QoE-Aware Resource Allocation for Crowdsourced Live Streaming: A Machine Learning Approach

QoE-A ware Resource Allocation for Cro wdsourced Li v e Streaming: A Machine Learning Approach Fatima Haouari, Emna Baccour , Aiman Erbad, Amr Mohamed, and Mohsen Guizani CSE department, College of Engineering, Qatar Univ ersity Abstract —Driven by the tremendous technological advance- ment of personal devices and the pr evalence of wir eless mobile network accesses, the world has witnessed an explosion in cro wdsourced liv e streaming. Ensuring a better viewers quality of experience (QoE) is the key to maximize the audiences number and increase streaming pro viders’ profits. This can be achieved by advocating a geo-distributed cloud infrastructure to allocate the multimedia resour ces as close as possible to viewers, in order to minimize the access delay and video stalls. Mor eover , allocating the exact needed resources befor ehand avoids over -provisioning, which may lead to significant costs by the service providers. In the contrary , under -pro visioning might cause significant delays to the viewers. In this paper , we introduce a prediction driven resour ce allocation framework, to maximize the QoE of viewers and minimize the resour ce allocation cost. First, by exploiting the viewers locations av ailable in our unique dataset, we implement a machine learning model to predict the viewers number near each geo-distributed cloud site. Second, based on the predicted results that showed to be close to the actual values, we formulate an optimization problem to pr oactively allocate r esources at the viewers proximity . Additionally , we will present a trade-off between the video access delay and the cost of resource allocation. Index T erms —QoE, Crowdsour ced live video, Resource allo- cation, Cloud computing, Machine learning. I . I N T RO D U C T I O N Crowdsourced liv e video streaming is on the rise, and it continues to gro w ev ery single day . As per Cisco mobile video traf fic statistics, mobile video content is predicted to present 82% of the global Internet traffic in 2021 as opposed to 73% in 2016 [1]. The rise in popularity of cro wdsourced liv e streaming can be attributed to technological advancement, proliferation of smartphones and wireless network availability , which have led crowdsourcers to broadcast their li ve videos to various content providers. One of the most popular live streaming platform is Facebook, which had 2.19 billion activ e users per month in the first quarter of 2018 [2]. As per [3] 78% of Facebook online users are watching live videos, and 1 out of 5 videos on Facebook is live. The industry and academia hav e shown an ov erwhelming interest in cro wdsourced streaming recently in terms of achie v- ing the best QoE as it is the key to increase the audiences number and the content providers’ rev enues. A series of recent studies have been conducted to determine the main factors that affect the viewers’ QoE [4, 5]. These studies re vealed that viewers QoE is primarily dependent on two factors: First, the video startup delay and playback buf fering stalls and second, the video quality which depends on the viewers’ internet connectivity quality and av ailable video representations. The authors in [5] highlighted that the higher the startup delay is, the more the viewers abandonment increases. The y also showed that viewers who experienced low QoE are less likely to revisit the content provider’ s application within a specific period of time. Therefore, video startup and rebuf fering delays hav e high impact on vie wers’ QoE. Ho wev er , the challenge is to serv e the vie wers with the best QoE possible, while minimizing the cost of resource allocation. Geo-distributed clouds are proposed to enhance the QoE. In this context, many efforts are working on presenting an efficient resource allocation by proposing heuristics and opti- misations. W u et al. [6] formulated an optimal vie wing request distribution in the geo-distributed clouds, they predicted users future demands based on their social influences using an epidemic model. He et al. [7] presented a resource allocation framew ork to allocate geo-distributed cloud service to crowd- sourcers for transcoding and serving viewers. K. Bilal et al. [8] presented a QoE-aware resource allocation optimization for crowdsourced multivie w live streaming to choose the optimal transcoding cloud site location, and the optimal set of video representations. The drawback of these traditional algorithms is the near optimal solutions they provide. The y lack the ability to allocate the exact resources needed beforehand. This may either lead to ov