P2PTV Traf?c Localization by Deep Packet Inspection
Hiep Hoang-Van,Takumi Miyoshi Graduate School of Engineering and Science Shibaura Institute of Technology
Saitama,Japan
Email:{nb12510,miyoshi}@shibaura-it.ac.jp
Olivier Fourmaux Laboratoire d’Informatique de Paris6 UPMC Sorbonne Universit′e s
Paris,France
Email:olivier.fourmaux@upmc.fr
Abstract—Recently,peer-to-peer has been becoming a promis-ing architecture for developing video streaming services because it can reduce the bottlenecks and the load at the server side.
However,P2P systems implement their own routing protocols on overlay networks,which are largely independent of the Internet routing.As a result,the traf?c generated by P2P systems is immense and unpredictable.Controlling the P2P traf?c is therefore becoming a big challenge for internet service providers (ISPs).Considering the P2P traf?c localization is one of the most ef?cient solutions to optimize the traf?c.The existing approaches, however,require several modi?cations of the application software. In this paper,we propose a novel traf?c localization method for P2P streaming application(P2PTV)by deep packet inspection. Based on the peer list and the geographical location of the listed peers,we propose two mechanisms for localizing the traf?c:the modi?cation of the peer list packets and the redirection of the video request packets at network routers.Since our method is implemented on the network routers,it can be applied for all P2P applications without any software modi?cations.The experiments evaluated on a popular P2PTV,namely PPStream,prove the effectiveness of our proposed method on the problem of traf?c localization.
I.I NTRODUCTION
According to the recent report of Cisco Systems Inc.,the Internet traf?c is increasing tremendously.The main reason is the rapid growth of video traf?c.It is reported that the sum of all forms of video including TV,video on demand(V oD), Internet,and P2P will be approximately86percent of global consumer traf?c by2016[1].Recently,video streaming ser-vices tend to shift from client/server models such as YouTube to P2P streaming systems such as PPTV[2],PPStream[3], SopCast[4],and Zattoo[5],which are more robust and have less congestion than the conventional model.
The problem of P2P systems is that ISPs have to face new challenges of traf?c control.P2P applications normally implement their own routing protocols on overlay networks, which are largely independent of the physical network infor-mation due to a lack of knowledge about the underlay network topology.As a result,they generate a large amount of unwanted traf?c on the Internet.For the intra-domain,P2P systems may cause unnecessarily traverse multiple links inside an ISP [6].For the inter-domain,the underlay topology-unaware peer selection strategies of many P2P applications may generate a signi?cant amount of cross traf?c among ISPs and autonomous systems(ASes).They affect bandwidth availability of other
(1) Request peer list
Peer list server
Fig.1:Protocol of PPStream.
applications and thus lead to the changes in traf?c engineering strategies of ISPs.
We focus on the problem how to reduce the inter-domain traf?c generated by P2PTV,i.e.,the problem of traf?c lo-calization.Many methods have been proposed for addressing the problem[6]–[12].However,almost all existing approaches require several modi?cations of application software.For example,each P2P application must be equipped with a locality-aware procedure to estimate the distance[7],[8],or an additional protocol to obtain physical network information from an oracle service[6],[9],[10].
In this study,we propose a novel approach to localize the P2PTV traf?c.In P2PTV,a peer generally receives a list of available peers as its neighbors from some peer list servers.The peer then contacts some peers in the peer list for exchanging video data by sending video data request packets. By deep packet inspection,we analyze every packet going through a gateway router to?nd which packet contains the peer list,and to parse the list of IP addresses of all peers in the peer list.The geographical locations of all peers are easily resolved by using several IP-to-location database services.We propose two mechanisms for localizing the traf?c as follows:?Modify the peer list packets:the peer lists are modi?ed for localizing before they arrive at the application;
in other words,the data payloads of the packets are
modi?ed.Since the application just connects to some
peers in the localized peer list,the traf?c will therefore
2013 14th ACIS International Conference on Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed
Computing
be localized.
