Moving Content to Wireless Edge

Transcription

1 Moving Content to Wireless Edge - Improving Backhaul-limited Small Cell Performance - Dragan Samardzija February, 2013 Discussions with Markos Tavares, Howard Huang, Mohammadali Ali, Ivica Rimac, Reinaldo Valenzuela

2 Motivation and Key Ideas Content Delivery for Small Cells via Wireless Backhaul Cache Hit Rate and Content Characteristics Appendix: User-generated Video Content Characteristics 2

3 Motivation Non-real-time and Delay-tolerant Content Dominates Non-real-time and delay-tolerant content. Most of downloaded data is video content (user-generated video, movies, series). Youtube, Hulu, Netflix Even braking news could be delayed for a few minutes. Source: Allot Mobile Trends 3

4 Motivation A Few Very Popular Web Sites High popularity of certain web sites. News: New York Times, BBC, Bloomberg, Daily Mail News portals : Drudge Report, Google News, Yahoo Sports: ESPN, NFL, NBA, NHL Video content: Youtube, Hulu Source: Allot Mobile Trends 4

5 Motivation Backhaul as Limiting Factor Key problems associated with small cell deployments: 1. siting, 2. backhauling, 3. powering. A single small cell may support many access technologies with many carriers, leading to high access data rate. An example based on correspondence with Jonathan Siegel, November 2012: LTE+3G+WiFi for Capacity Average rate: 61 MBPS. Peak rate: 299 MBPS. Configuration: 5 MHz 3G + 20 MHz LTE 2x2 + WiFi GHz. With a greater number of small cells backhaul becomes a limiting factor. 5

6 Key Ideas 1. Pre-load video content and most popular web-sites in small cells. A small cell contains content, and when a terminal requests it, it is readily available, i.e., no need for fetching it through backhaul. The content is periodically updated and expanded. 2. If applicable, use broadcast in backhaul so that multiple small cells are pre-loaded with the new and updated content. Multiplexing effect in the backhaul. Wireless MBMS, PON This effectively moves content to the very wireless edge, the small cell itself. Similar to Content Delivery Network in the general Internet world. Check tutorial on 6

7 Motivation and Key Ideas Content Delivery for Small Cells via Wireless Backhaul Cache Hit Rate and Content Characteristics Appendix: User-generated Video Content Characteristics 7

8 Wireless Backhaul Example For simplicity, but not necessarily, backhaul and access occupy different bands and could be different technologies. backhaul access access macro-cell backhaul hub backhaul small cell 2 small cell 4 backhaul access backhaul access small cell 1 small cell 3 backhaul access small cell 5

9 Small Cell Architecture Example macro-cell backhaul hub HSPA access backhaul LTE access WiFi access Backhaul Wireless Interface (e.g., 4x4 20 MHz TD-LTE on a separate carrier) Small Cell Controller (e.g., general up) Cache (e.g., a hard drive ) Wireless Access Interfaces WiFi n/ac LTE Carrier 1 LTE Carrier 2 HSPA 9

10 Wireless Backhaul Pre-loading and Unicast Period T PL 1- T PL time During this period content is being updated and expanded (i.e., preloaded). During this period backhaul transmission to individual small cells (i.e., unicast backhauling). The radio signal is broadcasted from the backhaul hub to each small cell. The received data is being stored in the small cell cache. Possibly all backhaul hubs broadcast the same signal creating a single frequency network (SFN). 10

11 How the System Operates ACCESS 1. A terminal may request data to be downloaded. 2. The small cell that the terminal is attached to checks whether the data is pre-loaded in its cache. 2.1 If the data is pre-loaded, it will be sent it to the terminal over the access link. 2.2 If data is not in the small cell cache, it will be first fetched over the unicast backhaul link, and then sent over the access link to the terminal. BACKHAUL There are two separate periods, repeated periodically. 1. Certain content is pre-loaded using a broadcast transmission over the backhaul channel to each small cell in the broadcast coverage range. 2. Individual small cells may request data to be fetch over the backhaul when a terminal requests data that is not already pre-loaded in the small cell cache. CONTENT that is pre-loaded is decided on based on many possible criteria employed by the network operator. 11

