1

AutoSight: Distributed Edge Caching in ShortVideo Network

Yuchao Zhang, Member, IEEE, Pengmiao Li, Student Member, IEEE, Zhili Zhang, IEEE Fellow,Bo Bai, Member, IEEE, Gong Zhang, Senior Member, IEEE,

Wendong Wang, Member, IEEE, Bo Lian, Ke Xu, Senior Member, IEEE

Abstract—Recent years have witnessed a rapid increase ofshort video traffic in Content Delivery Network (CDN). While thevideo contributors change from large video studios to distributedordinary end users, edge computing naturally matches the cacherequirements from short video network. But the distributed edgecaching exposes some unique characteristics: non-stationary useraccess pattern and temporal and spatial video popularity pattern,which severely challenge the edge caching performance. While theQuality of Experience (QoE) in traditional CDN has been muchimproved, prior solutions become invalid in solving the abovechallenges. In this paper, we present AutoSight, a distributed edgecaching system for short video network, which significantly boostscache performance. AutoSight consists of two main components,solving the above two challenges respectively: i) the CoStorepredictor, which solves the non-stationary and unpredictability oflocal access pattern, by analyzing the complex video correlations;ii) a caching engine Viewfinder, which solves the temporal andspatial video popularity problem by automatically adjustingfuture horizon according to video life span. All these inspirationsand experiments are based on the real traces of more than28 million videos with 100 million accesses from 488 serverslocated in 33 cities. Experiments results show that AutoSightbrings significant boosts on distributed edge caching in shortvideo network.

Index Terms—Short Video Network, Edge Computing, Con-tent Delivery Network, Caching Policy, Video Correlation, Non-stationary Access Pattern, Temporal and Spatial Characteristic.

I. INTRODUCTION

In recent years, many short video platforms are developingat an incredible speed (such as Kuaishou[1], Youtube Go[2],Instagram Stories[3] and so on), these platforms allow usersto record and upload very short videos (usually within 15seconds). As a result of the massive short videos uploaded

Yuchao Zhang, Pengmiao Li and Wendong Wang are with Beijing Univer-sity of Posts and Telecommunications, Beijing, China. Wendong Wang is alsowith the State Key Laboratory of Networking and Switching Technology.

Zhili Zhang is with University of Minnesota, Minneapolis, US.Bo Bai and Gong Zhang are with Huawei Company, Hong Kong, China.Bo Lian is with Kuaishou Company, Beijing, China.Ke Xu is with Tsinghua University and Beijing National Research Center

for Information Science and Technology, Beijing, China.Yuchao Zhang (yczhang@bupt.edu.cn) and Wendong Wang (wd-

wang@bupt.edu.cn) are the corresponding authors.The work of Yuchao Zhang was supported in part by the National Natural

Science Foundation of China under Grant 61802024, the Huawei Autonomousand Service 2.0 Project under Grant A2018185, and in part by the CCF-Tencent Rhinoceros Creative Foundation under Grant IAGR20190103. Thework of Ke Xu was in part supported by the National Key R&D Program ofChina with No. 2018YFB0803405, China National Funds for DistinguishedYoung Scientists with No. 61825204, Beijing Outstanding Young ScientistProgram with No. BJJWZYJH01201910003011.

by distributed users, caching problem is becoming morechallenging compared with the traditional centralized ContentDelivery Network (CDN) whose traffic is dominated by somepopular items for a long period of time. To handle scalability,fairness[4] and improve users’ Quality of Experience (QoE),the emerging edge computing naturally matches the demandfor distributed storage, the above mentioned platforms thushave resorted to employ edge caching servers to store anddeliver the massive short videos, so as to avoid that all requestshave to be fetched from the backend/origin server, whichusually introduces extra user-perceived latency.

