Future Network & MobileSummit 2010 Conference Proceedings Paul Cunningham and Miriam Cunningham (Eds) IIMC International

Information Management Corporation, 2010 ISBN: 978-1-905824-18-2

Functional Enhancements of Cooperative Session Control for Minimum Core Architecture

Takayuki WARABINO, Youji KISHI and Hidetoshi YOKOTA KDDI R&D Laboratories, 2-1-15 Ohara, Fujimino City, Saitama Prefecture 356-8502, JAPAN E-mail: {warabino, kishi, yokota}@kddilabs.jp
Abstract: This paper proposes functional enhancements of cooperative session control for Minimum Core architecture. In past work, the authors invented a cooperative session control in which the call setup time is guaranteed through cooperating core and overlay networks while minimizing processing and traffic load on the core network. In the session control, each peer selects the core or overlay network for session establishment based on the measured latency on the overlay network. Proprietary technology (the so-called Blacklist method) was applied for latency decision on the overlay network. However, the performance of the native Blacklist method degrades during peer join and leave events. Accordingly, this paper proposes two approaches: Voluntary Latency Acquisition (VLA) and Blacklist Sharing (BS) as enhancements of the Blacklist method. These approaches allow the efficient detection of new blacklist entries that emerge due to peer join and leave events. This paper also reports simulation results that reveal the basic properties of the invented session control in a large-scale network and its performance during peer join and leave events. Keywords: New Generation Network, Minimum Core, Overlay Network, P2PSIP, DHT, Chord.

1. Introduction
ICT (Information and Communication Technology) is adopting the role of a new infrastructure for production activities since it overcomes constraints of time and location by exchanging a huge amount of information at once [1]. In particular, mobile communication systems have experienced remarkable growth and they have become prevalent as social infrastructure. Since their development will be further accelerated thanks to innovative wireless technologies, the performance of mobile systems will be comparable to that of fixed-line communications. On the other hand, another trend will see user devices continue to evolve. In addition to existing personal computers and cellular phones, new types of devices such as home appliances and car navigation and sensor devices are connecting to the network. In such a ubiquitous environment, the number of network devices will overwhelm that of current mobile subscribers. Based on the above prospects, the authors considered that the New Generation Network (NWGN) over the next few decades should: x Converge fixed and mobile networks while supporting mobility as one of its fundamental functionalities x provide high-speed access (i.e., over 100Mbps - 1Gbps) even when user devices are connected through wireless links x have the capability for a tremendous volume of network devices
Copyright 2010 The authors www.FutureNetworkSummit.eu/2010 Page 1 of 9

Signaling

Data

SBS: Small Base Station, MM: Mobility Management, AS: Application Server

Network functionalities are minimized by shifting them to access network Registrar/MM

Minimum Core

AS

AAA

Overlay Network (Virtual)

Access/ Transport Network

AS SBSs Network functions are distributed into a vast number of SBSs

User device

User device

Fig 1. Minimum Core architecture

The authors invented Minimum Core as a key concept toward NWGN [2]. In the concept, network functionalities in a core network are minimized by shifting them to access networks. This avoids traffic concentration on the core network by localizing the signaling and data traffic within the access networks. This approach ensures the high scalability that is vital for future network architecture. In addition, a session control method was invented as one of the key technologies for Minimum Core [2]. In this method, call setup time is guaranteed through cooperating core and overlay networks, while minimizing processing and traffic load on the core network. In this method, proprietary technology (the so-called Blacklist method) was applied for latency decision on the overlay network. However, the performance of the native Blacklist method degrades during peer join and leave events. In order to handle these situations, this paper proposes two approaches: Voluntary Latency Acquisition (VLA) and Blacklist Sharing (BS) as enhancements of the Blacklist method. These approaches allow the efficient creation of new blacklist entries that emerge due to peer join and leave events. This paper also reports simulation results that reveal the basic properties of the invented session control in a large-scale network and its performance during peer join and leave events.