12 Data Rate Gain when Using Content Pre-loading Backhaul-limited Case GAIN N SC N R R SC BH SC R SC 1 T 1 1 T 1 PL N SC PL R BH _ SC N SC R BH _ SC probability that the requested data will be in the small cell cache, i.e., the cache hit rate. N SC number of small cells per macro cell. R SC access small cell rate. R BH_SC per-small-cell backhaul rate, R BH_SC < R SC. R BH aggregate backhaul rate, R BH =N SC R BH_SC. T PL fraction of time dedicated to pre-loading though the backhaul broadcast. 12

13 Rate Gain as a Function of the Cache Hit Rate N SC = 10 T PL = 0.1 R SC =2 R BH =1 Greater the small cell access rate R SC vs. backhaul rate R BH is, greater the gain. For a modest cache hit rate of 20%, =0.2, the rate gain is 4.7X or 2.7X, depending on R SC and R BH. R SC =1 R BH =1 In other words, for a target access rate, the backhaul rate may be lowered 4.7X or 2.7X, depending on R SC and R BH. 13

14 Use SFN Goal is to increase spectral efficiency of the pre-loading phase. Use SFN, which results in much higher rates and/or coverage than a single macrocell, i.e., backhaul hub transmission. Multiple backhaul hubs transmit the same broadcast signal. backhaul access small cell 4 macro-cell backhaul hub 1 backhaul small cell 2 access macro-cell backhaul hub 2 backhaul small cell 1 access backhaul small cell 3 access macro-cell backhaul hub 3 14

15 Reuse-1 vs. SFN Edge SNR= 20 db 23 db Significantly better performance of SFN, which is best suited for the pre-loading broadcast backhauling. ~23 db better median SINR. 99% coverage ~ 10 db. Shannon AWGN rate ~3.5 bps/hz. 50% achievable (overhead, imperfections). 4x4 with ¾ MIMO efficiency. W= 10 MHz. Data rate at 99% of locations better than 4 x 0.75 x 0.5 x 10 6 log(1+snr) ~ 50 MBPS

16 LTE for Backhaul Multimedia Broadcast and Multicast Service (MBMS) In multi-cell mode LTE offers the possibility to transmit multicast/broadcast over a single frequency network (MBSFN). A time-synchronized common waveform is transmitted from multiple cells for a given duration. The cyclic prefix is used to cover the difference in the propagation delays, which makes the MBSFN transmission appear to the UE as a transmission from a single large cell. MBMS transmission types: 1. Transmission on a dedicated carrier for MBSFN with the possibility to use a longer CP with a subcarrier bandwidth of 7.5 khz is supported. 2. Transmission of MBSFN on a carrier with both MBMS transmissions and unicast transmissions using time division multiplexing. Well-suited for the content pre-loading and unicast backhaul. 16

17 Motivation and Key Ideas Content Delivery for Small Cells via Wireless Backhaul Cache Hit Rate and Content Characteristics Appendix: User-generated Video Content Characteristics 17

18 Cache Hit Rate The rate gain is increasing with the probability that the requested content is already in the small cell cache. 1 = (R PL T PL, Q) R PL pre-loading rate. R PL T PL T PL fraction of time dedicated to pre-loading though backhaul broadcast. Q cache size. The cache hit rate is a function that increases with the amount of information downloaded during the pre-loading period of the backhaul transmission (i.e., the cache update rate), and overall size of the cache. It strongly depends on the pre-loaded content popularity, which is why the functions is not strictly defined. 18

19 News Portals Size and Update Estimates A typical HTML news page, with images, is less than 1 MB (based on random sampling).. The front news portal page has typically less than 100 articles pointing to. Downloading the news front page and all articles it points to is expected to be less than 100 MB. Assumptions pre-load 100 news portals, and have a full update 3 times a day. * Off-the-shelf PC hard drive, 4 TB, $250, Best Buy January Memory size requirements 10 GB, or 0.25% of a 4 TB disk space*. Backhaul requirements 2.77 MBPS or 5.5% time when using 4x4 10 MHz LTE, SFN, 99% coverage.

