Optimizing Streaming Server Selection for CDN-delivered Live Streaming
0 likes239 views
This document discusses the optimization of streaming server selection in content delivery networks (CDNs) for live streaming applications. It highlights the unique challenges posed by live streaming compared to traditional web content delivery, such as the need for real-time processing and high bandwidth. The proposed solution, called 'sticky-dns,' enhances geo-DNS functionality to select optimal origin servers based on complete delivery paths, reducing transit costs while improving viewer experience.
![Optimizing Streaming Server Selection for
CDN-delivered Live Streaming
Zhenyun Zhuang1 and Shun Kwok1,2
1 Ying-Da-Ji Tech., 603 Overseas High-Tech Venture Park, Shenzhen 518057, China
2 Corresponding Author
Abstract. Content Delivery Networks (CDNs) have been widely used to deliver
web contents on today’s Internet. Gaining tremendous popularity, live streaming
is also increasingly being delivered by CDNs. Compared to conventional static
or dynamic web contents, the new application type of live streaming exposes
unique characteristics that pose challenges to the underlying CDN infrastructure.
Unlike traditional web-objects fetching, which allows Edge Servers to cache con-
tents and thus typically only involves Edge Servers for delivering contents, live
streaming requires real-time full CDN-streaming paths that span across Ingest
Servers, Origin Servers and Edge Servers.
DNS is the standard practice for enabling dynamic assignment of servers. GeoDNS,
a specialized DNS system, provides DNS resolution by taking into account the
geographical locations of end-users and CDN servers. Though GeoDNS effec-
tively redirects users to nearest CDN Edge Servers, it may not be able to select
the optimal Origin Server for relaying a live stream to Edge Servers due to the
unique characteristics of live streaming. In this work, we consider the require-
ments of delivering live streaming with CDN, and propose advanced design for
selecting optimal Origin Streaming Servers in order to reduce network transit
cost and increase viewers’ experience. We further propose a live-streaming spe-
cific GeoDNS design for selecting optimal Origin Servers to serve Edge Servers.
1 Introduction
Today’s web contents are increasingly being delivered by Content Delivery Networks
(aka Content Distribution Networks, or CDNs) [1,2]. a typical CDN infrastructure con-
sists of two types of servers based on their functionalities: Ingest Servers for accepting
customers’ web contents; and (ii) Delivery Servers for serving end web users. Live
streaming is becoming increasingly popular, and it is expected to be more pervasive in
the near future. Major CDN providers such as Akamai [1] and Level3 [2] are support-
ing delivering live streaming with their CDNs. When supporting live streaming, CDN
providers provide a layered infrastructure which consists of three types of streaming
servers: (i) Ingest Streaming Servers (ISS) which accept customers’ live streams; and
(ii) Origin Streaming Servers (OSS) which transit live streams from ingest servers to
Edge Streaming Servers; and (iii) Edge Streaming Servers (ESS) that directly serve the
end viewers. Depending on the scale of a CDN, the number of Origin Streaming Servers
varies from several to hundreds, and Edge Streaming Servers can be even thousands.
Unlike other traditional web-based applications, live streaming is associated with
unique characteristics and in turn presents special challenges to CDN infrastructure.](https://image.slidesharecdn.com/2012-160421130646/85/Optimizing-Streaming-Server-Selection-for-CDN-delivered-Live-Streaming-1-320.jpg)
![(a) Default Delivery (b) Optimized Delivery
Fig. 1. Live streaming delivered with CDN
First, live streaming has to be delivered from the CDN Ingest Server to Edge Servers in a
real-time fashion, which is in sharp contrast with traditional CDN-based web delivering
where typically only the last hop (i.e., from CDN Edge Servers to web users) needs
to be optimized. Such uniqueness requires the optimization of the entire CDN transit
path which consists of two transit hops of Ingest-Origin and Origin-Edge. Second, live
streaming often features high-bandwidth requirement. A typical live videos can easily
consumes up to 3 Mbps per stream, thus there is a higher need for reducing transit
cost on CDN networks. Third, live streaming typically is time-sensitive, and the CDN
infrastructure design needs to consider such timeliness by choosing the shortest path
from the streaming source to the end users.
After a live stream is ingested into the CDN infrastructure through an Ingest Server,
when an end viewer requests to play a live stream, an appropriate Edge Server and
an Origin Server need to be chosen, so that the player pulls the stream from the Edge
Server and the Edge Server pulls the stream from the Origin Server. DNS is the stan-
dard practice for making the above server selections. GeoDNS [3], typically used by
CDNs, is a specialized DNS service by taking into account geographical locations of
CDN Origin/Edge servers and end users. It translates host names into IP address and
redirects the requests to optimal Origin/Edge servers. Though many factors such as
server loads contribute to this selection process, the critical factor is the proximity (i.e.,
geographical) information. For instance, in a CDN that deploys Edge Servers on both
coasts of USA, an end user coming from California most likely will be served by an
Edge Server located in west coast rather than east coast. Traditional GeoDNS solutions
work effectively by offering the end users lower response time and higher throughput.
Inside CDN infrastructure, GeoDNS is also used to determine the optimal Origin Server
to serve Edge Server requests, with the goal of offering lower transit cost and response
time.
When supporting live streaming, GeoDNS is used for dynamical assignment of Ori-
gin Streaming Servers in response to Edge Streaming Server requests. In this process,](https://image.slidesharecdn.com/2012-160421130646/85/Optimizing-Streaming-Server-Selection-for-CDN-delivered-Live-Streaming-2-320.jpg)
![that though physical distance is often used as the path cost, cost metrics may also in-
clude network delay, price, server load, reliability, loss rate, etc. Since different cost
metrics often belong to distinct dimensions and they are often prioritized. Though how
to prioritize them varies with different CDN providers and deployment scenarios, in
this work we provide the following guidelines: (i) Server Load is prioritized over all
other metrics. An Origin server is put into the DNS pool (i.e., available to be returned
by Sticky-DNS) only if it can take more loads (i.e., serving more requests); (ii) Loss
rate has to be under certain threshold. Since most live streaming (e.g., RTMP, HTTP)
use TCP to transmit, large loss rate cannot sustain a reasonable level of performance;
(iii) Reliability has to be above certain threshold. Live streaming requires a real-time
transmission path, thus reliability is important during the stream playback; (iv) Pricing
is prioritized over network delay and physical distance. Though pricing is often cor-
related with the latter two metrics, it is prioritized over them if they are not identical;
(v) Network delay and physical distance. The impact of these two metrics are buffering
time and playback delay. In reality, they can be used as the substitute for pricing metric
when the latter is not available or hard to define.