?Redirect the video request packets:we collect peers located in the same area from all the obtained peer
list packets,e.g.,all peers in Japan,and mark them
as localized list.After that,video data request packets
sent to the peers that do not exist in the localized list
are modi?ed to redirect them to peers in the localized
list.In other words,the control information of the
packets,the destination IP addresses in particular,will
be modi?ed.Since the application just sends queries
to some peers in the localized list,the video traf?c
will therefore be localized.
Because our proposed method is implemented on the gate-way routers,it does not require any modi?cation of existing P2P applications.Therefore,the proposal can be easily applied for all P2P applications.We believe that our router-aided approach will make a contribution to the problem of traf?c localization.
The remainder of this paper is organized as follows.We introduce the related work in Section II.Section III explains the peer list modi?cation and video request packet redirection method.The implementation of two mechanisms is described in Section IV.Section V shows experimental results.Finally, the conclusions and ongoing work are provided in Section VI.
II.R ELATED W ORK
Network distance estimation by end systems is one way to solve the traf?c localization problem.Costa et al.pro-posed PIC,a practical coordinate-based mechanism to estimate round-trip time(RTT)or network hops between two arbitrary nodes on the Internet[7].PIC maps each node to a point in a d-dimension Euclidean space.When a new node joins the system,it computes the coordinates of its corresponding point based on a set of landmarks,L,where L≥d+1and their coordinates must be computed in advance.The idea of PIC is similar to GNP[13],but the number of landmarks in PIC is not?xed for all nodes that join the system,and therefore PIC has more practicality.
Wong et al.proposed a Meridian framework for performing node selection based on network location[8].In Meridian, each node maintains a track of small?xed number of neighbors and organizes them into multi-resolution rings according to the distance from the node.When a Meridian node receives a request to?nd the closest node to a given target node,it ?rst computes the latency d between itself and the target,and then forwards the request to one of its ring members if the latency between the ring member and the target is less than β×d,whereβis an acceptance threshold.The above steps are repeated with the new node.If no new nodes meet the threshold,then the routine stops and the currently closest node is chosen.
Although P2PTV can adapt to the above approaches for developing neighbor peer selection strategies,this is not a trivial task due to a requirement of many modi?cations of existing software.Moreover,when some overlay networks are running on one node simultaneously,such measurement-based approaches are very dif?cult to reach high accuracy.
Recently,Xie et al.introduced P4P,a promising frame-work for traf?c optimization,where the underlay network information is not estimated by P2P applications but provided by ISPs[6].In the P4P architecture,each ISP maintains an iTracker in its own network to provide the p-distance interface, representing the logical distance and link costs among PIDs (aggregation nodes)based on physical network information. The P2P applications can use these p-distances as a guidance to shape their connectivity and to choose neighbor peers more ef?ciently.The experimental results evaluated by using simulation and real network on PlanetLab prove that P4P can improve both network ef?ciency and P2P applications performance.The Internet Engineering Task Force(IETF)has formed a working group for standardizing a protocol to enable P2P applications to access information provided by ISPs, known as Application Layer Traf?c Optimization(ALTO)[14], [10].The main goal of ALTO service is to make a win-win situation between network providers and P2P applications. ISPs,by offering ALTO service,can reduce the amount of inter-ISP traf?c,and optimize the intra-domain traf?c.On the other hand,P2P applications,by using ALTO guidance, can apply better-than-random peer selection strategy due to a deep understanding about underlay network.Furthermore,P2P systems are not limited in bandwidth by ISPs,and therefore improve the performance.
However,both P4P and ALTO have the following prob-lems:(1)Physical network information disclosure:for ef?cient traf?c localization,ISPs might need to open some detailed information to external entities.This raises a security problem;
(2)The need of server systems to provide network information to P2P applications;(3)Some modi?cations of P2P application software required for implementing the protocol to commu-nicate with the servers.(4)The need of the trust and good cooperation between ISPs and the P2P users.