There have been tremendous efforts towards better cachingperformance in traditional CDN, and these caching algorithmscan be classified into two categories. The simple but effectivereactive caching algorithms, such as First In First Out (FIFO),Least Recently Used (LRU), Least Frequently Used (LFU), k-LRU and their variants, and the proactive caching algorithms,such as DeepCache[5]. These prior solutions work well intraditional centralize-controlled CDN, but become invalid inthe emerging short video network, where is naturally im-plemented with distributed edge caching. Here are the twoessential differences.

(1) User access pattern is non-stationary. The basicassumption of traditional caching policies is the stationary useraccess pattern, i.e., recently requested or frequently requestedcontend in the past should be kept in cache because suchpolicies assume that these contents have greater chance ofbeing visited in the future. A study in [6] shows that in shortvideo network, popular contents always become expired veryquickly (within tens of minutes), indicating that the popularityin the past could not represent that in the future, and this is theroot cause of the failure of these reactive policies (see Section§III-A).

(2) Video popularity pattern has spatio-temporal char-acteristics. Our study on the workload of Kuaishou shows thatvideo popularity changes in different patterns during differenttime periods. For example, during peak hours, it takes less than1 hour for a popular video to become unpopular, while duringlate midnight, such invalidation takes more than 3 hours.Existing caching policies that try to predict future contentpopularity always focus on a fixed horizon, thus are failedin the edge caching scenarios.

To address the above challenges, this paper presents Au-toSight, a distributed caching mechanism that works in edgecaching servers for short video network. AutoSight allows edgeservers to retain respective local caching sight to adapt to

2

local video access pattern and video life span. AutoSight’sdistributed design is built on two empirical observations: (1)although the historical video access data is non-stationaryon individual edge server (see Figure 1), making it difficultto make popularity prediction, there is sufficient correlationsamong videos within the same edge server, because users tendto request related videos, contributing much cross visits inedge servers, and thus improves distributed predictions. (2)temporal and spatial video popularity pattern brings challengefor future sight of caching policy (see Figure 2), but distributeddesign allows adaptive future sights, enabling edge servers tomake decisions according to different video expiration speeds.

We have implemented a prototype of AutoSight and evaluateit with Kuaishou data traces. Experiments show that AutoSightachieves much higher hit rate compared with both reactivecaching policies and the state-of-the-art proactive cachingpolicies.

Our contributions are summarized as followed:• Characterizing Kuaishou’s workload of short video net-

work to motivate the need of distributed edge cachingpolicy.

• Presenting AutoSight, a distributed edge caching mech-anism working in edge caching servers for short videonetwork, which solves the problem of non-stationaryuser access pattern and temporal/spatial video popularitypattern.

• Demonstrating the practical benefits of AutoSight by realtraces.

II. RELATED WORK

Here we discuss some representative caching policies intraditional CDN and some related edge caching systems.

Existing caching policies. The most representative cachingpolicies such as FIFO, LRU, LFU and their variations aresimple but effective in traditional CDN, where the frequencyof content visits can be modeled as Poisson distribution. Underthese policies, the future popularity of a content is representedby the historical popularity. But in short video network, useraccess pattern is non-stationary and is no longer in Poissondistribution, so these policies become inefficiency in shortvideo network. The same problem exists in the TTL-based(Time-To-Live) caching policies. For example, in the caching(feedforward) networks, where the access pattern is Markovarrival process, [7] gives joint consideration about both TTLand request models, and drives evictions by stopping times. [8]sets an optimal timer to maximize cache hit rate, but it onlyworks in the case of Pareto-distributed access pattern and Zipfdistribution of file popularity, thus certainly becomes invalidin short video network. [9] propose two caches named f-TTLwith two timer, so as to filter out non-stationary traffic, but itstill relies on locally observed access patterns to change TTLvalues, regardless of the future popularity.