2. Architecture and Techniques for Minimum Core concept

2.1 Minimum Core Architecture To embody NWGN (NeW Generation Network), the authors invented Minimum Core architecture (Fig. 1). In this architecture, network functionalities in the core network (e.g., mobility management, session control and application servers, etc.) are minimized by shifting them to the access networks. In current network architectures such as NGN, the system is composed of service and transport stratums and transport stratum is divided into core and access networks. In this architecture, all service nodes are allocated in/over the core network and all signaling and data traffic is necessarily routed via the core network even for communications between user devices. In contrast, the Minimum Core architecture avoids traffic concentration on the core network by localizing signaling and data traffic within access networks. This approach ensures high scalability as future network architecture. In our current design, the network functions are allocated in small base stations (SBSs) such as Wireless LAN and Femtocell access points. They have scalability advantages since
Copyright 2010 The authors www.FutureNetworkSummit.eu/2010 Page 2 of 9

R(x,y): RTT from Peer x to Peer y

7, R(3,7) 19, R(3,19) 60 37, R(3,7)
57

ID=33, R(29,33) 60 ID=29, R(24,29) 57 ID=24, R(19,24)

7

1. Request
57

60

2. Response
19

42

Ping/Pong
24

42

ID=33, R(29,33) ID=29, R(24,29)

24 29

7 If R(3,19) + R(19,24) > th., ID space from Peer 24 to Peer 33 is registered

42

19

19

ID=33, R(29,33)
37 33

Registered ID space
37 33 29

24

37

33

29

(a) Latency measurement

(b) Latency acquisition

(c) Blacklist registration

Fig. 2. Blacklist registration

a vast number of SBSs will be deployed in future networks. Our immediate goal involves efficiently providing various types of services in the Minimum Core architecture without degrading current usability such as mobility performance, service quality and security.

2.2 Cooperative Session Control The authors invented a session control method for Minimum Core [2]. The method integrates the core network with an overlay network composed of SBSs. The overlay network is a virtual network on top of IP networks. While session control on the overlay network have been developed so far (e.g., dSIP [3], P2PP [4] and P2PSIP [5]), these technologies are utilized by some systems such as Skype [6]. It is assumed that P2PSIP, which is currently standardized in IETF, is utilized for session control on an overlay network. In the P2PSIP, the functionalities of existing SIP servers are distributed into user devices and which compose the overlay network, whereupon a serverless systems is realized. For the overlay construction and management, P2PSIP utilizes prominent DHT (Distributed Hash Table) technologies such as Chord [7], Bamboo [8] and Tapestry [9]. DHT is an abstract hash table service realized by storing the hash table contents across a set of peers. Out standing features includes absence of centralized server, system scalability and autonomous network management etc. However, current P2PSIP technologies are based on a best-effort principle and lack mechanisms to guarantee the quality of call setup time (CST). In contrast, our invented method guarantees CST through cooperating core and overlay networks, while minimizing processing and traffic load on the core. In cooperative session control, each peer selects the core or overlay network as the system to be used for session establishment based on the measured latency on the overlay network. For system selection, the authors have invented the Blacklist method. In this method, each peer measures latency on the overlay network by utilizing maintenance traffic, etc. and registers an ID space, for which the latency exceeds the desired threshold, in a blacklist. When initiating a session, each peer checks whether the destination ID is included in the blacklist or not. If the destination ID is not included in the blacklist, the overlay network is used. Otherwise, the core network is used. 2.2.1 Basic Operations of Blacklist Creation In the following, the blacklist creation procedures are explained when Chord [7] is used as a DHT algorithm (Fig. 2). In the Chord, each peer maintains a neighbor table (a so-called finger table) for efficient overlay routing.

Copyright 2010 The authors

www.FutureNetworkSummit.eu/2010

Page 3 of 9

(a) Latency measurement Each peer periodically measures the RTT (Round Trip Time) to other peers registered in its own finger table (Fig. 2 (a)). In the measurement, a keep-alive message to the fingers, which represents one type of maintenance traffic in the Chord, can be utilized. (b) Latency acquisition Each peer acquires the measured RTTs on a message path by utilizing the maintenance and other messages on the overlay network. In the actual procedures, when a peer relays a response message, the measured RTT to the previous relayed peer is attached to the message as well as the ID of the previous peer. In Fig. 2 (b), Peer 29 adds RTT to Peer 33, while Peer 24 adds RTT to Peer 29. Consequently, the peers receiving and relaying response messages can acquire latency on the relayed path on a hop-by-hop basis. (c) Blacklist registration After (b), a peer conducts blacklist registration based on the acquired RTTs. A feature of blacklist registration is that it allows for the peers to register a range of ID space to which latency exceeds a preconditioned threshold. In the detailed method, the peer totals up the RTTs from the closer peers. During the addition, when the sum of the RTTs exceeds the threshold, the peer registers the ID space from a corresponding peer to a responding peer. For example, in Fig. 2 (c), since the sum of R(3,19) and R(19,24) exceeds the desired threshold, the ID space from Peers 24 to 33 is registered in the blacklist. The blacklist has three fields: From, To and Expiration time. In the above case, the IDs of Peer 24 and 33 are set at the From and To fields, respectively. The expiration time is set based on the registered time and used for blacklist maintenance. (d) Blacklist maintenance If the expiration time of a blacklist entry expires without any updates, a peer sends a dedicated message destined to an ID that is included in a To field. The peer then updates information on the entry when receiving a response message.