Thus, with the above guidelines in mind, we propose to enforce the cost determina-
tion by keeping track of the following state information.
– An array of server load for the Origin Servers, denoted by SL[M], where M is the
number of Origin Servers. SL[i] represents the current load of Origin Server SSOi;
– A 2-D array of loss rate for each Origin-Edge path, denoted by LR[M][N], where
M is the number of Origin Servers and N is the number of Edge Servers. LR[i][j]
represents the loss rate of the path between Origin Server SSOi and Edge Server
SSE j; Another array of loss rate for each Ingest-Origin pair, denoted by LRI[M],
assuming a single Ingest Server;
– A 2-D array of reliability for each Origin-Edge path, denoted by RE[M][N], where
M is the number of Origin Servers and N is the number of Edge Servers. RE[i][j]
represents the reliability of the path between Origin Server SSOi and Edge Server
SSE j; Another array of reliability for each Ingest-Origin pair, denoted by REI[M];
– Pricing is represented by PR[M][N] for each pair of Origin-Edge; Another array of
PRI[M] for each pair of Ingest-Origin;
– Similarly, network delay and physical distance are represented by ND[M][N] and
PD[M][N], respectively. Two other arrays of NDI[M] and PDI[M] are also used.
4.2 Server Cost Probing (SCP)
SCP updates the various cost state by proactively or reactively monitoring/probing the
network. Specifically,
– Origin Server Load (SL[M]). The load can be obtained by installing services which
monitor CPU/Memory utilization;
– Loss rate (LR[M][N] and LRI[M][N]). Loss rate monitoring between servers can
be obtained by observing existing connections’ behaviors (e.g., TCP loss recovery
behaviors) or proactively sending test packets. When a network path (e.g., Ingest-
Origin-Edge) consists of two network hops, the aggregate loss rate of path SSEi −
SSO j −SSI can be calculated as LRij = (1−(1−LR[j][i])(1−LRI[j]));](https://image.slidesharecdn.com/2012-160421130646/85/Optimizing-Streaming-Server-Selection-for-CDN-delivered-Live-Streaming-8-320.jpg)
![Core Algorithm:
Initialize all cost states;
For an incoming new viewer Vn requesting Lm, determine the Edge Server SSEmn;
1 If SSEmn is already serving Lm;
2 Return; //No Origin Server request since the Edge Server is serving Lm now;
3 Else
4 Determine the candidate Origin Server set based on load, loss rate, etc.;
5 For every Origin Server SSOi, if (SL[i] ≤ α and LRj,k ≤ β and REj,k ≥ γ, where
6 α, β,γ are the respective threshold values), SSOi is a candidate Origin Server;
7 Determine the optimal Origin Server SSOq that has the smallest aggregate
8 cost of CIO and COE for SSEmn;
9 NV[q][m]++;
10// Update AD[q][m];
11 If (NV[i][j] ≥ 1)
12 AD[i][j] = 0;
13 Else
14 If ( WB[i][j] ==0 ): AD[i][j] = initial cost;
15 Else : AD[i][j] = initialcost
WB[M][T]+1
If Edge Server SSEi is disconnecting from SSOj for live stream Lm:
1 // Update NV[I][J] and AD[I][J];
2 NV[j][m]–;
3 If (NV[j][m] ≥=1): AD[j][m] = Initialcost
(NV[j][m]+1) ;
4 Else : AD[j][m] = {initialcost}.
Fig. 5. Core Algorithm
– Similarly, reliability (RE[i][j] and REI[i][j]) can be obtained by monitoring the path
up/down time. The aggregate reliability of path SSEi −SSOj −SSI can be calculated
as REij = (RE[j][i]∗REI[j]);
– Pricing (PR[M][N] and PRI[M][N]) can be adjusted by business logics. Whenever
the CDN operators change the network pricing rates, the pricing used by Sticky-
DNS can be correspondingly adjusted.
– Network delays (ND[M][N] and NDI[M]) and physical distances (PD[M][N] and
PDI[M]) are easy to obtain with known techniques.
4.3 Edge-Origin Connection Monitoring (ECM)
ECM is used to keep the server streaming state maintained by OSS fresh, in the sense
that the additional cost associated with each Origin Server is up-to-date, and the popu-
larity state of a stream on all Origin Servers is up-to-date. For all Origin Servers, there is
a server streaming state. The server streaming state is an array, and each element corre-
sponds to an Origin Server and is defined as the Additional Cost of using the particular
Origin Server to serve a stream. The state can be realized by a 2-D array of AD[M][T],](https://image.slidesharecdn.com/2012-160421130646/85/Optimizing-Streaming-Server-Selection-for-CDN-delivered-Live-Streaming-9-320.jpg)
![(a) Original, no viewers (a) E1 is on (a) E2 is on
Fig. 6. An illustrative example (I)
Source: S
Ingest: I
Origin: O1
Origin: O2
Edge: E1
Edge: E2
Edge: E3
(a) E3 is on (a) E4 is on
Fig. 7. An illustrative example (II)
where M is the number of Origin Servers and T is the number of streams. AD[i][j]
records the additional cost of using Origin Server SSOi to serve stream Lj. The values
of AD[M][T] are based on whether an Origin Server serves a stream or not.
Two other 2-D arrays are used: NV[i][j] records the number of Edge Servers Origin
Server SSOi is serving stream Lj to. WB[M][T] is used to indicate the stream popular-
ity on a particular Origin Server. The indicator uses the number of “Would-be” Edge
Servers for each Origin Server. ECM monitors the connecting/disconnecting events be-
tween Origin and Ingest Servers for all live streams. Whenever the events occur, CVM
updates NV[M][T], WB[M][T] and AD[M][T] correspondingly.
4.4 Sticky DNS Engine (SDE)
Dynamic DNS Engine (SDE) performs the actually computation and comparison of
costs, then selects the most appropriate Origin Server given a live streaming DNS re-
quest. The algorithms are shown in Figure 5. SDE handles the joining of an Edge Server
(i.e., begins to serve a stream) and the leaving (i.e., begins to drop a stream) separately.
