Presentation is loading. Please wait.

Presentation is loading. Please wait.

Periodic Transfers in Mobile Applications: Network-wide Origin, Impact, and Optimization Feng Qian 1, Zhaoguang Wang 1, Yudong Gao 1, Junxian Huang 1 Alexandre.

Similar presentations


Presentation on theme: "Periodic Transfers in Mobile Applications: Network-wide Origin, Impact, and Optimization Feng Qian 1, Zhaoguang Wang 1, Yudong Gao 1, Junxian Huang 1 Alexandre."— Presentation transcript:

1 Periodic Transfers in Mobile Applications: Network-wide Origin, Impact, and Optimization Feng Qian 1, Zhaoguang Wang 1, Yudong Gao 1, Junxian Huang 1 Alexandre Gerber 2, Z. Morley Mao 1, Subhabrata Sen 2, Oliver Spatscheck 2 1 University of Michigan 2 AT&T Labs - Research April 18 2012

2 Introduction Typical testing and optimization in cellular network Little focus has been put on their cross-layer interactions Many mobile applications are not cellular-friendly. The key coupling factor: the RRC State Machine – Determine the resource management policy – Similar RRC state machines exist in 2G, 3G, and 4G networks Page 2 RRC State Machine ?

3 Background: Radio Resource Management in Cellular Networks RRC (Radio Resource Control) state machine [3GPP TS 25.331] – State promotions have promotion delay – State demotions incur tail times Tail Time Delay: 1.5s Delay: 2s RRC StateChannel Radio Power IDLENot allocated Almost zero CELL_FACH Shared, Low Speed Low CELL_DCH Dedicated, High Speed High UMTS RRC State Machine for a large US 3G carrier Page 3

4 Background: Radio Resource Management in Cellular Networks Promo Delay 2 Sec DCH Tail 5 sec FACH Tail 12 sec Tail Time Waiting inactivity timers to expire Page 4

5 Cellular Periodic Transfers What are periodic transfers? – A handset (mobile device) periodically exchanges some data with a remote server every t seconds. Why do they occur? – Keep-alive, periodic polling, periodic measurements… – Easy to program: java.util.Timer.scheduleAtFixedRate() Why are they bad in cellular networks? – They are small and short – Each transfer incurs a long tail We perform the first network-wide study of cellular periodic transfers to understand their Prevalence Resource impact Application semantics Potential for improving their inefficiency Page 5

6 Talk Outline: Cellular Periodic Transfers… Why are they bad in cellular networks? How prevalent are they? Why do they occur in mobile applications? What are their network-wide resource impact? How to make them more resource-efficient? Page 6

7 Talk Outline: Cellular Periodic Transfers… Why are they bad in cellular networks? How prevalent are they? Why do they occur in mobile applications? What are their network-wide resource impact? How to make them more resource-efficient? Page 7

8 Small Data Transfers: 3G vs. Wi-Fi Perform controlled experiments – Download a small HTTP object in 3G and WiFi – Using a power monitor to measure energy consumption Radio energy breakdown Promotion (~2 sec) Promotion (~2 sec) Data Transfer Data Transfer DCH Tail (5 sec) DCH Tail (5 sec) FACH Tail (12 sec) FACH Tail (12 sec) 3G: Data Transfer Data Transfer Wi-Fi Tail (250 ms) Wi-Fi Tail (250 ms) Wi-Fi: Page 8

9 Measurement Results Object A: 1 KB Object B: 9 KB Wi-Fi is 140x more efficient than 3G, because Wi-Fi has a much smaller RTT Wi-Fi radio power is only half of 3G radio power Wi-Fi has a much shorter tail time For transferring A (B) using 3G, 97.0% (94.3%) of radio energy belongs to the tail. Small traffic bursts are extremely resource-inefficient in cellular networks Root cause: promotion delay the tail time Transferring them periodically leads to even more serious resource inefficiencies Page 9

10 Talk Outline: Cellular Periodic Transfers… Why are they bad in cellular networks? How prevalent are they? Why do they occur in mobile applications? What are their network-wide resource impact? How to make them more resource-efficient? Page 10