3. Functional Enhancements for Handling Peer Join and Leave

In Minimum Core, the overlay network consists of SBSs such as wireless LAN and Femtocell access points. The peer join and leave rates are low compared with ordinary P2P applications since SBSs are usually powered on and stably connected to the access networks. However, once a join and leave event occurs, they require the blacklists to be updated on the overlay network. This section analyses the situations and proposes some extensions of the Blacklist method. 3.1 System Analysis of Impact of Peer Join and Leave When a peer joins or leaves, routing tables in other peers are automatically updated according to the maintenance mechanisms of DHT. Since these updates change the routing paths of overlay messages, the blacklists in the overlay network must be updated. There are two cases regarding blacklist updates: (a) the creation of a new blacklist entry and (b) an ID space change/deletion of an existing entry. Fig. 3 shows simple examples of both cases. In both cases, when Peer B leaves the network, Peer F is added in the routing table of Peer A instead of Peer B. In the case of (a), since the routing update causes an increase in latency to Peer E, the ID space that should be registered in the blacklist of Peer A newly emerges. In the case of (b), Peer A already has an entry from Peer C to Peer E. Since the routing update decreases the latency to Peer E, the ID space of the entry should be decreased. In addition, the expansion of the ID space of an existing entry and the deletion of the entry are included in case (b). Meanwhile, case (b) can be handled by the blacklist maintenance shown in Section 2.2.1 (d) since the latest status of existing entries is checked by sending dedicated messages
Copyright 2010 The authors www.FutureNetworkSummit.eu/2010 Page 4 of 9

No Blacklist Peer

When Peer B departs, Peer F is added in routing table of Peer A

Threshold of blacklist registration: 500 ms

Case (a)

250ms

(2)
100ms 100ms

RTT

100ms 100ms 100ms 100ms

Blacklist entry (from C to E)

Peer

A blacklist entry has to be created (from C to E)

Case (b)

RTT

250ms 250ms 100ms 100ms

200ms

(3)

100ms 100ms

ID space has to be updated (from D to E).

(4)

Fig. 3. Simple examples of impact of peer join and leave events

after expiration. However, case (a) cant be fully supported by current mechanisms. In the basic operations shown in Section 2.2.1, new blacklist entries are not created until maintenance or other messages are sent through the corresponding ID space. Then, when the peer join and leave rates increase, performance degrades. In addition, blacklist maintenance cant handle case (a) since it has no mechanisms to detect new blacklist entries. Accordingly, this paper proposes the following two approaches to realize efficient detection of new blacklist entries. 3.2 Proposed Enhancements of Blacklist Method 3.2.1 Voluntary Latency Acquisition (VLA) In this method, each peer selects an ID on the overlay network and sends a dedicated message destined to the selected IDs (destination ID). When the peer receives a response message, the blacklist registration procedures are conducted as shown in Section 2.2.1 (c). This method allows the detection of new blacklist entries that emerge due to peer join and leave events. For efficient blacklist creation, it is desirable to select the destination ID so that it can take long time for the message to reach the destination. As one can imagine, latency increases according to the number of message hops. Then, an estimation technique of the number of message hops is invented and utilized for the ID section. Fig. 4 shows the pseudo-code of the estimation method. (1) Each peer first selects a target ID in a random manner. (2) The peer derives the distance from its own ID to the target ID and the average ID space that one peer is responsible for. The average ID space can be calculated by dividing the ID space covered by successors registered in a successor list by the number of successors. (3) The peer estimates the number of message hops to the target ID by simulating the original Chord shortcut routing. In each shortcut in the overlay network, the distance to the target ID shortens by 2i (i is an integer decreasing from 1591 to 0). The shortcut continues until the distance becomes smaller than the average ID space. (4) If the estimated number of message hops is larger than the threshold, the peer sends a dedicated message to the selected target ID. Otherwise, the procedures are repeatedly conducted from (1) to (4). In summary, the method realizes efficient detection of new blacklist entries by utilizing the invented estimation technique of the number of message hops.
1