For both events, it updates the AD and NV states.](https://image.slidesharecdn.com/2012-160421130646/85/Optimizing-Streaming-Server-Selection-for-CDN-delivered-Live-Streaming-10-320.jpg)
![0.00
0.50
1.00
1.50
2.00
2.50
3.00
C D F H J K N O Q T
Default Optimized
Fig. 9. Stream Delivery Paths
packages, CDN providers may pay ISPs for the bandwidth used. We assume each
live stream is 3Mbps. In the above scenario where only 4 Origin Servers are deliver-
ing the same live stream, the saved bandwidth per day would be 0.20 TB*Kmiles (or
73.74 TB*KMiles per year). With 2000 streams, the saved bandwidth would be 400
TB*Kmiles (or 147.47 PB * Kmiles per year). Simply assuming the transit cost of
0.10 per GB/Kmile, optimized-DNS can help save $14.74 Million per year with 2000
streams, or about $7K per stream, as shown in Table 1.
6 Related Work
Content Delivery Networks (CDNs) have been carried out by various providers to ex-
pedite the web access [1]. The techniques used by CDNs for delivering conventional
web contents are explained in related writings [4]. Despite the popularity and perva-
siveness of CDNs, many research problems and system-building challenges persist for
optimizing CDN infrastructure and addressing challenges associated supporting various
types of applications. For instance, as we demonstrated in this work, how to efficiently
support the application of live streaming with CDNs deserves more study.
Though various network architectures have been proposed and utilized for different
deployment scenarios with the rapid development of Internet technology [5, 6], CDN
providers typically apply layered structure (e.g., ingest-layer, origin-layer and edge-
layer) due to the caching-property of delivering static web contents tracing back to
Day-1 of CDN industry. Realizing that CDNs and P2P can complement each other,
Table 1. Bandwidth/Cost Saving of Sticky-DNS
Number of streams Daily BW Saving Yearly BW Saving Yearly Cost Saving
1 0.20 TB 73.74 TB $7K
2000 400 TB 147.47 PB $14.74M](https://image.slidesharecdn.com/2012-160421130646/85/Optimizing-Streaming-Server-Selection-for-CDN-delivered-Live-Streaming-13-320.jpg)
![many research efforts also attempt to integrate these two realms into one [7]. Unlike
these works that combine the merits of both CDNs and P2P, our work takes an inside
perspective and attempts to refine CDN delivery infrastructure.
Live streaming has been studied from various perspectives [8–10]. The unique prop-
erties of live streaming, coupled with CDN-assisted delivery, justifies a specialized de-
sign of CDN infrastructure that specifically serves live streaming and saves CDN transit
cost. One work [11] proposes to combine the benefits of CDN and P2P to enhance live
streaming. Another work [12] proposes to build a flat-layer infrastructure (as opposed
to the traditional 3-layer infrastructure) for supporting live streaming with CDNs. To
our best knowledge, this work is the first to consider and address the problem of saving
transit cost for delivering live streaming with traditional layered-infrastructure CDNs.
7 Conclusion
In this work, we optimize CDN streaming server selection to better support live stream-
ing with CDN. The proposed design dynamically selects streaming servers based on the
popularity of the live stream. Significantly improving both the CDN transit cost saving
and end viewers’ experience, the design can be deployed with a specialized GeoDNS.
References
1. “Akamai technologies,” http://www.akamai.com/.
2. “Level 3 communications, llc.” http://www.level3.com/.
3. “Bind geodns,” http://www.caraytech.com/geodns/.
4. D. C. Verma, S. Calo, and K. Amiri, “Policy based management of content distribution net-
works,” IEEE Network Magazine, vol. 16, pp. 34–39, 2002.
5. I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H. Balakrishnan, “Chord: A scalable
peer-to-peer lookup service for internet applications,” in SIGCOMM: Proceedings of the
2001 conference on Applications, technologies, architectures, and protocols for computer
communications, San Diego,CA, USA, 2001.
6. S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Schenker, “A scalable content-
addressable network,” in SIGCOMM: Proceedings of the conference on Applications, tech-
nologies, architectures, and protocols for computer communications, San Diego,USA, 2001.
7. H. Yin, X. Liu, T. Zhan, V. Sekar, F. Qiu, C. Lin, H. Zhang, and B. Li, “Livesky: Enhancing
cdn with p2p,” ACM Trans. Multimedia Comput. Commun. Appl., August 2010.
8. K. Sripanidkulchai, B. Maggs, and H. Zhang, “An analysis of live streaming workloads on the
internet,” in Proceedings of the 4th ACM SIGCOMM conference on Internet measurement,
ser. IMC ’04, 2004.
9. J. He, A. Chaintreau, and C. Diot, “A performance evaluation of scalable live video streaming
with nano data centers,” Comput. Netw., vol. 53, pp. 153–167, February 2009.
10. C. Vicari, C. Petrioli, and F. L. Presti, “Dynamic replica placement and traffic redirection in
content delivery networks,” SIGMETRICS Perform. Eval. Rev., vol. 35, December 2007.
11. Z. H. Lu, X. H. Gao, S. J. Huang, and Y. Huang, “Scalable and reliable live streaming service
through coordinating cdn and p2p,” in Proceedings of the 2011 IEEE 17th International
Conference on Parallel and Distributed Systems, ser. ICPADS, Washington, DC, USA, 2011.