er-pro visioning of resources that may incur significant costs to the service pro viders, or under-pro visioning of resources that may cause delays to the viewers. Therefore, addressing such a trade-off proactively is a real challenge that requires some accurate prediction techniques. In this work, we are addressing the proactive resource allo- cation by adopting machine learning techniques for designing a predicti ve model for the vie wers’ locations. In particular, we consider predicting the number of viewers near each geo- distributed cloud site for each incoming live video, in order to proactiv ely allocate resources at the proximity of the vie wers. T o the best of our knowledge, there is no research work that applied machine learning techniques for resource allocation to maximize QoE and minimize the cost. Only a few studies adopted machine learning to improve the vie wers QoE, with their focus varies from dealing with the buf fering and the bitrate selection [9], to determining Adaptive Bitrate (ABR) best parameters in order to improve adaptive video streaming [10]. The authors in [9] proposed a video freeze predictiv e model to detect possible factors that lead to video stalling at the viewers side. A recent study by [10] proposed using decision trees to choose the best ABR parameters to improve the adaptiv e video streaming. Moreov er , few recent studies hav e used machine learning for predicting the viewers’ QoE. The authors in [11] predicted the users engagement score, by considering users engagement as a function of Quality of Service (QoS) factors and viewers preferences. Another work in [4] proposed a classification model for users engagement, where users engagement was quantified in terms of users number of visits and video watching time. The contributions of this paper are summarized as: • Using our collected Facebook 2018 li ve videos dataset [12] containing records of viewers’ locations for each video, we dev elop a regressi ve model using machine learning techniques that predicts the number of viewers near different geo-distributed cloud sites for each incom- ing liv e video. • T o serve the predicted viewers such that they experience the minimum startup delay with a minimal cost to the content provider , we formulate an optimization problem for allocating resources as near as possible to the viewers. The rest of this paper is or ganized as follo ws: Section II presents our system model composed of: (1) vie wers predicti ve model; (2) proactive resource allocation optimizer . W e e valuate our system and present a trade-off between minimizing latency and maximizing cost gain in Section III. Finally , section IV concludes the paper and discusses the future directions. I I . S Y S T E M M O D E L In our system, we adopt a geo-distributed cloud infrastruc- ture as shown in Fig. 1 that consists of multiple geographically distributed cloud sites owned by a content provider . Our predictiv e model and resource allocation optimizer are de- ployed in a centralized master server . A set of geo-distributed crowdsourcers broadcast their videos in real time, which will be allocated by default in their nearest cloud site. Each broadcaster cloud site will report the master server with the incoming li ve videos information. The predicti ve model will predict the number of viewers expected near each cloud site. Based on the predicted results, the optimizer will allocate liv e videos replicas across the geo-distributed cloud sites near the viewers proximity to minimize the delay and video stalls with the minimum possible cost. Moreov er , the optimizer determines from which cloud site the viewers should be served. In our work, we consider only the storage resources, while the computation resources for video transcoding are out of the scope of this paper . A. Predicting live video viewers 1) Dataset: in our work, we are using the Facebook 2018 liv e videos dataset collected by our team [12], containing more than two million Facebook liv e video streams. The acti ve video streams metadata are fetched every 3 minutes in dif ferent periods on January , February , March, May , June and July 2018. As a result, we obtained a list of fetches related to each video and containing the number of viewers at the recording time. The liv e videos are collected with many features such as creation time and day , broadcaster location, number of likes and most importantly the viewers’ locations. In this work, we selected six features for each video namely , the broadcaster name, content category , created time, created day , broadcaster Fig. 1: System model. location and the viewers’ locations as illustrated in Fig. 2. The viewers’ locations were selected from the video fetch with maximum number of viewers. Fig. 2: Predicti ve model input and output. 