Choffnes et al.proposed another approach for traf?c lo-calization,which requires no cooperation between ISPs and P2P applications[15].They proposed to recycle network views gathered at low cost from content distribution networks (CDNs)to drive biased neighbor peer selection.By using DNS redirection,they hypothesized that two peers are recognized as being close to each other if they are redirected to a similar set of replica servers.The idea is deployed as a java plugin named Ono to Azureus BitTorrent client.Although this method needs no support from ISPs,it needs the support from many subscribing peers distributed worldwide.
Bindal et al.proposed biased neighbor selection scheme applying for Bittorent?le sharing system[11].Instead of a random approach,the neighbor peers are selected as follows: only k peers from outside of the ISP and(35?k)peers in the same ISP,where k is a parameter.The authors envisioned two ways to implement the proposal:modifying trackers and clients,and modifying the traf?c shaping devices.The former is similar to ALTO or P4P that requires a lot of software modi?cation,whereas the latter does not require any modi-?cation of software.However,the authors have not given an implementation of the idea in detail,and just evaluated on Bittorent by using simulation.In contrast,we focus on P2PTV and evaluate our proposed method on real network.
Miyoshi et al.proposed P2P-DISTO(Delay Insertion Scheme for Traf?c Optimization)framework for P2P traf?c
Fig.2:The proposed idea of peer list modi?cation mechanism. localization[12].The idea is the insertion of additional delay into each packet according to geographical location of the destination at network routers.Since P2P-DISTO is imple-mented on network routers,it does not require any modi?cation of application software.However,to delay a large number of packets for a prede?ned time simultaneously,P2P-DISTO requires a very large buffer to hold the packets.Therefore,it is hard to deploy the proposal in a real router.
Our work follows the idea of P2P-DISTO,but with a differ-ent mechanism.Instead of adding longer delay to farther peers to limit the connections from these peers,we directly modify the peer list packets or the query video data request packets. Since our method does not require large buffer memory,it can be easily deployed in a real router.
III.P ROPOSED S CHEME
In a P2PTV system such as PPStream,once a user initial-izes the playback of a channel,the peer will join an overlay network constructed by all the peers playing the same channel. The protocol in detail is shown in Fig.1.Firstly,a PPStream peer sends a request to a peer list server to obtain the list of candidate destinations where the desired resource resides. Secondly,after obtaining the peer list,the peer then sends hello packets to some peers in the peer list to know if they are now active or not.Thirdly,the peer sends requests to some active peers to ask buffer-map messages;a buffer-map message indicates which video chunks a peer has already buffered and can share with other peers.Finally,the peer sends requests to available peers for video data chunks,and begins to exchange data.We found two steps of the PPStream protocol where we can intervene to make the traf?c localized.One is at the step to obtain the peer list.We thus propose to modify the peer list packets for localizing before they arrive at the application as shown in Fig.2.Since the application just connects to some peers in the modi?ed peer list,the traf?c is therefore localized.The other is at the step to request video data chunks. If all the video data requests sending to locality-unaware peers are redirected to local peers,the traf?c can also be localized. Therefore,we propose another mechanism to localize the P2P traf?c as shown in Fig.3.Here after,we refer to mechanism to redirect the video data request packets as packet redirection mechanism for short.
A.Peer list modi?cation mechanism
To deploy the proposed method without any software mod-i?cations,the peer list modi?cation and the packet redirection mechanism should be implemented from outside of the existing applications.Therefore,we introduce a router-aided approach. Figure4presents the router architecture for the proposed peer
Fig.3:The proposed idea of packet redirection mechanism. list modi?cation method.Two modules,a packet monitoring module and a peer list modi?cation module,are added to a common router.The packet monitoring module inspects every packet going through the router to check whether it contains the peer list or not.If the peer list is found in the packets, the IP addresses of all peers in the peer list are collected,and the packets are passed to the peer list modi?cation module, otherwise the packets are forwarded directly to the common routing function.In the peer list modi?cation module,the obtained IP addresses are?rst mapped to their countries by using several IP-to-location database services.Then,the peer list is modi?ed as follows:keep only k foreign peers in the peer list,where k is a parameter with value from0to the number of peers.Note that the format of the peer list packet had been checked by Wireshark[16],a well-known packet sniffer application,in advance.