12. Z. Zhuang and C. Guo, “Optimizing cdn infrastructure for live streaming with constrained
server chaining,” in Proceedings of the 2011 IEEE Ninth International Symposium on Paral-
lel and Distributed Processing with Applications, ser. ISPA, Washington, DC, USA, 2011.](https://image.slidesharecdn.com/2012-160421130646/85/Optimizing-Streaming-Server-Selection-for-CDN-delivered-Live-Streaming-14-320.jpg)
Optimizing Streaming Server Selection for CDN-delivered Live Streaming
- 1. Optimizing Streaming Server Selection for CDN-delivered Live Streaming Zhenyun Zhuang1 and Shun Kwok1,2 1 Ying-Da-Ji Tech., 603 Overseas High-Tech Venture Park, Shenzhen 518057, China 2 Corresponding Author Abstract. Content Delivery Networks (CDNs) have been widely used to deliver web contents on today’s Internet. Gaining tremendous popularity, live streaming is also increasingly being delivered by CDNs. Compared to conventional static or dynamic web contents, the new application type of live streaming exposes unique characteristics that pose challenges to the underlying CDN infrastructure. Unlike traditional web-objects fetching, which allows Edge Servers to cache con- tents and thus typically only involves Edge Servers for delivering contents, live streaming requires real-time full CDN-streaming paths that span across Ingest Servers, Origin Servers and Edge Servers. DNS is the standard practice for enabling dynamic assignment of servers. GeoDNS, a specialized DNS system, provides DNS resolution by taking into account the geographical locations of end-users and CDN servers. Though GeoDNS effec- tively redirects users to nearest CDN Edge Servers, it may not be able to select the optimal Origin Server for relaying a live stream to Edge Servers due to the unique characteristics of live streaming. In this work, we consider the require- ments of delivering live streaming with CDN, and propose advanced design for selecting optimal Origin Streaming Servers in order to reduce network transit cost and increase viewers’ experience. We further propose a live-streaming spe- cific GeoDNS design for selecting optimal Origin Servers to serve Edge Servers. 1 Introduction Today’s web contents are increasingly being delivered by Content Delivery Networks (aka Content Distribution Networks, or CDNs) [1,2]. a typical CDN infrastructure con- sists of two types of servers based on their functionalities: Ingest Servers for accepting customers’ web contents; and (ii) Delivery Servers for serving end web users. Live streaming is becoming increasingly popular, and it is expected to be more pervasive in the near future. Major CDN providers such as Akamai [1] and Level3 [2] are support- ing delivering live streaming with their CDNs. When supporting live streaming, CDN providers provide a layered infrastructure which consists of three types of streaming servers: (i) Ingest Streaming Servers (ISS) which accept customers’ live streams; and (ii) Origin Streaming Servers (OSS) which transit live streams from ingest servers to Edge Streaming Servers; and (iii) Edge Streaming Servers (ESS) that directly serve the end viewers. Depending on the scale of a CDN, the number of Origin Streaming Servers varies from several to hundreds, and Edge Streaming Servers can be even thousands. Unlike other traditional web-based applications, live streaming is associated with unique characteristics and in turn presents special challenges to CDN infrastructure.
- 2. (a) Default Delivery (b) Optimized Delivery Fig. 1. Live streaming delivered with CDN First, live streaming has to be delivered from the CDN Ingest Server to Edge Servers in a real-time fashion, which is in sharp contrast with traditional CDN-based web delivering where typically only the last hop (i.e., from CDN Edge Servers to web users) needs to be optimized. Such uniqueness requires the optimization of the entire CDN transit path which consists of two transit hops of Ingest-Origin and Origin-Edge. Second, live streaming often features high-bandwidth requirement. A typical live videos can easily consumes up to 3 Mbps per stream, thus there is a higher need for reducing transit cost on CDN networks. Third, live streaming typically is time-sensitive, and the CDN infrastructure design needs to consider such timeliness by choosing the shortest path from the streaming source to the end users. After a live stream is ingested into the CDN infrastructure through an Ingest Server, when an end viewer requests to play a live stream, an appropriate Edge Server and an Origin Server need to be chosen, so that the player pulls the stream from the Edge Server and the Edge Server pulls the stream from the Origin Server. DNS is the stan- dard practice for making the above server selections. GeoDNS [3], typically used by CDNs, is a specialized DNS service by taking into account geographical locations of CDN Origin/Edge servers and end users. It translates host names into IP address and redirects the requests to optimal Origin/Edge servers. Though many factors such as server loads contribute to this selection process, the critical factor is the proximity (i.e., geographical) information. For instance, in a CDN that deploys Edge Servers on both coasts of USA, an end user coming from California most likely will be served by an Edge Server located in west coast rather than east coast. Traditional GeoDNS solutions work effectively by offering the end users lower response time and higher throughput. Inside CDN infrastructure, GeoDNS is also used to determine the optimal Origin Server to serve Edge Server requests, with the goal of offering lower transit cost and response time. When supporting live streaming, GeoDNS is used for dynamical assignment of Ori- gin Streaming Servers in response to Edge Streaming Server requests. In this process,
- 3. traditional GeoDNS only considers the properties (e.g. proximity) of Origin Servers with respect to a particular Edge Server, which may result in sub-par performance. Consider an example shown in Figure 1 where Source S is ingesting a live stream into CDN’s Ingest Server I. Assuming Edge Server E needs to serve User U the live stream originating from Ingest Server I. E has two optional Origin Servers to retrieve the stream from: O1 and O2. Since O1 is closer to E, default GeoDNS will assign O1 to E’s DNS request, thus resulting in a CDN transit path of I − O1 − E, as shown in Figure 1(a). However, as shown in Figure 1(b), there exists a better CDN transit path of I −O2 −E, which has smaller CDN transit cost in terms of stream delivery distance. In addition to less transit cost, U’s viewing experience is also expected to be better, as the entire live stream path of S−I −O2 −E −U is shorter. The CDN transit path shown in Figure 1(b) can be achieved by an optimized GeoDNS solution which assigns to E the Origin Server of O2, as opposed to O1. Though effectively addressing all challenges associated with live streaming requires advanced design/redesign of many aspects of CDN infrastructure, in this work, we focus on optimizing the Origin Streaming Server selection through GeoDNS, and propose a design called Sticky-DNS for supporting live streaming. The key idea of Sticky-DNS is that when performing GeoDNS and selecting Origin Servers to serve Edge Servers, the optimal Origin Server is determined with the consideration of two CDN hops of both Ingest-to-Origin and Origin-to-Edge rather than the single hop of Origin-to-Edge as in traditional GeoDNS. In the following paper, we first provide some background information and motivate our design of Sticky-DNS in Section 2. We then present the problem definition and detailed design in Section 3 and Section 4, respectively. We perform prototype-based evaluation and show the results in Section 5. We present related works and conclude the work in Section 6 and Section 7, respectively. 2 Background In this section, we first provide some backgrounds about CDN, GeoDNS, and live streaming. We then present some initial results about the popularity of live streams on Ustream.tv, a major live streaming web site. Finally, we summarize this section by providing insights into the design of new GeoDNS for live streaming. 2.1 Backgrounds CDN: Content Delivery Networks (CDNs) can be used to deliver both static web con- tents and live streams. A typical CDN infrastructure consists of Ingest Servers (for ac- cepting customer contents), Origin Servers (for serving edge servers) and Edge Servers (for serving end users directly), forming a tree-like structure. Various servers are orga- nized into POPs (i.e., Point Of Presence). GeoDNS: CDNs aim to deliver contents directly from Edge Servers that are closer to end users, thus minimizing the adverse impact of dynamic network path and reducing network transit cost. For this, a specialized DNS service is typically used for redirecting requesting end users to appropriate Edge Servers.
