Improve network efficiency by 1000 times! Energy-Efficient Congestion Control Opportunistically reduce link capacity to save energy Lingwen Gan1, Anwar Walid2, Steven Low1 1Caltech, 2Bell Labs Network links consume a lot of electricity Electricity consumption of network links > Electricity consumption of the United Kingdom Fiber optics, copper cable yearly growth rate 20% 15% 10% 5% 0% network link electricity consumption world electricity consumption Reduce electricity consumption of network links. Exploit low link utilization link utilization 40% 30% 20% 10% 0% on average off-peak What we do: dynamically manage link capacity. Technologies to change link capacity Sleep mode [Gupta’03] Voltage and frequency speed scaling [Pillai’01] Link bundle Link bundle router .. . 2~20 router component link to sleep Linear power consumption Identical component links 5 4 Power consumption (units) 3 2 1 0 0 1 2 3 4 # active component links energy saving reduced capacity 5 Outline • • • • Challenge Goals Algorithm Simulations Challenge: interaction with TCP capacity throughput congestion TCP reacts Reduce traffic throughput Two approaches Adjust capacity slowly. Routing time scale. [He’06] [Fisher’10] Adjust capacity fast, but TCP friendly. Packet time scale. [Francini’10]… Flow time scale. • Fast response • Small overhead This work Goals Dynamic Bandwidth Adjustment (DBA) Algorithm, such that 1) Operate at flow time scale. 2) Do not reduce throughput. 3) Save as much energy as possible. 4) Throughput does not oscillate---stability. Recall TCP transmission rate TCP packet loss probability at steady state transmission rate packet loss probability Recall Random Early Discard (RED) link link capacity incoming traffic buffer size packet drop probability buffer size Recall network solves NUM Thm [Kelly’98, Low’99]: The network model solves the Network Utility Maximization problem: Transmission rates Ideal throughput Throughput on the links Ideal capacity Bottleneck & non-bottleneck links Bottleneck link: • Do not reduce capacity packet drop probability buffer size Non-bottleneck link: • Reduce capacity • Keep 0 packet drop packet drop probability buffer size Keep the buffer at the “right” place packet drop probability target buffer buffer size DBA Algorithm (for each link) 1. Pick a target delay satisfying 2. At any time, set target buffer size and update capacity as capacity current buffer size zero throughput reduction & maximum energy saving Thm: Network under DBA algorithm, modeled by Current network architecture converges to (original) target throughput (zero throughput reduction) with minimum energy consumption (maximum energy saving) Model network delay transmission rate TCP sources packet loss No network delay Global stability incoming traffic Links packet drop With network delay delay Local stability under network delay Thm: Network (with DBA) is locally asymptotically stable, in the presence of network delay modeled as provided some mild conditions hold. Goals Dynamic Bandwidth Adjustment (DBA) Algorithm, such that 1) Operate at flow time scale. 2) Do not reduce throughput. 3) Save as much energy as possible. 4) Throughput does not oscillate---stability. ns2 simulation to verify. ns2 is a standard and accurate simulation software. Standard simplifying assumption s Simulation setup TCP Source 1 TCP sink 1 1Mb/s 1Mb/s Node 1 Node 2 1Mb/s TCP Source 20 1Mb/s Compare two configurations • static:50Mb/s • DBA: 5~50Mb/s TCP sink 20 20 additional TCP flows come and go abruptly. Zero throughput reduction 50 Fast recovery 40 static DBA instant increase throughput preservation 30 Throughput does not oscillate. 20 Initial dip throughput 10 preservation 0 150 200 TCP flows come throughput TCP flows preservation go 250 time (s) 300 350 Maximum energy saving 50 short transient same as throughput static DBA 40 capacity ramps down slowly 30 capacity ramps up fast 20 10 same as throughput same as throughput 0 150 TCP flows come 200 TCP flows go 250 time (s) 300 350 Concluding remarks Network link is lightly utilized, can reduce capacity to save energy. Propose DBA to adjust link capacity in TCP flow time scale. Optimality: zero throughput reduction, maximum energy saving. Stability: locally asymptotically stable. Verified by ns2 simulations.