One of the promising attempts in recent years is learning-based proactive prediction-based policy. [5] trains a char-acteristics predictor to predict object future popularity andinteroperates with traditional LRU and LFU, boosting thenumber of cache hits. However, it looks a fixed length into

the future to predict object popularity (1-3 hours, 12-14 hoursand 24-26 hours), ignoring the temporal and spatial videopopularity pattern, thus cannot handle the varies life spansin short video network. Pensieve [10] trains a neural networkmodel as an adaptive bitrate (ABR) algorithm, it complementsour AutoSight framework, in the sense that the proposeddynamic control rules can give us a reference when facingthe various life span problem, but it ignores temporal patterninformation and only works in live streaming. Besides, [11]introduces reinforcement learning for making cache decisions,but it works in the case that user requests comply with theMarkov process, while the proposed AutoSight in this papercan work under arbitrary non-stationary user access patterns.

Edge caching systems. Edge computing was proposed toenable offloading of latency-sensitive tasks to edge serversinstead of cloud, and has achieved rapid development in manyares such as 5G, wireless, mobile networks [12] and videostreaming [13]. [14] studies the content placement problem inedge caching to maximize energy efficiency, the analysis inthis work gives us references to design our AutoSight networktopology, but what we considered under such topology is thecaching of short video network rather than energy-saving. [13]proposes a geo-collaborative caching strategy for mobile videonetwork, suggesting that joint caching over multiple edges canimprove QoE, which provides strong proof for our AutoSightdesign. While this paper tries to reveal the characteristics ofdifferent mobile videos, we focus the short video networkon edge servers with unique user access pattern and videopopularity pattern. [15] considers a network caching setup,comprising a parent node connected to several leaf nodesto serve end user file requests, proposes an efficient cachingpolicy leveraging deep RL, and shows capable of learning-and-adapting to dynamic evolutions of file requests, and cachingpolicies of leaf nodes, which provides strong support to ourAutoSight design.

III. BACKGROUND

Before introducing the caching problem, we start by char-acterizing the short video network, illustrating the essentialdifferences with traditional CDN, then we show the limitationsof existing schemes and draw lessons from real-world traces toinform the design of AutoSight. The findings are based on realdatasets from Kuaishou’s caching network collected during 4days, from 9 Oct. 2018 to 12 Oct. 2018.

A. Characteristic and Challenges

Short video platforms allow users to upload seconds ofvideos (usually within 15 seconds) to the network, suchconvenience on content uploading and accessing finally leadsto a revolution in the way network works and the way videosare cached.

Non-stationary user access pattern: In traditional cachingsystem, user access pattern is stationary, in other words, thepopular contents at present still have a higher likelihood ofreceiving more accesses in the next moment. But a study in[6] shows that due to the short life cycle of short videos,popular contents always become expired very quickly (within

3

1 2 3 4 5 602468

1 01 21 41 61 8

Acce

ss Nu

mber

T i m e ( m i n )

v 1 v 2

Figure 1: Non-stationary video access pattern

0 5 0 1 0 0 1 5 0 2 0 00

1 02 03 04 05 06 07 0

Acce

ss Nu

mber

T i m e ( m i n )

v 1 v 2

Figure 2: Temporal and spatial video popularity pattern

tens of minutes), indicating that the popularity at present couldnot represent that in the future. Figure 1 shows the requestsnumber of two particular videos v1 and v2, from which wecan see the non-stationary of the user access pattern (withsudden increase and decrease). In the first two minutes, v1receives less requests while v2 receives more, but the accesspattern reverses at 2nd min (request burst on v1 while valley onv2). There is also a similar reverse at the 5th min. Therefore,traditional caching policies become invalid due to the non-stationary access pattern, and a new popularity predictionalgorithm in short video network is in urgently needed.

Spatio-temporal video popularity pattern: In short videonetwork, video popularity changes in various patterns. Accord-ing to our study on Kuaishou workloads, it takes less than 1hour for a popular video to become unpopular during peakhours, while such invalidation takes more than 3 hours duringlate midnight. Figure 2 shows the life spans of 2 videos froman edge server but appear in different time periods, which aresignificantly different from each other. This figure illustratesthe variability of video life span during different time period,and similar situations also occur among different edge servers,which means that in some temporal and spatial, videos getexpire at different speeds, this finding motivates us to equip our

caching engine with auto-adjusted future horizons, and herewe present the definition of Future Horizon (The algorithmViewfinder will be described in Section §IV-C.)