- 4. 0 200 400 600 0 10 20 30 40 50 60 Ustream.TV Live Streams Fig. 2. Number of viewers for the first 55 streams on Ustream.tv Live Streaming: Live streaming is a new type of application that is increasingly delivered by CDNs. Featuring different properties from static contents, live streaming typically has high bandwidth requirement which exerts pressure on the CDN infrastruc- ture. Unlike static web contents, live streams cannot be cached and have to maintain a full path from the customer source to the end viewers (i.e., end users). 2.2 Popularity of live streams On today’s Internet, vast amount of live streams are being supported by CDNs. Though the sizes of viewer set vary greatly, many live streams today feature small viewer sets. Specifically, we consider a major live streaming website of Ustream.tv and study the live streams that fall into the category of “most-total-views” at a particular time. We count the numbers of viewers for each of the first 55 streams, and the results are shown in Figure 2. Note that the figure only shows the active live streams, and other streams are off-air at the particular time. We observe that though the most popular stream has more than 500 viewers, only 6 of them have more than 100 viewers. The average number of live viewers is only about 53. 2.3 Observations As we observed from the illustrate example shown in Figure 1, CDN transit cost (i.e., Ingest-Origin and Origin-Edge paths) can be saved by assigning appropriate Origin Servers to requesting Edge Servers for live streaming applications. Instead of always assigning the nearest Origin Server to the Edge Server that issues DNS request, a specially-designed GeoDNS could determine the assignment based on the entire CDN delivery path. Such a special GeoDNS can significantly save CDN transit cost for sce- narios where a live stream is not popular as they incur shorter CDN delivery paths. Though the popularity of streams (i.e., viewer set size) varies greatly, many live streams today feature small viewer sets. On the other hand, today’s major CDN providers typically have hundreds or even thousands of Edge/Origin streaming servers for sup- porting live streaming. The mismatch between the sizes of viewer sets and server sets
- 5. motivates our design of a specialized DNS for less-popular live streams, with the benefit of reduced CDN transit cost. For popular live streams, however, the same approach needs to be adapted to re- utilize delivery path at Ingest-Origin level. Specifically, default GeoDNS practically encourages path reutilization at Ingest-Origin level since each Origin Server can ag- gregate the streaming requests from closer-by Edge Servers, thus reducing the CDN transit cost. Allowing each Edge Server to connect to far-away Origin Servers, on the other hand, discourages path reutilization, thus could result in larger aggregate delivery distance inside CDN. Therefore, an appropriate design of GeoDNS has to consider the popularity of a live stream, and adapts its behavior correspondingly. 3 Design of Sticky-DNS In this section, we first formalize the problem we attempt to address. We then present the design overview of Sticky-DNS, its software structure and the entire process. 3.1 Problem Definition We consider scenarios where a CDN provider has three types of streaming servers for supporting live streaming: Ingest Streaming Servers, Origin Streaming Servers, and Edge Streaming Servers. We use SSI, SSO and SSE to denote the three server sets, re- spectively. The CDN provider delivers a set of live streams {Lm}. We further assume each stream Lm has a single ingesting source which connects to an Ingest Server SSIm. The problem we are addressing with Sticky-DNS design can be defined as follows. For a live stream denoted by Lm and a viewer Vn, we assume the Edge Server that serves viewer Vn is SSEmn, which is determined by regular GeoDNS. Typically SSEmn is the closest Edge server to the end viewer. Sticky-DNS needs to decide the optimal Origin Server SSOmn to serve SSEmn such that the CDN transit cost can be minimized and the end viewers’ experience is not compromised. We use CIO and COE to denote the transit costs corresponding to the paths of Ingest-Origin and Origin-Edge, respectively. With these assumptions, the input, output and objective function of Sticky-DNS are presented in Figure 3. 3.2 Design Overview Sticky-DNS is a specialized DNS service designed for live streaming applications. It aims at saving CDN transit cost but not compromising end user experience. 3 Sticky- DNS achieves the cost savings by assigning proper Origin Servers to serve Edge Servers with shorter CDN delivery paths. Sticky-DNS’s operations are per-stream based, thus it maintains appropriate states for all live streams. At a higher level, Sticky-DNS adjusts its behavior as the particular live stream changes popularity. Specifically, when a live stream is not popular, Origin 3 Note that for ease of presentation, we use the distance as the primary cost metric; but it would be easy to extend the cost metric to include other factors such as pricing.
- 6. Input: 1 {Lm}: Live stream set of individual stream Lm; 2 Vm = {Vmn}: Viewer set for live stream Lm; 3 SSIm: The Ingest Server for live stream Lm; 4 SSOm: CDN Origin Server set for Lm; 5 SSEm = {SSEmn}: CDN Edge Server set for Lm; // Note that each Edge Server 6 // Emn for each individual viewer Vn is determined by regular GeoDNS; 7 CIO: Cost arrays of all Ingest-Origin pairs; 8 COE: Cost arrays of all Origin-Edge pairs; Output: 1 Optimal Origin Server set SSOm = {SSOmn} for Edge Server set SSEm, where each SSOmn is serving the Edge Server of SSEmn ; Objective Function: 1 Total CDN transit cost Σ(CIO +CEO) for all live streams {Lm} is minimized; 2 Non-compromised viewer experience; Fig. 3. Problem Definition Servers are chosen based on the lengths of entire Edge-Origin-Ingest paths. For this, Sticky-DNS maintains the cost values of every Edge-Origin and Origin-Ingest pairs. Initially, all the costs are assigned according to certain cost-determination methods. Whenever an Origin Server is arealdy serving the live stream, the corresponding Origin- Ingest cost is set to zero to reflect the fact that serving an additional Edge Server with this Origin Server will incur no additional cost on the Origin-Ingest path. However, as a stream becomes more popular, Sticky-DNS needs to adapt back to the default behavior determined by regular DNS for reusing Ingest-Origin paths. The reason is that when serving popular streams, since Origin Servers are more likely serving that live stream, the behavior of Sticky-DNS will more likely be that of regular GeoDNS, and the proximity-closest Origin Server is typically chosen for each Edge Server so that the Ingest-Origin paths are reutilized. To encourage this, we introduce the concept of “would-be” behavior. When an Origin Server is determined by regular DNS for Origin request issued by an Edge Server, the particular Origin is the “would-be” Origin for the Edge, while the Edge is the ”would-be” Edge for the Origin. Even if a different Origin is determined by Sticky-DNS, the “would-be” state is recorded, which is used to adjust the corresponding Origin-Ingest cost. Specifically, with more “would-be” Edge Servers retrieving from an Origin, the corresponding Origin-Ingest cost is reduced further. The idea behind this treatment is that the Origin-Ingest cost is shared by all downstream Edge Servers, thus encouraging the reuse of the particular Origin-Ingest path. These cost values will be used to choose optimal Origin Servers for incoming Origin Server DNS requests so that the smallest Σ(CIO +CEO) will result. To achieve this, Sticky-DNS uses the notion of “Additional Cost” when deciding on an Origin Server. Specifically, given an incoming viewer Vn and the corresponding Edge Server SSEmn, Sticky-DNS always return the Origin Server that causes minimal “Additional Cost”.