11 Cellular Measurement Dataset Packet traces collected from a large U.S. cellular carrier Collected at the cellular core network without sampling 1.5 billion packets for 1.25 hours in December 2010 Recorded IP and TCP(UDP) headers and timestamps Extracted 2.8 million user sessions Page 11

12 The Periodicity Detection Algorithm Existing approaches: DFT and auto-correlation – Used to estimate RTT (the “clock” of TCP flows) – Not work well in our scenario: periodic transfers have much fewer samples than RTTs We proposed a simple heuristics-based approach – Input: a session; output: periodicities and periodic transfers – Key idea: exhaustively search for repetitions of data transfers spaced by a fixed time period Page 12 t seconds

13 Measurement Results Key algorithm parameter: – θ (minimal repetitions to be observed before declaring a periodicity) – We use θ = 4 (more details in paper) What is the distribution of periodicities? A particular value of one-minute dominates the periodicity. Likely to be set by developers in an ad-hoc manner. Page 13

14 Measurement Results (cont.) Prevalence of periodic transfers? – A long session: at least one minute – They occur in about 20% of long sessions Typical size and duration of a periodic transfer? They are Small: 25 th / 50 th / 75 th percentiles: 0.2 KB, 1.1 KB, 1.8KB 97% of periodic transfers < 10 KB They are short: 90% are less than 7 seconds Page 14 Cellular periodic transfers are: Prevalent: they occur in 20% of long sessions Small: median size is 1.1 KB Short: 90% are less than 7 seconds At least 70% of periodicities are one minute

15 Talk Outline: Cellular Periodic Transfers… Why are they bad in cellular networks? How prevalent are they? Why do they occur in mobile applications? What are their network-wide resource impact? How to make them more resource-efficient? Page 15

16 Understanding Origins of Cellular Periodic Transfers Leverage a database generated by the same carrier – Contains IP address  content provider mappings – Based on the data collected on the same day / same location as the packet trace was collected IPs of 46% of detected periodic transfers have meaningful content provider names in the database Page 16 IP  content provider database 184.73.206.70  lt.andomedia.com 170.149.173.1  www.nytimes.com 212.58.244.69  www.bbc.co.uk …

17 Origins of Cellular Periodic Transfers Content Provider / Applications % Periodic transfers Remarks facebook.com48.4%Keep connection alive for pushing andomedia.com15.5%Pandora’s audience measurement medialytics.com4.9%User behavior monitoring DNS14.1%DNS lookups Advertisements3.3%Advertisement update gmail.com1.4%Checking emails pinger.com1.4%Polling to fetch updates for SMS (Other)11.1%e.g., periodically check the weather (within 46% of all detected periodic transfer instances) Many periodic transfers are either unnecessary or overly aggressive Page 17

18 Case Studies Facebook – Periodic transfers (60s) used as keep-alive messages – Prevent a TCP connection from being closed by the NAT – For the four largest U.S. cellular carriers, the timeout of cellular NAT is at least 4.25 minutes [Wang et al, SIGCOMM 2011] Pandora (music streaming) – Periodic audience measurement uploaded to andomedia.com every one minute – Even when Pandora is running in the background Page 18

19 Talk Outline: Cellular Periodic Transfers… Why are they bad in cellular networks? How prevalent are they? Why do they occur in mobile applications? What are their network-wide resource impact? How to make them more resource-efficient? Page 19

20 Quantifying the Resource Impact of Periodic Transfers Use our trace-driven RRC state machine simulator with a handset radio power model [Qian etal, Mobisys 11] Three metrics of resource consumption – D: radio resource consumption Quantified by the CELL_DCH occupation time – S: signaling load, quantified by the total promotion delay – E: handset radio energy consumption Computed using a handset radio power model Page 20

21 Quantifying the Resource Impact of Periodic Transfers Compute the impact Quantify the impact at four study scopes – All sessions in the dataset – All sessions contain periodic transfers – All Facebook sessions – All Pandora sessions ΔE = (E R – E 0 ) / E 0 E 0 : Radio energy consumption in original sessions E R : Radio energy consumption in modified sessions with periodic transfers removed ΔE : Radio energy impact of periodic transfers (the value is negative) Page 21