Definition 1: Future Horizon. It represents how far into thefuture to plan the caching, it is the length of future time ∆t,during which period of time, the predicted video popularitycould represent the popularity of a current video.

B. Limitations of existing solutions

Realizing caching improvement of short video networkhas some complications. As a first order approximation, weplanned simply borrow existing techniques from traditionalCDN. But the above two characteristics results in inefficiencyof existing approach that will be described below.

Key Observation 1: The non-stationary access pattern makesthe heuristic reactive caching policy invalid.

Explanation: In the traditional stationary network scenario,heuristic reactive caching policies, such as LRU and LFU, canwork well by replacing the least used videos. But in shortvideo network with non-stationary access pattern, as the caseshown in Figure 1, those policies will eject v1 at the 2nd min,because v1 is less recently used and less frequently used in thefirst 2 minutes. However, v1 got more accesses than v2 duringthe 3rd min, resulting in less caching hit and lower overall hitrate. This is because under the non-stationary access pattern,video popularity in the past could not represent that in thefuture, and the ejected video at the 2nd minute should be v2rather than v1. This simple example shows the invalidity ofthe heuristic reactive caching policies.

Key Observation 2: Spatio-temporal video popularity pat-tern on different edge servers and at different time periodsmake the fixed-horizon proactive caching policy inefficient.

Explanation: When making video popularity prediction,existing learning-based proactive caching policies always looka fixed length into the future, i.e., future horizon = ∆t, whichis a fixed length window, and their output is a sequenceof k future popularity probabilities, where k represents thenumber of probabilities to predict during the future ∆t time.Such policies work well in traditional caching systems, butin the short video network scenario, the average video lifespan on different edge servers or during different time periodssignificantly varies from each other, there is no “one size fitsall”. In the case shown in Figure 2, if future horizon ∆t isset to 2 hours, even through those policies can have 100%prediction accuracy, the one would be ejected at 50th min isv1 rather than v2, because the predicted popularity of v1 isless than that of v2 in the next 2 hours. But in reality, v1 ismuch more popular than v2 in the near future. Ejecting v1 willobviously downgrade the performance. This example showsthe invalidity of the fixed-horizon proactive caching policy.

IV. AutoSight DESIGN

The core of AutoSight is a distributed caching algorithmthat makes adaptive caching for edge servers. There aretwo main components in AutoSight: a correlation analyzernamed CoStore, which solves the non-stationary access patternproblem by analyzing the videos correlations, and a caching

4

S1S2

YCoStore

Viewfinder∆t

Access seq.Life span

Cache Mem

AutoSight① ② ③ ④

Visits from users and other servers①

Cross visits to other edge servers②

Download videos from backend server③

Upload local videos to backend server④

①

Figure 3: The design of distributed edge caching and the framework of AutoSight

�����������������������������������

����

����

����

����

����

����������

��������

�����������������������������

Figure 4: The power of Viewfinder

engine named Viewfinder, which solves the temporal/spatialpopularity problem by automatically adjusting horizon to adaptto different edge caching server during different time periods.

A. System Overview

AutoSight takes an explicit stance that it works the dis-tributed edge caching servers to handle the non-stationary useraccess pattern and temporal/spatial video popularity pattern,significantly boosting the cache hit rate in the setting of shortvideo network. AutoSight uses a correlation-based predictorthat predicts the number of times a video will be requested bymaking real time analysis of videos’ cross visits, and uses acaching engine with adaptive future horizon to make cachingdecisions. The framework of AutoSight is shown in Figure 3.The workflow can be summarized as follows.

• AutoSight is implemented in edge servers. It receives bothvisits from local users and cross visits from other edge

�����������������������������������

���

���

���

���

��������

��������

����������������������� �

Figure 5: Hit rate comparison

servers.• When the requested videos are stored in local cache

memory (called “cache hit”), the edge server sends backthe videos directly. When the requested videos are notstored in local cache, the edge server makes a cross visitto other edge servers, and then sends the obtained videosto the requestor.

