EERA Energybased Rate Adaption for 802 11 n
EERA: Energy-based Rate Adaption for 802. 11 n Chi-yu Li*, Chunyi Peng*, Songwu Lu*, Xinbing Wang+ *University of California, Los Angeles, +Shanghai Jiaotong University ACM MOBICOM 2012 Istanbul, Turkey
Increasing Popularity of 802. 11 n 2 802. 11 n chipset shipment 450 M+ units in 2010, >1 billion in 2012 (expected) Annual growth > 15% Wi-Fi Chipset Shipments, by Protocol (ABI Research, May 2010) 2500 2000 1500 802. 11 n 1000 500 2015 2014 2013 2012 2011 2010 2009 2007 2006 2005 2008 802. 11 a/g 0 2004 Shipments (Million)
Increasing Power by 802. 11 n 3 Higher power consumption compared with legacy 802. 11 a 3 x 3 MIMO RX: 2 x during active 3 x 3 MIMO RX: 1. 5 x during idle Even higher if more antennas are used (up to 8 for 802. 11 ac)
802. 11 n Rate Adaptation 4 RA is the popular mechanism to boost wireless performance Select the best 3 -tuple MIMO setting over timevarying channel Modulation and coding scheme (MCS): 6. 5 Mbps, …, 600 Mbps Number of activated antennas: 1, …, 4 Stream modes: SS, DS, TS, QS Traditional design goal: Highest goodput What about energy efficiency?
Goal for this Work 5 Energy perspective for RA design in 802. 11 n NIC Limitation Design of traditional RA in energy savings of EERA: Energy-based RA
Outline 6 Case Study on 802. 11 n RA Finding, General scenarios Highest goodput ≠ Energy efficiency EERA design Single root cause client, multiple clients Evaluation Comparison Conclusion with 3 other schemes
Case Study 7 2 MIMO RA algorithms ARA: Atheros RA Mi. RA [Mobicom’ 10] Up to 3 x 3 antennas, triple-stream (TS) mode Software: ath 9 k open source driver & Host. AP Power meter: Agilent 34401 A Zigzags between MIMO modes 802. 11 n NIC: Atheros AR 9380 2. 4/5 GHz MIMO chipset Excludes half of rates to reduce search space An accuracy of 100 u. W or 10 u. W Setting: AP mode; static; fixed-rate (30 Mbps) UDP
Limitation 1: Energy Inefficiency 8 EE ARA Mi. RA Goodput (Mbps) 35. 4 52. 5 2 -min Energy (J) 69. 0 106. 2 105. 6 Pet-bit-energy Eb (n. J/bit) Gap (%) 19. 2 29. 7 29. 4 - 54. 5% 52. 9% High goodput, but not energy efficiency
Root Cause: Highest Goodput ≠ Energy Efficiency 9 EE v. s. Highest-Goodput (HG) settings The gap between EE and HG reaches 11. 1 n. J/bit Incurring energy waste 57. 8% using HG Major rates selected by ARA/Mi. RA EE HG 40. 5 SS 81 DS 108 DS 81 TS 121. 5 TS
Why HG ≠ EE? 10 More antennas and more streams activated for higher goodput Small goodput gain at a high energy cost 3 x 1/40. 5 SS 3 x 3/81 DS Slow down can save energy, while still accommodating traffic source
Limitation 2: Slow Convergence 11 Multiple rounds to reach HG setting by ARA and Mi. RA Root cause: Sequential search Scaling issue in many-antenna 802. 11 x: 360 options in 8 -antenna 802. 11 ac vs. 48 options in 3 -antenna 802. 11 n
In General Scenarios 12 Generally, HG ≠ EE Locations # of activated AP antennas Traffic source rate Power saving schemes SMPS: one receiver antenna; PSMP: sleep mode HG: 3 x 3/81 DS 3 x 1 40. 5 SS 3 x 2 81 DS Data source rates HG: 3 x 3/81 DS 3 x 1 40. 5 SS 3 x 1 54 SS Power-saving schemes
Quantify NIC Energy Efficiency 13 Per-bit energy consumption: Eb Eb = # bits Active = § § Energy Non-active Pa × Ta + Pna × Tna S × (Ta + Tna) § § Rate setting Active Power model § Rate setting § Idle power model § Power save scheme Rate setting goodput Traffic source rate Tradeoff between power consumption and goodput
EERA: Energy-Based RA for 802. 11 n 14 Idea: Slow down to save energy tradeoff goodput for energy efficiency but still accommodate the data source How to locate slow rate for energy saving? How to locate it faster? How to control the degree of slowdown?