2) Prepr ocessing: as our objective is to predict the viewers number near v arious geo-distrib uted cloud sites, there was a need to preprocess our ra w data. First, we mapped the viewers locations into 10 Amazon W eb Services (A WS) cloud sites locations [13] namely , Asia-Mumbai, Asia-Seoul, Asia- Singapore, China-Ninxgia, Europe-Frankfurt, Europe-P aris, South America-Sao paulo, US East-Ohio, US East-V irginia and US W est-California. This was done by calculating the shortest distance between the viewers locations and the 10 A WS cloud sites locations. Furthermore, we calculated the number of viewers near each cloud site for each video. W e did the same to the broadcaster location, where we mapped his location into the nearest A WS cloud site. Moreov er , we clustered the created time into 6 time periods. Finally , we applied the categorical one-hot encoding to the time period, created day and broadcaster location features, while we used feature hashing introduced by [14] to transform the high- cardinality features namely broadcaster name and content category into hashed feature vectors. 3) Predictive model: the dataset used to train our models included 224,839 live video records collected in March, May Fig. 3: Models validation. Fig. 4: Models testing. (a) Asia Seoul. (b) Europe Frankfurt. (c) China Ningxia. Fig. 5: Hourly actual vs predicted viewers number . and June 2018. 80% of the records were randomly selected for training and 20% were used for validation. W e trained our regression models to produce 10 outputs as illustrated in Fig. 2, each represents the number of viewers near the 10 A WS cloud sites mentioned previously . W e adopted three different ML algorithms namely , Multilayer -perceptron (MLP), Deci- sion trees (DT) and Random F orest (RF). W e b uilt sev eral models using each ML algorithm, as there is no method to predetermine the best combination of hyperparameters, such as the number of hidden layers and neurons for MLP models, number of forests for RF models and the max depth for DT models. Finally , the best models were selected considering the best determination coefficient ( R 2 ) values, which is used to assess the goodness of fit of our regression models. R 2 values approaching 1 indicate that the model provides accurate predictions, and it is calculated according to Eq. (1): R 2 = 1 − P m i =1 ( A i − P i ) 2 P m i =1  A i − ¯ A  2 (1) where m is the number of videos, A i is the actual number of viewers for video i , P i is the predicted number of viewers for video i , and ¯ A is the mean of the actual number of viewers of all videos. 4) Predictive model results: after training the models, the validation results, depicted in Fig. 3, showed that RF outper- forms the other ML algorithms by achie ving for example an R 2 of 0.91 for Seoul, 0.89 for Sao Paulo, 0.85 for Ohio, 0.86 for California and 0.74 for China. The DT model achiev ed the lowest R 2 as opposed to MLP and RF . The results showed that increasing the number of layers for the MLP models improves the results. Howe ver , due to the complexity of the models, and because we noticed that there is a slight difference between the performance of the 5 layers model and the 7 layers model, we did not increase the layers above 7. The results also showed that for all ML models, the predicted number of viewers near some regions achieved a higher R 2 compared to other regions, China achieved the lowest, while Seoul and Sao paulo achiev ed the best R 2 . W e further tested our models on unseen data of li ve videos collected from July 1 to July 6, 2018. The models performed the same as with validation data in some regions, slightly less or higher in other regions as shown in Fig. 4. W e then extended our experiments by performing the predictions on hourly basis for 24 hours using the live videos of July 3, 2018. The RF and MLP 7 layers models were used for prediction, since they performed better than other models. The predicted number of vie wers for the hourly incoming liv e videos versus the actual number of viewers for Seoul, Frankfurt and China cloud sites are presented in Fig. 5. Since our results demonstrate that the RF predictions are the closest to the actual values, we will adopt this model in our system. B. Proactive live video allocation and viewers serving In this section, we formulate the problem of proactiv e resource allocation, to deriv e the optimal number of video allocation cloud sites and the nearest cloud site to serve the viewers, with an objective of minimizing the cost constrained by the access delay . W e then, present our proactiv e resource allocation algorithm. 