B.Packet redirection mechanism
This method requires to know which packet is the video data request packet for the redirection.By analyzing Wire-shark’s packet capture?les,we found many similar UDP packets that have large data payloads,more than1000bytes. These packets are likely to carry the video data chunks.We assume that the video data request packets are sent to some peers in the peer list and followed by the packets carrying video data chunks.Such kind of packets are found in UDP format,and have very small data payloads,less than100bytes.
Figure5shows the router architecture for the proposed packet redirection mechanism.It is quite similar to the archi-tecture of the peer list modi?cation mechanism,but the peer list modi?cation module is replaced by a packet redirection module.The packet monitoring module inspects every packet going through the router:if the packet contains the peer list,the IP addresses of local peers in the peer list,e.g.,all IP addresses in Japan are collected.If the packet is the video data request type,it is forwarded to the redirection module,otherwise the packet is sent directly to the common routing function.In the redirection module,the destination IP address of the packet is modi?ed to redirect it to one of the local peers.
Because of deep packet inspection,we just intervene in P2P traf?c,and therefore pose no threat to the bandwidth availability of other applications.On the other hand,both mechanisms have one common drawback that the format of
Packet
monitoring module
Common
routing function
Peer list modification module
Input Out put Peer list packets
Fig.4:A router architecture for the peer list modi?cation mechanism.
peer list packet must be known in advance.This makes our proposed method dependent on the protocol of applications.However,we believe that ?nding peer list formats of other P2P applications is not dif?cult by using several strong packet analysis tools.
IV.
I MPLEMENTATION
We set up a desktop PC as a software router.The hardware con?guration is as follows:Intel Core i7-26003.4GHz CPU,12GB of DDR3memory,and two 1Gbps Ethernet network interface cards,operated under Linux Ubuntu 12.04with 3.2.0-29generic kernel.
In this study,we utilized libipq [17],a library for iptables packet queueing,to capture all packets ?owing through the router.Libipq provides a mechanism to intervene to modify the packets before they arrive at the application.First,all packets are pushed into QUEUE by using iptables command,e.g.,iptables –A FORWARD –i eth0–o eth1–j QUEUE .Then libipq is used to read every packet from the QUEUE and to send it back to user space.The packet can be modi?ed by calling an API function provided by libipq library:ipq_set_verdict with reasonable pa-rameters.In the peer list modi?cation mechanism,we modify the data payload of the packet containing the peer list.In the packet redirection mechanism,the control information of the packet is modi?ed;in particular,we change the destination IP address of the packet.Note that all the check sum ?elds should be recomputed in modifying the content of packets.
For the IP-to-location mapping,in this study,we just check the countries where peers are located by utilizing GeoLite Country database,a free IP geo-location database created by MaxMind [18].
V.
E XPERIMENTAL R ESULTS
A.Experimental Setting
We performed the experiments using an existing P2PTV application,namely PPStream on the Internet.Figure 6shows our experimental network environment.For the network con-nection,we subscribed to FLET’S HIKARI NEXT,a 100Mbps optical access service on the next generation network (NGN),and plala HIKARI Mate with FLET’S as an ISP in Japan.The proposed router is placed as a gateway router.A measurement host connects to the Internet via the gateway router.The hardware con?guration of the measurement host is as follows:Intel Core i5-2440CPU 3.1GHz,4GB of memory,and a 100Mbps network interface card,operated under 64-bit Windows 7.
An on-demand drama was selected to run on the measure-ment host.For the statistical information,Wireshark was
Packet monitoring module
Common routing function
Input Output
Video data request packets
Redirection module
Fig.5:A router architecture for the packet redirection mech-anism.
installed on the measurement host.All the measurement were conducted in January 2013.