It is derived based on the length of the ID in the Chord (160 bits).

www.FutureNetworkSummit.eu/2010 Page 5 of 9

Copyright 2010 The authors

targetID = randID(); // generate a random target ID distance = targetID - ownID ; // calculate ID distance from own ID to target ID. interval = cal_interval(); // calculate average ID interval which one peer is responsible for. eHop = 1; // initialize an expected hop count. for (i=159; i > 0; i--) { if (distance < interval) break ; // if the distance is smaller than interval, finish increment of eHop if (distance >= pow (2,i) ) { eHop ++; // increment eHop for shortcut on Chord network. distance - = pow (2,i);} } if (eHop > threshold) request_sendto(targetID ); // if the eHops is larger than threshold, send an overlay message to target ID. Otherwise, go back to first line of pseudo-code.

(1)

(2)

(3)

(4)

Fig. 4. Pseudo-code of estimation of message hops for Voluntary Latency Acquisition (VLA)

3.2.2 Blacklist Sharing (BS) In addition, blacklist sharing (BS) enables efficient blacklist creation by sharing blacklist information between peers. The procedures of blacklist sharing are as follows (Fig. 5): (1) Each peer periodically sends blacklist sharing requests toward its own fingers, which carry the IDs of both the destination finger (Peer 19) and its next finger (Peer 37). (2) A peer receiving the request searches the entries existing between the two IDs attached in the request within its own blacklist. The peer sends a response message to the originator (Peer 3) with the corresponding entries. (3) When the originator receives the blacklist sharing response, it compares received entries with entries registered in its own blacklist and checks whether the response includes new blacklist entries. (4) If the new entries exist, the originator sends dedicated request messages destined to IDs, which are set at the To field of the entries, in order to reflect the most recent status of the overlay network. Corresponding response messages are returned to the originator. (5) Finally, when the originator receives a response message, it conducts the blacklist registration procedures as shown in Section 2.2.1 (c). (6) Each peer repeatedly conducts the above procedures (i.e., from (1) to (5)) to all fingers. This method has two effects for efficient blacklist creation. When a peer detects a new blacklist entry, the method allows the peer to quickly share the information with other peers. In addition, the method is effective when a new peer joins the network. Since the new peer has no blacklist, the method helps the peer create a network-wide blacklist immediately after joining.

4. Simulation Evaluations
4.1 Evaluation Setup The authors have developed a simulator based on Oversim [10] (an open-source overlay simulator) and have conducted evaluations to validate the proposed cooperative session control and additional functionalities of the Blacklist method. The evaluation parameters are shown in Table 1. In these evaluations, the authors evaluated the basic properties when changing the number of peers up to 16,000 and the impact of peer join and leave events.
Copyright 2010 The authors www.FutureNetworkSummit.eu/2010 Page 6 of 9

Existing blacklist entries

Peer 3(black) Peer 19 (gray)

3
(1) Blacklist sharing request (Peer19, Peer37) (2) Blacklist sharing response (a,b,c,d)
(3) Detection of new entries (b, c)

19

37

(4) Dedicated request/response messages (to b and c)

(5) Blacklist registration (b, c)

Fig. 5. Procedures of Blacklist sharing

Stub domain

Table 1. Evaluation parameters

Number of peers Evaluation period Target CST (Blacklist registration threshold) Session initiation model DHT algorithm Stabilize interval Fix finger interval Network model up to 16,000 2 [hour] 3 [sec] (1.5 [sec]) Poisson distribution Chord 15 [sec] 120 [sec] Transit-Stub model T-T link: 100 [ms] T-S link: 10 [ms] S-S link: 1[ms]

Core Transit domain (Core transport)

Transit-Transit link Transit-Stub link Stub-stub link

Small BSs (Peer)

Stub domain (Access transport)