1) Problem formulation: the set of incoming live videos at period t is denoted by V ( t ) = { v 1 , v 2 , v 3 ,.... v m } . The set of regions is represented by R = { r 1 , r 2 , r 3 ,.... r n } . Let r b , r a and r w denote the broadcasting region, video allocation region and video serving re gion respecti vely . The round trip delay from r a to r w is represented by d r a r w . Let P ( t ) = { P v 1 , P v 2 , ...P v m } represent the set of predicted viewers for the incoming videos at period t . As each video has predicted vie wers in different regions, let P v = { p 1 , p 2 , p 3 ,.... p n } denote the set of the number of predicted viewers at different regions for each video v . The broadcasters’ regions for the incoming videos at period t is denoted by B ( t ) = { r b 1 , r b 2 , ...r b m } . Due to the fact that some videos do not hav e any viewers near some cloud sites, let E ( v , r w ) present a binary variable, equal to 1, if video v has predicted viewers near the region r w , and 0 otherwise. W e consider renting S3 storage [15] servers at each cloud site. Three types of costs are taken into account: (1) the storage cost at each cloud site; (2) the migration cost of a video replica from one cloud site to another and (3) the cost of serving viewers. W e assume that the storage capacity can be provisioned based on the application demand. On allocation cloud site at region r a , let α r a be the storage cost per GB, which varies based on site location and the storage thresholds fixed by Amazon S3. For example, Amazon charges 0.023$ per GB for the first 50TB, while it charges 0.021$ when exceeding 500TB in the case of US East V irginia region [15]. Given that κ is the video size, the total storage cost S can be calculated as presented in Eq. 2. Giv en that η r b is the cost to migrate a copy of a video from the broadcaster region r b to allocation region r a , which is the data transfer cost from one cloud site to another per GB, the total migration cost M is calculated as presented in Eq. 3. Giv en that ω r a is the serving request cost from region r a , which is the data transfer cost from that region to the internet per GB, the total serving request cost R is calculated as presented in Eq. 4. The overall cost C to serve vie wers is shown in Eq. 5. p r w is the predicted number of viewers at region r w . S = X v ∈ V ( t ) X r a ∈ R α r a ∗ κ ∗ A ( v , r a ) (2) M = X v ∈ V ( t ) X r a ∈ R η r b ∗ κ ∗ A ( v , r a ) (3) R = X v ∈ V ( t ) X r a ∈ R X r w ∈ R ω r a ∗ κ ∗ p r w ∗ W ( v , r a , r w ) (4) C = S + M + R (5) Our objectiv e is to minimize the cost for period t as shown in Eq. 6: min A ( v ,r a ) W ( v,r a ,r w ) C (6) Subject to the following constraints: Every video is allocated by default in the broadcaster nearest cloud site. A ( v , r b ) = 1 ∀ v ∈ V ( t ) , ∀ r b ∈ B ( t ) (6a) A video v can be served from region r a to viewers at region r w , only if it is allocated at region r a . W ( v , r a , r w ) ≤ A ( v , r a ) ∀ v ∈ V ( t ) , ∀ r a ∈ R , ∀ r w ∈ R (6b) A video v can be served from region r a to r w only if there exists vie wers at r w . W ( v , r a , r w ) ≤ E ( v , r w ) ∀ v ∈ V ( t ) , ∀ r a ∈ R , ∀ r w ∈ R (6c) If there exists viewers for video v at region r w , they can only be served from one region. X r a ∈ R W ( v , r a , r w ) = E ( v , r w ) ∀ v ∈ V ( t ) , ∀ r w ∈ R (6d) The average serving request delay for each video should not exceed a threshold D . P r a ∈ R P r w ∈ R p r w ∗ d r a r w ∗ W ( v , r a , r w ) P r w ∈ R p r w ≤ D ∀ v ∈ V ( t ) (6e) Binary decision variables that can be set to 0 or 1. A ( v , r a ) , W ( v , r a , r w ) ∈ { 0 , 1 } (6f) The decision variable A ( v , r a ) is equal to 1, if video v is allocated in region r a , and 0 otherwise. While the decision variable W ( v , r a , r w ) is equal to 1, if viewers at region r w are served from region r a and 0 otherwise. The problem formulation notations are presented in T able I. 2) Proactive r esour ce allocation: the proposed proactiv e resource allocation algorithm is presented in Algorithm 1. In fact, at each period t, the system receives a set of incoming videos, which will be an input to the viewers predictiv e model. Based on the predicted viewers, the optimal number of allocation cloud sites and the nearest cloud site to serve the viewers will be decided by the optimizer . The storage resources at each cloud site is reserved based on the allocation decisions, and released for ended live videos from the previous periods. Moreover , the viewers are served from their closest cloud site based