EERA Design 15 Single-client: How to locate the low-energy MIMO setting Search over multi-level tree Ternary search over each branch Simultaneous pruning by leveraging MIMO features Multi-client: on top of single-client design How to prevent each client from affecting others due to its slowdown Ensure fair share of airtime by each client Tradeoff between energy efficiency and fairness
Multi-dimensional Search Problem 16 On 4 dimensions # of transmit antennas (Nt) # of receiver antennas (Nr) # of data streams (Nss) MCS options (NMCS) L 1: Nt L 2: Nr 1 L 3: Nss SS L 4: SS 3 Heuristic: AP uses the maximum number of antennas 2 3 DS SS DS TS … … … MCS 13. 5 M… 135 M 27 M… 270 M 40. 5 M… 405 M
Ternary Search over Each Branch 17 Unimodal function: Eb w. r. t. MCS rate Binary search not applicable Example: 3 x 2/DS branch Eb (n. J/bit) MCS 3 x 2/DS 0 27 1 54 25. 2 2 81 24. 1 3 108 23. 2 4 162 23. 8 5 216 6 243 7 270 4 Steps
Simultaneous Pruning of Branches 18 Pruning over multiple branches during search: - High-Loss pruning: loss increases - Low-loss pruning: The lower (A) decreasing Nr, given the same bound of a setting’s per-bit MCS and Nss energy from loss-free (B) Nss, given the same MCS and Nr goodput 3 x 3/81 SS (26. 4 n. J/bit) 3 x 3/108 SS (∞ n. J/bit) Eb 3 x 3/SS 26. 6 26. 5 26. 4 ∞ Prune 15 settings Eb 3 x 3/DS Eb 3 x 3/TS Eb 3 x 2/SS Eb 13. 5 27 40. 5 13. 5 EERA takes 17 probes to locate the 27 54 81 27 3 x 1/40. 5 SS 121. 5 51. 3 40. 5 29. 0 most 81 EE, 40. 5 34. 8 54 108 ∞ are pruned) 162 22. 7 54 ∞ (31 settings ∞ 81 162 243 81 108 216 324 108 Sequential search needs 35 probes 121. 5 243 364. 5 121. 5 135 270 405 Prune 8 settings 3 x 2/DS Eb 3 x 1/SS 27 54 81 108 162 216 243 270 19. 2 ∞ 13. 5 27 40. 5 54 81 108 121. 5 135
Is this Enough? C 1 -Gput (Mbps) 19 Slow down by EERA clients might hurt others C 2 Source rate at C 1 (Mbps) C 1 … ARA ✔ … ARA ✗✗ ARA EERA
EERA+: Multi-Client Operation 20 Idea: An EERA+ client slows down only if other clients do not get hurt Isolation via fair share of airtime for each client Tair Phase I: get the temporal air C 1 time for each client S 1/G 1 (ARA) – Traditional MIMO RA An epoch of time (Tep) S 2/G 2 C 2 (EERA+) S 3/G 3 C 3 (EERA+)
EERA+: Multi-Client Operation 21 Phase II: fairly allocate extra air time to EERA+ clients Fair share of airtime (Fi) Tair S 1/G 1 C 1 (ARA) S 2/G 2 C 2 (EERA+) S 3/G 3 C 3 (EERA+) Fi An epoch of time (Tep)
EERA+: Multi-Client Operation 22 Phase III: Client i selects the most EE setting given the constraint Fi Prune the settings which are too slow to accommodate Si (EERA operation) Tair S 1/G 1 C 1 (ARA) S 2/G 2 C 2 (EERA+) S 3/G 3 C 3 (EERA+) Fi An epoch of time (Tep)
Evaluation 23 Comparing EERA with ARA, Mi. RA, and MRES[ICNP’ 11]: improve EE by adjusting the number of antennas on top of RA Scenarios Single Client Static, mobility, interference, power-saving modes, wireless configurations, … Multi-Client Multiple EERA/EERA+ clients Coexistence with EERA/EERA+ and non-EERA clients
Single Client 24 Static UDP: at different locations, with varying AP antennas# and PS modes Application: Web, Vo. IP, FTP, and Video streaming Static interference, mobility, trialsscenarios EERA can. TCP, locate the EE settings in field various ARA well to Mi. RA MRES rate TCP/App gain: adapt dynamic source Static UDP (13. 4 – 35. 6) % (14. 3 – 36. 1) % (5. 8 – 26. 8) % Mobility gain: locate the EE settings quickly with Static TCP (5. 1 – 20. 5) % (10. 4 – 32. 3) % (7. 3 – 23. 8) % low probing cost Application (26. 5 – 33. 9) % (26. 6 – 35. 2) % (6. 7 – 36. 5) % Mobility 27. 8 % 30. 1 % 20. 3 % Field Trials 31. 7 % 33. 1 % 24. 1 %
Multi-Client 25 EERA+ does not hurt coexisting non-EERA clients C 1: ARA (10 Mbps 50 Mbps); C 2: ARA EERA+ C 2: ARA C 2: EERA+ 10 20 30 40 50 Source rate at C 1 (Mbps) 3 x 1 108 SS C 2 -Eb (n. J/bit) 3 x 2 3 x 3 162 DS 243 TS 10 20 30 40 50 Source rate at C 1 (Mbps) Slowdown overhead: delay increase Multiple EERA clients: < 0. 2 ms per packet (< 14. 2%) Coexistence of EERA/ARA: <0. 08 ms per packet (<5. 3%)