• When the requested video are not stored in other edgeservers either (called “cache miss”), such videos wouldbe downloaded from backend servers (where all videosare stored).

• To ensure the video completeness of the back-end servers,all of the uploaded videos would be stored in backendservers.

The ultimate goal of AutoSight is to increase the numberof cache hit and eliminate cache miss, so as to reducethe request response time. We therefore define the hit rate

5

by cache hit/(cache hit + cache miss). In the followingCoStore and Viewfinder design, we try to improve the overallhit rate.

B. Correlation-based predictor: CoStore

Although short video network has non-stationary popularity,i.e., video popularity in the past could not represent that in thefuture, we also found that past video correlations could providemuch information to make future popularity prediction [6].This is because the video correlations in short video networkis much informative due to the frequent cross-visits. For aparticular video with low popularity, it possibly gets higherpopularity if its correlated videos already have high popularity.

Inspired by the long short term memory (LSTM) networkthat has already shown its dominance in natural languageprocessing (NLP), machine translation and sequence predic-tion, CoStore is built on LSTM, using video correlationsas input features, and predicting request numbers within thefuture horizon (Section §IV-C). In particular, for video vi attime t, the input consists of two sets of access sequence:S1 = {r1vi , r

2vi , ..., r

tvi} and S2 = {r1vj

, r2vj , ..., rtvj}, where

rkvi denotes the request number of vi at time k, and vj is themost related video to vi at time t. The output of CoStore isthe expected request number of vi during the future horizon∆t.

C. Caching engine: Viewfinder

As edge caching servers are experiencing temporal andspatial video popularity pattern (shown in Section §III-B),inappropriate future horizon ∆t (too short-sighted or too long-sighted) always leads to inefficient or even wrong cachingdecisions. We therefore design Viewfinder, which can adjustfuture horizons automatically. Viewfinder aims at finding asuitable horizon for future prediction. In different time frames,it outputs different horizons through on a reinforcement algo-rithm, and this horizon is the time window for CoStore topredict video popularity.

The challenge here is that there are too many options to ex-plore (in the granularity of seconds/minutes), which introducesunacceptable overhead. Viewfinder changes it into a classifi-cation problem that chooses ∆t from a predefined set: ∆T ={60min, 120min, 160min, 180min, 200min, 360min}.

This significantly reduces the computation overhead, andthe experiment results show that Viewfinder works well in edgecaching servers, disclosing that the quicker videos get expired,the shorter-sight the caching policy should be.

V. EVALUATION

In this section, we evaluate our approach AutoSight usingreal traces, and show the results of applying AutoSight on themversus the existing representative policies.

A. Experiment setting.

Algorithms: We compare AutoSight with four existing so-lutions: FIFO, LRU, LFU and LSTM-based prediction schemewithout the auto-adjusted future horizon.

Datasets: We analyze the traces from 2 cities with1,128,989 accesses to 132,722 videos in 24 hours. Each traceitem contains the timestamp, anonymized source IP, videoID and url, file size, location, server ID, cache status andconsumed time. Thus we can deploy and evaluate differentcaching policies.

B. Performance comparison

We first provide dataset introduction and analysis aboutthese 2 cities, then we compare the overall hit rate on edgeservers among 5 different caching policies, after that, we lookinto the proposed AutoSight and show the power of Viewfinderwith auto-adjusted future horizons.

The power of Viewfinder. To evaluate the effect of thecaching engine Viewfinder with the auto-adjusted future view,we show the cache hit rate with Viewfinder set to fixed ∆t,Figure 4 shows the corresponding cache hit rate under differentfuture horizons. The optimal value for each period various withtime, i.e., during late midnight when video life span is longer,Viewfinder tends to be long-sighted, while during the leisuretime (e.g., 20:00-21:00pm) when video life span is shorter,Viewfinder also tends to be short-sighted. These results furtheremphasize the necessity of Viewfinder with adaptive futurehorizon.