- 7. Fig. 4. Software Architecture The additional cost is defined as the extra CDN transit cost that needs to be paid when serving the Edge Server with a particular Origin Server. Recall that CDN transit cost contains two parts: the Ingest-Origin cost and Edge-Origin cost. Whenever an Origin Server first connects to an Ingest Server to retrieve a live stream, CDN pays Ingest- Origin cost, but succeeding retrieval of the same live stream from the same Ingest Server to the same Origin Server will incur zero additional cost. 3.3 Software Architecture The software architecture of Sticky-DNS is shown in Figure 4. The heart is Sticky-DNS Engine (SDE), which takes the DNS request, interacts with other components, and then decides the optimal Origin Server. Sticky-DNS assigns Origin Servers based on the cost associated with the live streaming delivery path, and the cost is CDN-specific and de- termined by another component of Cost Metric Determination (CMD) and updated by Server Cost Probing (SCP). Meanwhile, to accommodate network dynamics, a com- ponent of Edge-Origin Connection Monitoring (ECM) is used to update the cost state. Sticky-DNS maintains the server streaming state which basically records information of which Edge Server is serving which streams and which Origin Server is retrieving which streams. This state is updated by the events of viewers joining and leaving. The overall process of how Sticky-DNS works is as below. When a new live stream- ing DNS request comes, the optimal Edge Server will be determined by regular GeoDNS. CDN will then attempt to determine the optimal Origin Server for the determined Edge Server. For this, the Edge Server issues DNS request to DNS service asking for the proper Origin Server for retrieving the live stream from. Sticky-DNS takes the Origin- Server request and returns the determined Origin Server. Meanwhile, it updates the server streaming states. 4 Sticky-DNS Components 4.1 Cost Metric Determination (CMD) In order to perform Sticky-DNS operations, each delivery hop (i.e., Ingest-Origin and Origin-Edge) on CDN infrastructure has to be associated with quantifiable cost. Note
- 8. that though physical distance is often used as the path cost, cost metrics may also in- clude network delay, price, server load, reliability, loss rate, etc. Since different cost metrics often belong to distinct dimensions and they are often prioritized. Though how to prioritize them varies with different CDN providers and deployment scenarios, in this work we provide the following guidelines: (i) Server Load is prioritized over all other metrics. An Origin server is put into the DNS pool (i.e., available to be returned by Sticky-DNS) only if it can take more loads (i.e., serving more requests); (ii) Loss rate has to be under certain threshold. Since most live streaming (e.g., RTMP, HTTP) use TCP to transmit, large loss rate cannot sustain a reasonable level of performance; (iii) Reliability has to be above certain threshold. Live streaming requires a real-time transmission path, thus reliability is important during the stream playback; (iv) Pricing is prioritized over network delay and physical distance. Though pricing is often cor- related with the latter two metrics, it is prioritized over them if they are not identical; (v) Network delay and physical distance. The impact of these two metrics are buffering time and playback delay. In reality, they can be used as the substitute for pricing metric when the latter is not available or hard to define. Thus, with the above guidelines in mind, we propose to enforce the cost determina- tion by keeping track of the following state information. – An array of server load for the Origin Servers, denoted by SL[M], where M is the number of Origin Servers. SL[i] represents the current load of Origin Server SSOi; – A 2-D array of loss rate for each Origin-Edge path, denoted by LR[M][N], where M is the number of Origin Servers and N is the number of Edge Servers. LR[i][j] represents the loss rate of the path between Origin Server SSOi and Edge Server SSE j; Another array of loss rate for each Ingest-Origin pair, denoted by LRI[M], assuming a single Ingest Server; – A 2-D array of reliability for each Origin-Edge path, denoted by RE[M][N], where M is the number of Origin Servers and N is the number of Edge Servers. RE[i][j] represents the reliability of the path between Origin Server SSOi and Edge Server SSE j; Another array of reliability for each Ingest-Origin pair, denoted by REI[M]; – Pricing is represented by PR[M][N] for each pair of Origin-Edge; Another array of PRI[M] for each pair of Ingest-Origin; – Similarly, network delay and physical distance are represented by ND[M][N] and PD[M][N], respectively. Two other arrays of NDI[M] and PDI[M] are also used. 4.2 Server Cost Probing (SCP) SCP updates the various cost state by proactively or reactively monitoring/probing the network. Specifically, – Origin Server Load (SL[M]). The load can be obtained by installing services which monitor CPU/Memory utilization; – Loss rate (LR[M][N] and LRI[M][N]). Loss rate monitoring between servers can be obtained by observing existing connections’ behaviors (e.g., TCP loss recovery behaviors) or proactively sending test packets. When a network path (e.g., Ingest- Origin-Edge) consists of two network hops, the aggregate loss rate of path SSEi − SSO j −SSI can be calculated as LRij = (1−(1−LR[j][i])(1−LRI[j]));
- 9. Core Algorithm: Initialize all cost states; For an incoming new viewer Vn requesting Lm, determine the Edge Server SSEmn; 1 If SSEmn is already serving Lm; 2 Return; //No Origin Server request since the Edge Server is serving Lm now; 3 Else 4 Determine the candidate Origin Server set based on load, loss rate, etc.; 5 For every Origin Server SSOi, if (SL[i] ≤ α and LRj,k ≤ β and REj,k ≥ γ, where 6 α, β,γ are the respective threshold values), SSOi is a candidate Origin Server; 7 Determine the optimal Origin Server SSOq that has the smallest aggregate 8 cost of CIO and COE for SSEmn; 9 NV[q][m]++; 10// Update AD[q][m]; 11 If (NV[i][j] ≥ 1) 12 AD[i][j] = 0; 13 Else 14 If ( WB[i][j] ==0 ): AD[i][j] = initial cost; 15 Else : AD[i][j] = initialcost WB[M][T]+1 If Edge Server SSEi is disconnecting from SSOj for live stream Lm: 1 // Update NV[I][J] and AD[I][J]; 2 NV[j][m]–; 3 If (NV[j][m] ≥=1): AD[j][m] = Initialcost (NV[j][m]+1) ; 4 Else : AD[j][m] = {initialcost}. Fig. 5. Core Algorithm – Similarly, reliability (RE[i][j] and REI[i][j]) can be obtained by monitoring the path up/down time. The aggregate reliability of path SSEi −SSOj −SSI can be calculated as REij = (RE[j][i]∗REI[j]); – Pricing (PR[M][N] and PRI[M][N]) can be adjusted by business logics. Whenever the CDN operators change the network pricing rates, the pricing used by Sticky- DNS can be correspondingly adjusted. – Network delays (ND[M][N] and NDI[M]) and physical distances (PD[M][N] and PDI[M]) are easy to obtain with known techniques. 4.3 Edge-Origin Connection Monitoring (ECM) ECM is used to keep the server streaming state maintained by OSS fresh, in the sense that the additional cost associated with each Origin Server is up-to-date, and the popu- larity state of a stream on all Origin Servers is up-to-date. For all Origin Servers, there is a server streaming state. The server streaming state is an array, and each element corre- sponds to an Origin Server and is defined as the Additional Cost of using the particular Origin Server to serve a stream. The state can be realized by a 2-D array of AD[M][T],