B.Results of peer list modi?cation mechanism
To prove the ef?ciency of the peer list modi?cation mech-anism,we set up two experiments.In the ?rst experiment,the peer list was in real time modi?ed while playing a prede?ned channel.The channel was played totally for 10minutes.During the ?rst two minutes,the peer list was not modi?ed.After 2minutes elapsed,we modi?ed the peer list as follows:all foreign peers were replaced by peers in Japan that had been accumulated from all the previous peer lists,i.e.,the parameter k =0.We selected 2minutes as the time period to modify the peer list because 2minutes is enough for PPStream to start exchanging video data with some peers in some different countries.In the Fig.7,we can see the change of throughput in the connection of PPStream.During the ?rst two minutes,without modifying the peer list,the application connects to peers located in many countries including China,Japan,the United States and other countries,which are bundled into one group.After the second minute,the application tends to change the connection to peers in Japan,and the traf?c from Japan is gradually increasing in particular.From the 8th minute,all the traf?c is downloaded from Japan.We can also see that the traf?c from Japan does not change immediately right after applying the method,but gradually increases.It is possible to infer that PPStream remains the connection with some peers for few minutes before switching to other peers in the new peer list.
In the second experiment,the same part of the video on demand was played three times on the measurement host.The peer list is modi?ed according to different scenarios as shown in the Fig.8.In scenario (a),without any modi?cation of the peer list,we can see that the traf?c comes from many countries including China,Japan,the United States,and other countries.In scenario (b),the peer list was modi?ed by replacing all foreign peers by peers in Japan,i.e.,the parameter k =0.As expected,almost all traf?c comes from Japan as shown in Internet
Gateway router
Fig.6:Experimental network environment setting.
Fig.7:Temporal change of throughput for PPStream with real-time the peer list modi?cation while playing a channel.
Fig.8:Temporal change of throughput for PPStream in three scenarios:(a)without peer list modi?cation,(b)replace all peers by peers in Japan,(c)remove all peers in Japan from the peer list.
the localization problem by removing all Japan peers in the peer list,i.e.,the parameter k=N,where N is the number of peers in the peer list.The traf?c from Japan is almost zero as a consequence;the green color corresponding to traf?c from Japan almost disappeared in the Fig.8(c).Note that three scenarios watch the same part of the video during the same period,5minutes,but the amounts of traf?c in three scenarios are quite different.We performed the experiments ?ve times and checked the data cached by PPStream for?ve minutes.We found that the data cached in scenario(a)was usually larger than that of the other scenarios.It means that the peer list modi?cation affects the download speed of the video. Therefore,the selection of the parameter k with a reasonable value should be taken into account.Nevertheless,the above results prove that the peer list modi?cation mechanism can be successfully applied for addressing the traf?c localization problem.
C.Results of packet redirection mechanism
We setup the experiments in a similar manner to those of the peer list modi?cation mechanism.Figures9and10 demonstrate the results of the packet redirection mechanism. It is easy to see that the results are very similar to those of the peer list modi?cation mechanism.In scenario(a)of the second experiment,without applying the packet redirection,video data were exchanged with peers in some countries such as Japan, the United States,and other countries.In scenario(b),all video data request packets sent to foreign peers were redirected to the peers in Japan.The list of Japan peers had been collected from all the previously obtained peer lists.As expected,Fig. 10(b)shows that almost all traf?c is presented in green color corresponding to the traf?c from Japan.In scenario(c),to test the negative side of the localization,all video data request packets were redirected to peers in the United States,which is very far from Japan.Figure10(c)shows that almost all traf?c comes from the United States.This demonstrates that the packet redirection mechanism realizes traf?c?ow control on PPStream application.
VI.C ONCLUSION
We proposed a router-aided approach for P2P traf?c opti-mization.By deep packet inspection,we proposed two mech-anisms:peer list modi?cation and packet redirection,which intervene in the peer selection.The former modi?es the peer list packets,whereas the latter redirects video data request packets to local peers.We also presented implementation meth-ods for both mechanisms by utilizing iptables,libipq, GeoLite Country database on a PC-based router.The experi-ments evaluated on a popular P2PTV,PPStream proved that our proposed method successfully realizes traf?c localization. Since our method is implemented on network routers,it does not require any modi?cation of existing application software. This proposal can be easily deployed on traf?c-shaping devices to help ISPs control the P2P traf?c.
There remain several problems in the current implementa-
Fig.9:Temporal change of throughput for PPStream with real-time redirection of video data request packets while playing a channel.