Overall cache hit rate. As analyzed in Section §III-B, thenon-stationary access pattern would make the reactive cachingpolicy inefficient, and the temporal/spatial video popularitypattern also invalidates learning-based policies with fixed-length future horizon. Figure 5 shows the overall cache hitrate of applying the 5 policies. AutoSight outperforms all theexisting algorithms.

VI. CONCLUSION

In this paper, we analyze Kuaishou dataset and use trace-driven experiments to motivate and investigate edge cachingperformance for short video network. We first disclose thecharacteristics on non-stationary user video access pattern andtemporal/spatial video popularity pattern, and illustrate theinvalidation of existing caching policies by giving two realcases. Then we design AutoSight, a distributed edge cachingsystem for short video network, with CoStore and Viewfinder.Results show that enabling AutoSight in edge caching serverscould significantly outperforms the existing algorithms.

REFERENCES

[1] kuaishou, “Kuaishou,” https://www.kuaishou.com, 2019.[2] Y. Go, “Youtube go,” https://youtubego.com, 2019.[3] I. Stories, “Instagram stories,” https://storiesig.com, 2019.[4] Y. Zhang, K. Xu, H. Wang, Q. Li, T. Li, and X. Cao, “Going fast

and fair: Latency optimization for cloud-based service chains,” IEEENetwork, vol. 32, no. 2, pp. 138–143, 2017.

[5] A. Narayanan, S. Verma, E. Ramadan, P. Babaie, and Z.-L. Zhang,“Deepcache: A deep learning based framework for content caching,” inProceedings of the 2018 Workshop on Network Meets AI & ML. ACM,2018, pp. 48–53.

[6] Y. Zhang, P. Li, Z. Zhang, B. Bai, G. Zhang, W. Wang, and B. Lian,“Challenges and chances for the emerging short video network,” in 2019IEEE Conference on Computer Communications (Infocom). IEEE,2019, pp. 1–2.

6

[7] D. S. Berger, P. Gland, S. Singla, and F. Ciucu, “Exact analysis of ttlcache networks: the case of caching policies driven by stopping times,”ACM SIGMETRICS Performance Evaluation Review, vol. 42, no. 1, pp.595–596, 2014.

[8] A. Ferragut, I. Rodrı́guez, and F. Paganini, “Optimizing ttl caches underheavy-tailed demands,” in ACM SIGMETRICS Performance EvaluationReview, vol. 44, no. 1. ACM, 2016, pp. 101–112.

[9] S. Basu, A. Sundarrajan, J. Ghaderi, S. Shakkottai, and R. Sitaraman,“Adaptive ttl-based caching for content delivery,” ACM SIGMETRICSPerformance Evaluation Review, vol. 45, no. 1, pp. 45–46, 2017.

[10] H. Mao, R. Netravali, and M. Alizadeh, “Neural adaptive video stream-ing with pensieve,” in Proceedings of the Conference of the ACM SpecialInterest Group on Data Communication. ACM, 2017, pp. 197–210.

[11] A. Sadeghi, F. Sheikholeslami, and G. B. Giannakis, “Optimal andscalable caching for 5g using reinforcement learning of space-timepopularities,” IEEE Journal of Selected Topics in Signal Processing,vol. 12, no. 1, pp. 180–190, 2018.

[12] X. Li, X. Wang, P.-J. Wan, Z. Han, and V. C. Leung, “Hierarchical edgecaching in device-to-device aided mobile networks: Modeling, optimiza-tion, and design,” IEEE Journal on Selected Areas in Communications,vol. 36, no. 8, pp. 1768–1785, 2018.

[13] G. Ma, Z. Wang, M. Zhang, J. Ye, M. Chen, and W. Zhu, “Understandingperformance of edge content caching for mobile video streaming,” IEEEJournal on Selected Areas in Communications, vol. 35, no. 5, pp. 1076–1089, 2017.