on the serving decisions. T ABLE I: Notations for the formalized problem. Notation Description V ( t ) Set of incoming live videos at period t R Set of regions B ( t ) Set of broadcasters regions for videos at period t r a Region of video allocation r w Region of serving r b Region of broadcasting P ( t ) Set of predicted viewers for liv e videos at period t P v Set of predicted viewers at different R for video v S U Set of storage used at each region W ( v, r a , r w ) Binary decision v ariable that indicates the serving site A ( v , r a ) Binary decision variable that indicates the allocation site E ( v, r w ) Binary variable that indicates viewers existence d r a r w Round trip delay between r a and r w RT T Matrix for round trip delay between the different R D Delay threshold κ V ideo size α r a Storage cost per GB at region r a η r b Migration cost per GB from broadcaster region r b ω r a Serving request cost per GB from r a S T otal storage cost M T otal migration cost R T otal serving request cost C Overall cost Algorithm 1 Proactiv e resources allocation 1: Input: R , { α 1 , ..., α n } , { η 1 , ..., η n } , { ω 1 , ..., ω n } , RT T , κ 2: Storage usage at each cloud site initialization: S U 1 = 0 , ..., S U n = 0 3: for t ∈ { 1 , .., T } do 4: - Receive videos informations V ( t ) and their broadcasters B ( t ) 5: - Run predictive model to predict P ( t ) for videos V ( t ) 6: - Derive optimal solution min A ( v,r a ) W ( v ,r a ,r w ) C as per Eq. 6 7: for j ∈ { 1 , .., n } do 8: - Update S U j based on allocation decisions A ( v , r a ) 9: - Release storage resources for ended videos from previous 10: periods. 11: - Serve viewers based on serving decisions W ( v , r a , r w ) I I I . P E R F O R M A N C E E V A L UA T I O N A. Simulation settings In this section, we ev aluate the performance of our system using the RF hourly predicted viewers of July 3, 2018 to get the hourly optimal resource allocation for T =24(hours) and t =1(hour). The number of hourly incoming videos, and the hourly predicted viewers used in our simulation are presented in Fig. 6. In our system, we assume that the video duration is 4 hours, which is the maximum video duration for a Facebook liv e video. W e assume that if a video is allocated in a set of cloud sites at period t , it will be allocated in the same cloud sites for the remaining time periods of streaming. Moreover , because video quality is out of the scope of this paper we assume that the viewers are served with the best video quality , where we set the video size κ to 0.738 Gbit. W e constructed our round trip time (R TT) matrix d r a r w by calculating the av erage R TT from the different cloud sites using [16] accessed on September 19, 2018. The storage and data transfer prices of Amazon S3 [15] are considered in our simulation to model α , ω and η . W e varied the latency thresholds constraints D for serving a video to 8.8ms, 60ms, 120ms, 171ms, 220ms and 371ms. 8.8ms is the latency needed to serve a viewer from its closest cloud region [8]. Fig. 6: Hourly incoming videos/ Hourly predicted viewers. B. Simulation r esults Fig. 7(a) shows that we can establish a trade-of f between the video access delay and the resource allocation cost. Indeed, the hourly optimal cost is high when the system is forced to serve the viewers from their region by setting the latency threshold to 8.8ms. Relaxing the threshold leads to minimizing the cost. Therefore, the content provider can sacrifice in terms of cost to enhance the QoE or the opposite based on his requirements. It is worth mentioning that the optimal cost is higher in some periods as opposed to others, because as illustrated in Fig. 6 the number of incoming videos and predicted viewers varies from period to another . In order to e valuate the total system cost ov er the 24 hours with various latency thresholds, we calculated the hourly total cost, as presented in Fig. 7(b). The hourly total cost is defined as the sum of the network cost at period t and the cost of storage of still running videos, which is presented in Eq. 7, giv en that S U n is the storage usage at region n until period t . Hourly total cost (t) = C ( t ) + X r a ∈ R α r a ∗ S U r a (7) The system total cost is calculated as shown in Eq. 8: System total cost = T X t =1 Hourly total cost(t) (8) Furthermore, we calculated the hits percentages, which represents the percentage of videos served from the same region of viewers as shown in Fig. 7(c). Setting the latency to 8.8ms resulted in hits percentage of 100% in e very hour , as all viewers will be served from their region. While it is in the range of 20% to 30% with 60ms latency