22 The Resource Impact of Periodic Transfers Study Scope ΔV Traffic Volume ΔE Radio Energy ΔS Signaling Overhead ΔD Radio Energy All sessions0.4%-7.9%-8.9%-6.5% Periodic sessions0.7%-20.4%-25.3%-15.6% Facebook sessions1.7%-30.5%-30.4%-30.5% Pandora sessions0.5%-28.7%-35.0%-20.5% Huge disparity between traffic volume and resource consumption of periodic transfers All sessions: ΔE is 20 times of the ΔV Pandora: ΔE, ΔS, and ΔD are 40 to 70 times higher than ΔV Page 22 The state-of-art cellular periodic transfers are extremely resource inefficient The root cause: tail time and promotion overhead

23 Talk Outline: Cellular Periodic Transfers… Why are they bad in cellular networks? How prevalent are they? Why do they occur in mobile applications? What are their network-wide resource impact? How to make them more resource-efficient? Page 23

24 Optimizing Periodic Transfers Periodic transfers are delay-tolerant – Not initiated by user inputs – Applications usually have the flexibility over a time window in scheduling each transfer Can existing optimization approaches effectively reduce their resource impact? – Perform “what-if” analysis for existing opt. techniques – Resource reduction computed as – Focus on Pandora and Facebook sessions ΔE = (E after_opt – E before_opt ) / E before_opt Page 24

25 What-if Analysis: Increasing Periodicity Increasing the periodicity to 5 min: reduce the resource consumption by 22%~ 28% Tradeoff between resource saving and application semantics Study Scope ΔE Radio Energy ΔS Signaling Overhead Facebook-30.5%-30.4% Pandora-28.7%-35.0% Page 25

26 What-if Analysis: Fast Dormancy Fast dormancy [3GPP Release 7 R2-075251] – Handset actively requests a state demotion after data transfer – Controlled by an independent fast dormancy timer ----- Without FD ----- With FD The Fast Dormancy Timer

27 What-if Analysis: Fast Dormancy Fast dormancy on only periodic transfers Fast dormancy on all transfers Blindly applying fast dormancy on all traffic is not recommended Doing that only at the end of periodic transfers is acceptable Tradeoff between resource saving and signaling overhead Page 27

28 Summary: Cellular Periodic Transfers… Why are they bad in cellular networks? – They are small and short – The tail effect How prevalent are they? – 20% of long sessions (> 1min) What are their key characteristics? – Small and short – 60 seconds dominates the periodicity Page 28

29 Summary: Cellular Periodic Transfers… Why do they occur in mobile applications? – Keep alive, measurement, polling, advertisement… – Many are too aggressive or even unnecessary What are their network-wide resource impact? – Up to 30% for popular apps – Resource impact is 20~70x of bandwidth usage impact How to make them more resource-efficient? – Existing techniques are effective, incurring various tradeoffs Page 29

30 Research Impact Our finding reached Pandora and Facebook The ARO (mobile Application Resource Optimizer) tool for profiling smartphone apps http://web.eecs.umich.edu/~fengqian/ Page 30

31 Backup Slides

32 The Periodicity Detection Algorithm Existing approaches: DFT and auto-correlation – Used to estimate RTT (the “clock” of TCP flows) – Not work well in our scenario: periodic transfers have much fewer samples than RTTs We proposed a simple heuristics-based approach – Input: a user session – Output: periodicities and periodic transfers – Assume each periodicity is associated with the same server – Allow multiple periodicities appear in a session Page 32

33 The Periodicity Detection Algorithm For packets of each server IP address 1.Discretize timestamps using a slot length of ω 2.Search for periodicity of t min <= t <= t max slots (2 conditions) 3.Identify packets associated with each periodic transfer Evaluate by performing manual inspection Slot len ω sec A marked slot contains at least one packet of the target IP t: the detected periodicity Cond1: Observe at least θ marked slots spaced by a fixed number of t slots (e.g., θ = 3) Cond2: no marked slots between periodic seeds Periodic Seed 1 Periodic Seed 2 Periodic Seed 3 Page 33

34 Optimization Techniques ----- Without FD ----- With FD The Fast Dormancy Timer Page 34


Download ppt "Periodic Transfers in Mobile Applications: Network-wide Origin, Impact, and Optimization Feng Qian 1, Zhaoguang Wang 1, Yudong Gao 1, Junxian Huang 1 Alexandre."

Similar presentations


Ads by Google