- 10. (a) Original, no viewers (a) E1 is on (a) E2 is on Fig. 6. An illustrative example (I) Source: S Ingest: I Origin: O1 Origin: O2 Edge: E1 Edge: E2 Edge: E3 (a) E3 is on (a) E4 is on Fig. 7. An illustrative example (II) where M is the number of Origin Servers and T is the number of streams. AD[i][j] records the additional cost of using Origin Server SSOi to serve stream Lj. The values of AD[M][T] are based on whether an Origin Server serves a stream or not. Two other 2-D arrays are used: NV[i][j] records the number of Edge Servers Origin Server SSOi is serving stream Lj to. WB[M][T] is used to indicate the stream popular- ity on a particular Origin Server. The indicator uses the number of “Would-be” Edge Servers for each Origin Server. ECM monitors the connecting/disconnecting events be- tween Origin and Ingest Servers for all live streams. Whenever the events occur, CVM updates NV[M][T], WB[M][T] and AD[M][T] correspondingly. 4.4 Sticky DNS Engine (SDE) Dynamic DNS Engine (SDE) performs the actually computation and comparison of costs, then selects the most appropriate Origin Server given a live streaming DNS re- quest. The algorithms are shown in Figure 5. SDE handles the joining of an Edge Server (i.e., begins to serve a stream) and the leaving (i.e., begins to drop a stream) separately. For both events, it updates the AD and NV states.
- 11. 4.5 An illustrative example We use the following example to illustrate how Sticky-DNS works. Assuming there are one Ingest Server (I) and two Origin Servers (O1 and O2), as shown in Figure 6. When the first end-viewer joins and is bound to Edge Server E1, E1 needs to decide which Origin Server (O1 or O2) for E1. Assuming the distance is the cost metric, O2 is chosen since the additional cost of pulling from O2 (i.e., sum of I −O2 and O2 −E1) is smaller. When the second Edge Server E2 requests an Origin Server, though it is closer to Origin Server O1, the additional cost of pulling from O1 (i.e., I − O1 and O1 − E2) is larger than that of pulling from O2 (i.e, only O2 − E2, since I − O2 has cost of be- ing zero). Thus E2 chooses O2 as the optimal Origin Server. Note that since E2 is the “would-be” edge of O2, the cost of I − O1 is reduced in half to reflect the benefit of path reuse. When the third Edge Server E3 joins, as shown in Figure 7, it compares the additional costs of pulling from O1 and pulling from O2. Since the cost of I −O1 is not in half, assuming the cost sum of I −O1 and O1 −E3 is smaller than the cost of O2 −E3, E3 chooses O1 to be the optimal Origin Server. Finally, when Edge Server E4 joins, it chooses O1 as the Origin Server. 5 Evaluation To gain concrete understanding of GeoDNS regarding the impact on both CDN and viewers, we consider the following simplified scenario to quantify the saving on CDN transit cost and the performance experienced by end viewers. We assume an USA mainland CDN provider which has 1 Ingest POP (Point of Presence) and 10 Origin POPs, where each POP has only a single streaming server. We further assume that Edge Servers have to retrieve live streams from Origin Servers, which then retrieve from In- gest Server. For simplicity, we consider a performance metric of physical distance in the unit of thousand-miles (i.e., KMiles) for both CDN transit-cost saving and live-stream viewers. Thus, the larger the distances are, the higher cost both CDN and viewers pay. We built a prototype with Adobe Flash Media Server (FMS) 4.0. The prototype consists of 11 servers deployed in 11 cities, all running Windows Server 2008 R2. One of the servers serves as the Ingest Server and receives a live stream from a stream source, and the other 10 servers serve as Origin Servers and are relaying the live stream to end viewers. We then invoke 10 stream viewers from different locations. Since viewers are always bound to the closest Edge Servers, we use viewers to mimic the edge servers in typical CDN setup as they need to decide the origin servers to pull stream from. We first consider the default regular DNS, which directs each viewer to the closest Origin Server. For comparison, we also consider an optimized DNS which directs each Edge POP to the closest Origin POP that leads to smallest CDN transit path (i.e., from Ingest Server to Edge Server). CDN Transit Cost: For both scenarios, we count the lengths of Ingest-Origin paths, Origin-Edge paths, and aggregated CDN delivery path. As shown in Figure 8, the blue bars show the default scenario, while the red bars show the optimized scenario. The ag- gregated CDN delivery distance is the sum of all Ingest-Origin and Origin-Edge paths. Specifically, first, the aggregated Ingest-Origin distances are 10.48 Kmiles and 1.81
- 12. 0.00 5.00 10.00 15.00 Origin Transit Cost Edge Transit Cost Aggregate Transit Cost Default GeoDNS Optimized GeoDNS KM Fig. 8. CDN Transit Cost Kmiles for two scenarios, respectively. The reason for reduced distance in optimized- DNS scenario is that some Origin POPs are not serving any Edge Servers, thus they are free from stream delivery. Second, the aggregated Origin-Edge distances are 2.92 Kmiles and 5.40 Kmiles, respectively. Since Edge Servers in optimized-scenarios may connect to far-away Origin POPs, the Origin-Edge paths are inflated. Third, the aggre- gated path distances, which correspond to the CDN transit cost, are 13.41 Kmiles and 7.17 Kmiles, respectively. We see that for the particular setup, optimized-DNS results in much smaller CDN transit cost (more than 46% reduction). These results suggest that CDN is able to gain significant transit cost saving by op- timizing GeoDNS inside CDN. Note that in this simple scenario, we assume each POP only has a single Edge Server. With more servers, the saved transit cost can be more sig- nificant