threshold. Moreover , when the latency threshold was set to 120ms, 171ms, 220ms and 371ms, less than 20% of videos were served from the same region of vie wers. The hits percentage was very lo w with high latency thresholds, as the system is not forced to serve the viewers from their closest region. Finally , to ev aluate the accuracy of our resource allocation framew ork, we calculated the hourly average latenc y using the proactive serving decisions with variant latency thresholds D . In fact, we calculated the latency of serving the actual number of vie wers based on our proactiv e video allocation and we compared it to the latency deriv ed from the predictive model. The results as shown in Fig. 8 proved that the average latency to serve the actual viewers is very close to the average latency serving the predicted viewers. Moreov er , the av erage REFERENCES (a) Hourly optimal cost. (b) T otal cost vs latency thresholds. (c) Serving hits percentages. Fig. 7: Simulation results. Fig. 8: Predicted vs actual hourly average latency . latency to serve the actual vie wers did not exceed the latency thresholds D . I V . C O N C L U S I O N In this paper , we propose a proacti ve resource allocation framew ork. First, we adopt machine learning to build a pre- dictiv e model that captures the viewers number near each geo- ditributed cloud site. Then, based on the predicted results, we formulated our resource allocation model as an optimiza- tion problem to optimally allocate resources across the geo- distributed cloud sites based on the viewers proximity . For the future work, we plan to design a distributed proactiv e resource allocation framework. W e are also interested in implementing predictiv e models for the number of incoming liv e videos, the liv e video duration, the liv e videos viewing time and the computation resources. A C K N O W L E D G M E N T This publication was made possible by NPRP grant 8-519- 1-108 from the Qatar National Research Fund (a member of Qatar Foundation). The findings achieved herein are solely the responsibility of the author(s) R E F E R E N C E S [1] Cisco V isual Networking Index: Global Mobile Data T raf fic F orecast Update, 2016-2021 White P aper . Mar . 2017. U R L : https : / / www. cisco . com / c / en / us / solutions / collateral / service - provider / visual - networking - index - vni / mobile - white - paper - c11- 520862.html. [2] F acebook users worldwide 2018 . 2018. U R L : https : / / www . statista . com / statistics / 264810 / number - of - monthly - acti ve - facebook- users- worldwide/. [3] F acebook Statistics for 2018 . U R L : https : / /www. w ordstream . com/blog/ws/2017/11/07/facebook- statistics. [4] Athula Balachandran et al. “Developing a predictiv e model of quality of experience for internet video”. In: ACM SIGCOMM . V ol. 43. 4. 2013, pp. 339–350. [5] S Shunmuga Krishnan and Ramesh K Sitaraman. “V ideo stream quality impacts viewer behavior: inferring causality us- ing quasi-experimental designs”. In: IEEE/ACM T ransactions on Networking (TON) 21.6 (2013), pp. 2001–2014. [6] Y u W u et al. “Scaling social media applications into geo- distributed clouds”. In: IEEE/ACM T ransactions on Network- ing (TON) 23.3 (2015), pp. 689–702. [7] Qiyun He et al. “Coping with heterogeneous video contributors and vie wers in crowdsourced liv e streaming: A cloud-based approach”. In: IEEE T ransactions on Multimedia 18.5 (2016), pp. 916–928. [8] K Bilal, A Erbad, and M Hefeeda. “QoE-aware distributed cloud-based live streaming of multisourced multiview videos”. In: J ournal of Network and Computer Applications 120 (2018), pp. 130–144. [9] Stefano Petrangeli et al. “A machine learning-based framew ork for preventing video freezes in HTTP adaptive streaming”. In: Journal of Network and Computer Applications 94 (2017), pp. 78–92. [10] Anh Minh Le. “Improving Adaptiv e V ideo Streaming through Machine Learning”. In: (2018). [11] Guowei Zhu et al. “User Mapping Strategies in Multi-Cloud Streaming: A Data-Driv en Approach”. In: GLOBECOM, 2016 IEEE , pp. 1–6. [12] F acebookV ideosLive18 Dataset . U R L : https://sites.google.com/ view/f acebookvideosliv e18/home. [13] Amazon W eb Services— A WS . U R L : https: // a ws. amazon. com/ about- aws/global- infrastructure/. [14] Kilian W einberger et al. “Feature hashing for large scale multitask learning”. In: Pr oceedings of the 26th annual in- ternational confer ence on machine learning . A CM. 2009, pp. 1113–1120. [15] Cloud Stora ge Pricing — S3 Pricing by Re gion — Amazon Simple Storag e Service . U R L : https : / / aws . amazon . com / s3 / pricing/. [16] Global Ping Statistics . U R L : https://wondernetwork.com/pings.