Fig.10:Temporal change of throughput for PPStream in three scenarios:(a)without packet redirection,(b)redirect all video data request packets to peers in Japan,(c)redirect all video data request packets to peers in the United States.
tion.First,the format of the peer list packet must be known in advance.This makes our method dependent on the protocol of the applications.Secondly,using only country information of peers is insuf?cient,and might cause performance degrada-tion,e.g.the degradation of video quality.Further study will improve this method using more information of the peers such as ISP,AS,and quality of connection links to achieve?ner-grained result.We are also planning to deploy this method to other P2PTV applications such as PPTV,SopCast,and Zattoo.
A CKNOWLEDGMENT
This study was partly supported by a Grant-in-Aid for Young Scientists(B)No.23760344from the Japan Society for the Promotion of Science(JSPS).
R EFERENCES
[1]Cisco System,Inc.“Cisco visual networking index:forecast and
methodology,2011-2016,”White paper,May2012.
[2]PPTV.[Online].Available:https://www.doczj.com/doc/b72284590.html,/.
[3]PPStream.[Online].Available:https://www.doczj.com/doc/b72284590.html,/.
[4]SopCast.[Online].Available:https://www.doczj.com/doc/b72284590.html,/.
[5]Zattoo.[Online].Available:https://www.doczj.com/doc/b72284590.html,/.
[6]H.Xie,Y.Yang,A.Krishnamurthy,Y.Liu,and A.Silberschatz,“P4p:
provider portal for applications,”in Proc.ACM SIGCOMM2008,Aug.
2008,pp.351–362.
[7]M.Costa,M.Castro,R.Rowstron,and P.Key,“Pic:practical internet
coordinates for distance estimation,”in Proc.24th Int.Conf.Distributed Comput.Syst.(ICDCS’04),March2004,pp.178–187.
[8] B.Wong, A.Slivkins,and E.G.Sirer,“Meridian:a lightweight
network location service without virtual coordinates,”in Proc.ACM SIGCOMM2005,Aug.2005,pp.85–96.
[9]V.Aggarwal,A.Feldmann,and C.Scheideler,“Can isps and p2p
users cooperate for improved performance?”ACM SIGCOMM Comput.
Commun.Rev.,vol.37,no.3,pp.29–40,July2007.
[10]R.Alimi,R.Penno,and Y.Yang,“Alto protocol,”in Internet draft,
draft-ietf-alto-protocol-10.txt,Oct.2011.
[11]R.Bindal,P.Cao,W.Chan,J.Medved,G.Suwala,T.Bates,and
A.Zhang,“Improving traf?c locality in bittorrent via biased neigh-
bor selection,”in Proc.IEEE Int.Conf.Distributed Comput.Syst.
(ICDCS2006),July2006.
[12]T.Miyoshi,Y.Shinozaki,and O.Fourmaux,“A p2p traf?c localization
method with additional delay insertion,”in Proc.4th Int.Conf.Intelli-gent Networking and Collaborative Syst.(INCoS2012),Sept.2012,pp.
148–154.
[13]T.Ng and H.Zhang,“Predicting internet network distance with
coordinates-based approaches,”in Proc.IEEE INFOCOM2002,June 2002.
[14]J.Seedorf,S.Kiesel,and M.Stiemerling,“Traf?c localization for p2p-
applications:the alto apprach,”in Proc.IEEE Int.Conf.Peer-to-Peer Comput.(P2P2009),Sept.2009,pp.171–177.
[15] D.Choffnes and F.Bustamante,“Taming the torrent-a practical
approach to reducin cross-isp traf?c in peer-to-peer systems,”in Proc.
ACM SIGCOMM2008,Aug.2008,pp.363–374.
[16]Wireshark.[Online].Available:https://www.doczj.com/doc/b72284590.html,/.
[17]Linux man page.[Online].Available:https://www.doczj.com/doc/b72284590.html,/man/3/libipq.
[18]MaxMind and GeoIP,IP address location technology.[Online].
Available:https://www.doczj.com/doc/b72284590.html,/app/ip-location/.