opportunity: In principle, existing switching hardware can do it 7

8 Slow path Assume knowledge of the future traffic Slow path rate as a function of number of heavy-hitters A few thousand heavy hitters enough to keep slow path rate low Limited variations across traces Controller traffic load (PPS) ISP (median) ISP (max) MAWI (median) MAWI (max) IXP (median) IXP (max) Number of heavy hitters Opportunity: with a few thousand flows, slow path rate can be kept 8 low

9 Churn Assume knowledge of the future traffic What is the best-case churn rate? Proportional to number of heavy-hitters Changes per time bin (median) ISP MAWI IXP Number of heavy hitters Challenge: keep churn low while keeping most of the traffic on the fast path 9

10 Agenda Motivation Revisiting IP routers architecture Opportunities and challenges Evaluation Conclusion

11 Traffic Offloading Tame natural churn of heavyhitters Traffic Offloading (TFO) Algorithm 1 Round-robin select popular entries next current FIB cache entries FIB cache changes to be applied - Monitor traffic at multiple time-scales 2 Compute difference 4 Select high-value FIB cache changes - Select heavy-hitters that are expected to lead to low churn top 10sec 1min 10min additions removals top additions top removals - Trade-off offloading gain with churn Full FIB sorted by prefix popularity 3 Sort by popularity 11

12 Churn Traditional caching - Always replace entry upon miss - Leads to high churn for low number of heavy-hitters TFO keeps churn much lower than bin-optimal and caching When number of heavy-hitters high, combination of caching and TFO is ideal Changes per second (median) Number of heavy hitters LFU LRU Bin optimal (10s bin) TFO 12

13 Slow path LFU shows importance of heavy-hitters dynamics over short time-scales LRU and TFO close to optimal Slow path rate low for a few thousand heavy-hitters Controller traffic load (PPS) LFU LRU FIBpredict Optimal (10s bin) Number of heavy hitters 13

14 TFO: churn Churn depends on traffic aggregation - IXP: a few changes per second - ISP: 10 s of changes per second - Transcontinental link: up to 100 of changes per second Changes per second (median) ISP MAWI IXP TFO tames the churn Number of heavy hitters Feasible on today s OpenFlow-enabled switches 14

15 TFO: slow path Load can be handled by commodity PC Could be done on better embedded switch CPU Scaling up Routebricks Packetshader Traditional router Controller traffic load (PPS) ISP (median) ISP (max) MAWI (median) MAWI (max) IXP (median) IXP (max)... Number of heavy hitters Feasible on today s commodity hardware

16 Agenda Motivation Revisiting IP routers architecture Opportunities and challenges Evaluation Conclusion

17 Conclusion Revisiting router architecture through SDN - Leverage traffic properties (Zipf) - Combine open-source routing with fast and cheap switches TFO algorithm - Beyond traditional caching: carefully select the right heavy-hitters - Keep both churn and slow path rates low Scale up routers: fast switching hardware + fast software routers