Negative Impact on Device-Level Energy? 26 Slowdown may increase energy of other components: Two dominant components Display: its energy independent of NIC status CPU: its status only slightly changed due to slowdown Quantify the impact with applications Applications: Web, Vo. IP, FTP, and Video streaming There are negligible impacts on all of them except FTP Why FTP? FTP stops once a file transfer completes
Summary 27 Limitations of goodput-optimizing RAs Goodput ≠ Energy Efficiency @NIC Slow convergence due to sequential search EERA: Energy-based RA for 802. 11 n NIC Ternary search + simultaneous branch pruning Slow down limited by fair share of airtime Insights: Tradeoff between speed and energy Tradeoff between fairness and energy
Backup 28
Power Save Mechanisms in 802. 11 n 29 Spatial Multiplexing Power Save (SMPS) Static SMPS: the client statically retains a single receive chain Dynamic SMPS: the client switches to multiple receive chains during data transmission, but shifts back to one chain afterwards. Power Save Multi-Poll (PSMP) Scheduled PSMP (S-PSMP): AP periodically initiates a PSMP sequence to schedule the transmission Unscheduled PSMP (U-PSMP): AP starts an unscheduled sequence and delivers to those wakeup clients
Experimental Floorplan 30
802. 11 n Receiver Power Model 31 Goodput is affected by Number of receive chains (Nr), number of streams (Nss), and MCS rates (R) The power of an 802. 11 n receiver Active power model Pra = (a 1 · Nr + f(Nss)) · BW + a 2 · Nr + a 3 · R + Pf Idle power model Pri = i 1 · Nr · BW + i 2 · Nr + Pf Number of receive chains, number of streams, and MCS rates affect both
Power Model of an 802. 11 n Receiver 32 Active power model Pra = (a 1 · Nr + f(Nss)) · BW + a 2 · Nr + a 3 · R + Pf Idle power model Pri = i 1 · Nr · BW + i 2 · Nr + Pf Platform a 1 a 2 a 3 Atheros 9380 2. 3 19. 8 Intel 5300 3. 0 195. 0 Nr: Number of receive chains Nss: Number of streams BW: Channel bandwidth (MHz) R: MCS Rate (Mbps) f(Nss) Pf i 1 i 2 SS DS TS (m. W) 0. 3 0. 6 4. 6 7. 0 429. 0 2. 3 19. 8 0. 3 3. 3 4. 1 4. 3 496. 8 2. 9 195. 0 MOBICOM 2012
Power Measurement and Estimation 33 Power v. s. Nss (Nr/R/BW) Power v. s. BW (Nr/RNss) Power v. s. Nr (RNss/BW) Power v. s. R (Nr/Nss/BW) MOBICOM 2012
Eb Estimation 34 Eb = Pa × Ta + Pna × Tna = S × (Ta + Tna) Pa − Pna G + Pna S Pa, Pna: obtained from power models G: estimated from probing S: estimated from buffer change Eb = (1 -a) Eb (t) + a Eb
Other Issues in EERA 35 Coexistence of EERA and other MIMO RA clients Greedy clients EERA sets the pre-configured parameter Ri : how much goodput the client is willing to give up for energy saving Uplink case EERA has an option to revert to goodput-optimizing RA mode EERA seek to minimize (Pa(tx) – Pi) / GUL AP calculates fair share for each uplink/downlink client, and then notify it of its uplink airtime share Ad-hoc mode: not supported due to two challenges How to allocate fair share of airtime in the multihop setting? How to coordinate RA operations among multiple clients in a fully distributed manner?
Device-level Energy Efficiency 36 Any impact on the energy consumption of other device components? Consider Display and CPU: the dominant portion of device’s energy consumption Device: ASUS F 8 S laptop with Intel Core 2 Duo T 8300 CPU • Display energy consumption is independent of the NIC status • CPU status can be slightly changed due to slower transmission CPU State C 0 C 1 C 2 C 3 EE (3 x 1/40. 5 SS) 5. 8% 0% 26. 0% 66. 4% HG (3 x 3/81 DS) 5. 5% 0% 42% 52. 0% Power@800 MHz (W) 16. 8 | 21. 3 10. 3 | 13. 0 9. 8 | 12. 4
In Real Application Scenarios 37 The EE setting has negligible impact on the devicelevel energy consumption except in the FTP case FTP in HG stops consuming more energy once a file transfer completes The other applications include UDP flow (30 Mbps), Web, Vo. IP, Video streaming
- Slides: 37