20 Wikipedia Size and Update Estimates Jerry takes his 246 KB. User-generated content (text and still images). The typical size of an article is 0.5 MB (based on random sampling). The number of articles in English 4,162,172 (reported by Wikipedia). The number of new articles appearing on average is 814 per day. a:size_of_wikipedia Memory size requirements 2 TB or ~52% of a 4 TB disk space. Possibly selecting a certain portion of Wikipedia content. Backhaul requirements to send all new articles KBPS or 0.07% time when using 4x4 10 MHz LTE, SFN, 99% coverage.

21 Youtube Size and Update Estimates No information on the absolute size of Youtube. 72 hours of video are uploaded to Youtube every minute. Assumed coding rate is size of all video clips is 30 MB. A screen size 640 x minutes playback time MBPS coding rate. A typical user-generated video is reported to be encoded at 300 to 500 KBPS. Professionally-produced Full HD quality videos are closer to 6 MBPS (e.g., Gangam Style is 5.5 MBPS) Memory size requirements to record content generated in 1 day 62 TB, or sixteen 4 TB disks. Backhaul requirements to send the updates 5.76 GBPS. Only a fraction of Youtube content can be downloaded. 21

22 User-generated Video Content Popularity CCDF of Views per Day Experimental example of content popularity. The top 1% and top 10% of the videos (in terms of views received in the 2 week period) account for roughly 40% and 80%, respectively, of the total views. Characterizing Web-Based Video Sharing Workloads, Siddharth Mitra, Mayank Agrawal, Amit Yadav, Niklas Carlsson, Derek Eger, Anirban Mahanti 22

23 User-generated Video Content Popularity contd. The exact shape depends on a particular experimental sample set. 20% cache hit rate, for Youtube. 870 GB to store two weeks of Youtube videos, occupying 21% of the 4 TB disk space MBPS for update, or 11.5% time when using 4x4 10 MHz LTE, SFN, 99% coverage. 23

24 Estimate of Rate Gain N SC = 10 T PL = 0.2, i.e., 20% of backhaul for pre-loading. R SC =3 R BH =1 R SC =2 R BH =1 ~6% for news and Wikipedia, ~11% for Youtube, ~3% for other preloads 5 full-length HD movies, each 3 GB, per day. Cache size of ~ 4 TB, preferably larger. R SC =1 R BH =1 Expected region of gains. 24

25 Improving the Cache Hit Rate Make customers aware of the pre-loaded content and entice them to use it. Through a portal-like access to news, video clips, movies Selection of pre-loaded content is service-provider-specific, allowing for market-place differentiation. May allow service providers to pre-load exclusive content for particular content providers (e.g., a deal between VzW and Disney). A way for service providers to increase revenues. Customers will experience significantly faster downloads for the exclusive content compared to a more general one. Walled-garden, at its best. If in coverage, other small cells to store data that is backhauled to individual small cells. Opportunistic broadcasting and caching (to be investigated). 25

26 Conclusions Potential for significant data rate gains if content pushed into small cells. Use of SFN when backhaul broadcast. Is this the right place for LTE/LTE-A for backhauling? How to select the right content for preloading and entice users to use it? Should customers or service providers be the owners of the small cells? The MiFi model or conventional small-cell ownership? Should we place small cells in cars, buses? 26

27 Motivation and Key Ideas Content Delivery for Small Cells via Wireless Backhaul Cache Hit Rate and Content Characteristics Appendix: User-generated Video Content Characteristics 27