[14] F. Gabry, V. Bioglio, and I. Land, “On energy-efficient edge caching inheterogeneous networks,” IEEE Journal on Selected Areas in Commu-nications, vol. 34, no. 12, pp. 3288–3298, 2016.

[15] G. W. A. Sadeghi and G. B. Giannakis, “Deep reinforcementlearning for adaptive caching in hierarchical content deliverynetworks,” IEEE Transactions on Cognitive Communications andNetworking, vol. abs/1902.10301, 2019. [Online]. Available: http://arxiv.org/abs/1902.10301

Yuchao Zhang received her PhD degree from Com-puter Science Department at Tsinghua Universityin 2017. Before that, she received B.S. degree inComputer Science and Technology at Jilin Univer-sity in 2012. Her research interests include largescale datacenter network, content delivery network,data-driven network and edge computing. She iscurrently with the Beijing University of Posts andTelecommunications as an Assistant Professor.

Pengmiao Li received M.A. in computer scienceand technology department, North China Universityof Science and Technology, Tangshan, China, in2018. She is currently a Ph.D. candidate in softwareengineering from Beijing University of Posts andTelecommunications, Beijing, China. Her currentresearch interests include CDN and edge computing.

Zhili Zhang graduated with B.S. in Computer Sci-ence with highest distinction from Nanjing Univer-sity, Nanjing, China. Then received his M.S. andPh.D. degrees in computer science from the Univer-sity of Massachusetts, Amherst in 1992 and 1997.He joined the Department of Computer Science andEngineering at University of Minnesota in January1997, where he is now a Full Professor. Zhi-Li is aFellow of IEEE.

Bo Bai (S’09M11SM’17) received the B.S. de-gree (Hons.) from the School of CommunicationEngineering, Xidian University, Xian, China, in2004, and the Ph.D. degree from the Departmentof Electronic Engineering, Tsinghua University, Bei-jing, China, in 2010. He was a Research Assistantand a Research Associate with the Department ofElectronic and Computer Engineering, Hong KongUniversity of Science and Technology, from 2009 to2010, from 2010 to 2012, respectively. From 2012to 2017, he was an Assistant Professor with the

Department of Electronic Engineering, Tsinghua University. He is currently aSenior Researcher with the Theory Lab, 2012 Labs in Huawei TechnologiesCo., Ltd., Hong Kong.

Gong Zhang is a Principal Member in Theory Labat Huawei. His research spans networks, distributedsystem and communication system architectures. Hehas contributed to more than 90 patents. He hadbeen a product development team leader of smartdevices in 2002 to pioneer new consumer businessfor Huawei. He then started to lead the researchon future Internet and cooperative communicationin 2005. He has been in charge of the AdvanceNetwork Technology Research Department, leadingresearch on future network, distributed computing,

database system, and data analysis since 2009. His recent research focuses onfuture network.

Wendong Wang (M’05) received his B.E. and M.E.degrees both from the Beijing University of Postsand Telecommunications, China, in 1985 and 1991,respectively, where he is currently a Full Professor inState Key Laboratory of Networking and SwitchingTechnology. He has published over 200 of papersin various journals and conference proceedings. Hiscurrent research interests are the next generation net-work architecture, network resources managementand QoS, and mobile Internet. He is a member ofIEEE.

Bo Lian received his B.S. from Northwestern Poly-technical University, Xi’an, China. He currentlyworks as a senior R&D engineer in Kuaishou Com-pany. Before that, he had worked in Tencent for morethan 10 years. His research interests include videostreaming, CDN, recommendation, and network in-frastructure.

Ke Xu (M’02-SM’09) received his Ph.D. from theDepartment of Computer Science and Technologyof Tsinghua University, where he serves as fullprofessor. He serves as Associate Editor of IEEE In-ternet of Things Journal and has guest edited severalspecial issues in IEEE and Springer Journals. His re-search interests include next generation Internet, P2Psystems, Internet of Things, network virtualization,and network economics. He is a member of ACM.