as each server will incur a separate delivery path. Note that with shorter delivery paths, end viewers also see shorter playback buffers and better viewing experiences. Viewers’ Experience: Since each live stream has to delivered from the stream source all the way to the end users in a real-time fashion, the end-users’ performance is heavily impacted by the length of the entire delivery path. In other words, though there are other factors impacting live-stream users’ performance, it is generally acceptable to claim that the longer path an end user relies on, the worse performance the user experiences. For both scenarios, since the distances from streaming sources to the Ingest Server and from Edge Server to end viewer are the same, we now focus only on the CDN delivery distance from Ingest Server to Edge Server for each viewer and show them in Figure 9. We observer that all paths in optimized-DNS scenario are shorter, which is not surprising since Edge Servers choose the Origin Servers that result in shortest paths. The aggregated distances are 13.41 Kmiles and 10.35 Kmiles, respectively, or 23% decrease. These results suggest that end viewer’s experience can also improved with optimized GeoDNS. Money Talks: Taking a more direct business perspective, we now compare the transit cost with respect to pricing of bandwidth usage. Depending on subscription
- 13. 0.00 0.50 1.00 1.50 2.00 2.50 3.00 C D F H J K N O Q T Default Optimized Fig. 9. Stream Delivery Paths packages, CDN providers may pay ISPs for the bandwidth used. We assume each live stream is 3Mbps. In the above scenario where only 4 Origin Servers are deliver- ing the same live stream, the saved bandwidth per day would be 0.20 TB*Kmiles (or 73.74 TB*KMiles per year). With 2000 streams, the saved bandwidth would be 400 TB*Kmiles (or 147.47 PB * Kmiles per year). Simply assuming the transit cost of 0.10 per GB/Kmile, optimized-DNS can help save $14.74 Million per year with 2000 streams, or about $7K per stream, as shown in Table 1. 6 Related Work Content Delivery Networks (CDNs) have been carried out by various providers to ex- pedite the web access [1]. The techniques used by CDNs for delivering conventional web contents are explained in related writings [4]. Despite the popularity and perva- siveness of CDNs, many research problems and system-building challenges persist for optimizing CDN infrastructure and addressing challenges associated supporting various types of applications. For instance, as we demonstrated in this work, how to efficiently support the application of live streaming with CDNs deserves more study. Though various network architectures have been proposed and utilized for different deployment scenarios with the rapid development of Internet technology [5, 6], CDN providers typically apply layered structure (e.g., ingest-layer, origin-layer and edge- layer) due to the caching-property of delivering static web contents tracing back to Day-1 of CDN industry. Realizing that CDNs and P2P can complement each other, Table 1. Bandwidth/Cost Saving of Sticky-DNS Number of streams Daily BW Saving Yearly BW Saving Yearly Cost Saving 1 0.20 TB 73.74 TB $7K 2000 400 TB 147.47 PB $14.74M
- 14. many research efforts also attempt to integrate these two realms into one [7]. Unlike these works that combine the merits of both CDNs and P2P, our work takes an inside perspective and attempts to refine CDN delivery infrastructure. Live streaming has been studied from various perspectives [8–10]. The unique prop- erties of live streaming, coupled with CDN-assisted delivery, justifies a specialized de- sign of CDN infrastructure that specifically serves live streaming and saves CDN transit cost. One work [11] proposes to combine the benefits of CDN and P2P to enhance live streaming. Another work [12] proposes to build a flat-layer infrastructure (as opposed to the traditional 3-layer infrastructure) for supporting live streaming with CDNs. To our best knowledge, this work is the first to consider and address the problem of saving transit cost for delivering live streaming with traditional layered-infrastructure CDNs. 7 Conclusion In this work, we optimize CDN streaming server selection to better support live stream- ing with CDN. The proposed design dynamically selects streaming servers based on the popularity of the live stream. Significantly improving both the CDN transit cost saving and end viewers’ experience, the design can be deployed with a specialized GeoDNS. References 1. “Akamai technologies,” http://www.akamai.com/. 2. “Level 3 communications, llc.” http://www.level3.com/. 3. “Bind geodns,” http://www.caraytech.com/geodns/. 4. D. C. Verma, S. Calo, and K. Amiri, “Policy based management of content distribution net- works,” IEEE Network Magazine, vol. 16, pp. 34–39, 2002. 5. I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H. Balakrishnan, “Chord: A scalable peer-to-peer lookup service for internet applications,” in SIGCOMM: Proceedings of the 2001 conference on Applications, technologies, architectures, and protocols for computer communications, San Diego,CA, USA, 2001. 6. S. Ratnasamy, P. Francis, M. Handley, R. Karp, and S. Schenker, “A scalable content- addressable network,” in SIGCOMM: Proceedings of the conference on Applications, tech- nologies, architectures, and protocols for computer communications, San Diego,USA, 2001. 7. H. Yin, X. Liu, T. Zhan, V. Sekar, F. Qiu, C. Lin, H. Zhang, and B. Li, “Livesky: Enhancing cdn with p2p,” ACM Trans. Multimedia Comput. Commun. Appl., August 2010. 8. K. Sripanidkulchai, B. Maggs, and H. Zhang, “An analysis of live streaming workloads on the internet,” in Proceedings of the 4th ACM SIGCOMM conference on Internet measurement, ser. IMC ’04, 2004. 9. J. He, A. Chaintreau, and C. Diot, “A performance evaluation of scalable live video streaming with nano data centers,” Comput. Netw., vol. 53, pp. 153–167, February 2009. 10. C. Vicari, C. Petrioli, and F. L. Presti, “Dynamic replica placement and traffic redirection in content delivery networks,” SIGMETRICS Perform. Eval. Rev., vol. 35, December 2007. 11. Z. H. Lu, X. H. Gao, S. J. Huang, and Y. Huang, “Scalable and reliable live streaming service through coordinating cdn and p2p,” in Proceedings of the 2011 IEEE 17th International Conference on Parallel and Distributed Systems, ser. ICPADS, Washington, DC, USA, 2011. 12. Z. Zhuang and C. Guo, “Optimizing cdn infrastructure for live streaming with constrained server chaining,” in Proceedings of the 2011 IEEE Ninth International Symposium on Paral- lel and Distributed Processing with Applications, ser. ISPA, Washington, DC, USA, 2011.