28 Appendix -User-generated Video Content Characteristics- 28

29 Youtube Statistics Over 800 million unique users visit Youtube each month. Over 4 billion hours of video are watched each month on Youtube. 72 hours of video are uploaded to Youtube every minute. 29

30 CCDF of Weekly Views Youtube video files. Distribution of added views at week i, for recently-uploaded and keyword-search videos. Characterizing and Modeling Popularity of User-generated Videos Youmna Borghol, Siddharth Mitra, Sebastien Ardon, Niklas Carlsson, Derek Eager, Anirban Mahanti 30

31 CCDF of Views After 32 weeks, the recently-uploaded data set. Characterizing and Modeling Popularity of User-generated Videos Youmna Borghol, Siddharth Mitra, Sebastien Ardon, Niklas Carlsson, Derek Eager, Anirban Mahanti 31

32 CDF of Time to Popularity Peak Randomly selected most recent Youtube video files. Characterizing and Modeling Popularity of User-generated Videos Youmna Borghol, Siddharth Mitra, Sebastien Ardon, Niklas Carlsson, Derek Eager, Anirban Mahanti 32

33 CCDF of Video Weekly Views For videos that are in their at-peak, before- and after-peak phases. Note that during the popularity peak, order of magnitude more views are recorded compared to the off-peak. Characterizing and Modeling Popularity of User-generated Videos Youmna Borghol, Siddharth Mitra, Sebastien Ardon, Niklas Carlsson, Derek Eager, Anirban Mahanti 33

34 CDF of Video Duration Characterizing Web-based Video Sharing Workloads, Siddharth Mitra, Mayank Agrawal, Amit Yadav, Niklas Carlsson, Derek Eger, Anirban Mahanti 34

35 Popularity vs. CCDF Plot Zipf distribution : the number of occurrences y relates to the rank of the object r as y r θ, where θ is the exponent of the Zipf distribution. Zipf-like behavior- with the presence of an approximate straight line when a logarithmic scale is used on both. While the rank-frequency plot shows most clearly the distribution of the lukewarm and cold objects, the CCDF plot shows most clearly the distribution of the hot and lukewarm objects. Characterizing Web-based Video Sharing Workloads, Siddharth Mitra, Mayank Agrawal, Amit Yadav, Niklas Carlsson, Derek Eger, Anirban Mahanti 35

36 Modeling It appears that no fit is qualitatively significantly better than the other fits. The middle region to be best modeled by a power law. The left-most region and part of the middle region appears to be better modeled by a lognormal distribution or by a power law with cut off. Characterizing Web-based Video Sharing Workloads, Siddharth Mitra, Mayank Agrawal, Amit Yadav, Niklas Carlsson, Derek Eger, Anirban Mahanti 36

37 CCDF of Views per Day Dailymotion videos. The top 1% and top 10% of the videos (in terms of views received in the 2 week period) account for roughly 40% and 80%, respectively, of the total views. Characterizing Web-based Video Sharing Workloads, Siddharth Mitra, Mayank Agrawal, Amit Yadav, Niklas Carlsson, Derek Eger, Anirban Mahanti 37

38 Observed Invariants The number of uploaded videos is several orders of magnitude smaller than the number of views to these videos. Heavy download vs. upload. Both the total number of views of videos, and the rate with which new views occur, follow the Pareto rule, with 20% of the most popular videos accounting for 80% or more of the views. The total views popularity distribution is heavy-tailed and may be modeled as power law or power law with exponential cut off, with power law exponent between 1.4 and 2.5. Neither a power law (or variants), nor lognormal distribution, appear to fit the entire distribution well. There are not many one-timers, according to views since upload. When considering views over a fixed-length trace period, the proportion of onetimers becomes more comparable to that with traditional web workloads. Characterizing Web-based Video Sharing Workloads, Siddharth Mitra, Mayank Agrawal, Amit Yadav, Niklas Carlsson, Derek Eger, Anirban Mahanti 38