Fig. 6. Network model for evaluations

As a network model, a traditional transit-stub model [11] was used since its architecture resembled the assumed future network (Fig. 6). Transit and stub domains correspond to core and access networks, respectively. In the model, all peers were allocated in stub domains and the packet transmission delay was configured based on previous studies [7][12]. The target CST was set to 3 [s] according to Japanese regulations [13]. Moreover, each measurement was conducted by starting the actual measurement after a 2-hour test run. During the test run, each peer conducted voluntary latency acquisition and intentionally sent overlay messages to destination IDs in order to create as many blacklist entries as possible.
4.2 Evaluation Results (a) Basic properties of proposed session control in a large-scale network First, the authors have evaluated CST in terms of both the proposed session control and a conventional P2PSIP in order to validate the former. In the P2PSIP method, the session being established is always conducted on the overlay network. Distributions of CST are shown in Figs. 7 when changing the number of peers. These figures indicate a similar tendency, while the CST in P2PSIP slightly increases according to the number of peers. Table 2 summarizes these results. For example, when the number of peers is 16,000, the percentage of CST exceeding the desired time reaches 12.97 [%] in P2PSIP. In contrast, the proposed methods suppress that to be less than 0.4 [%], which represents approximately one-thirtieth of that of P2PSIP. These results also show that the proposed methods satisfy the Japanese regulations [13], which stipulate that the probability
Copyright 2010 The authors www.FutureNetworkSummit.eu/2010 Page 7 of 9

30 25 Proposal P2PSIP

30 25 Proposal P2PSIP

Percentage [%]

20 15 10 5 0 0.0 0.6 1.2 1.8 2.4 3.0 3.6

Percentage [%]

20 15 10 5 0 0.0 0.6 1.2 1.8 2.4 3.0 3.6

Call Setup Time [sec]

Call Setup Time [sec]

(a) 8,000 peers

(b) 16,000 peers

Fig. 7. Distribution of the call setup time (CST) Table 2. Evaluation results of CST
Probability of CST exceeding 3 seconds P2PSIP Proposal 2.63 6.69 12.97 0.14 0.23 0.39 Average CST [sec] 4,000 P2PSIP 2.05 2.22 2.36 Proposal 1.86 1.86 1.82 8,000 16,000

Table 3 System utilization

Via overlay 89.06 [%] 79.61 [%] 70.32 [%] Via core 10.94 [%] 20.39 [%] 29.68 [%]

# of peers 4,000 8,000 16,000

of CST exceeding 3 [s] should be less than 1 [%]. In addition, Table 3 shows the percentage of selected systems for the proposed methods. As shown in Table 3, the proposed methods can reduce the processing burden of the core by up to roughly 70 [%]. When the core network is dimensioned based on system occupation which is defined as the average arrival rate divided by the average service rate in queuing theory. Given the constant system occupation, the service rate is proportional to the arrival rate. This implies server facilities dimensioned in the core network are proportional to the load imposed on the core network. Based on the above notion, the evaluation result shows that proposed approach can reduce server facilities in the core network by 70 [%] compared to existing centralized approach. (b) Performance characteristics when peers join and leave Next, the impact of peer join and leave events on the proposed session control was investigated. As explained in Section 3, the authors proposed two approaches (i.e., voluntary latency acquisition (VLA) and blacklist sharing (BS)) to handle peer join and leave events. Fig. 8 shows the evaluation results of CST, which is less than 3 [s]. In the evaluations, one peer joins and another leaves at constant intervals. The model maintains a constant number of joining peers. The number of peers is 10,000. The intervals of voluntary latency acquisition and blacklist sharing are 10 [s] and 15 [s], respectively. As shown in Fig. 8, each method improves the performance compared with the native Blacklist method (w/o extensions in Fig. 8). In particular, utilizing the two proposed methods achieves a probability of 99 [%] when the join/leave interval exceeds 1 [s]. The interval corresponds to each peer remaining within the network for 2.8 [hours]. In contrast, the proposal without extensions achieves a probability of 99 [%] when the join/leave interval exceeds 20 [s]. Throughout the evaluations, it was confirmed that the two proposed approaches can improve the performance of the Blacklist method when the peers constantly join and leave.

5. Conclusions
This paper proposes two approaches as enhancements of our proprietary Blacklist method: Voluntary Latency Acquisition (VLA) and Blacklist Sharing (BS). While the native procedures of the Blacklist method cant handle peer join and leave events, these methods allow the detection of new blacklist entries that emerge due to peer join and leave events.

Copyright 2010 The authors

www.FutureNetworkSummit.eu/2010

Page 8 of 9

100 99 98 97 96 1 10 100

Probability [%]

Proposal (w/o extensions) Proposal with VLA Proposal with BS Proposal with VLA and BS

Join/Leave interval [sec]

Fig. 8. Probability of CST less than 3 [sec] when changing join/leave interval (10,000 peers)

This paper also reports simulation results in a large-scale network. These results reveal the basic properties of the proposed session control and performance during peer join and leave events. For example, when the number of peers is 16,000, the probability is reduced from 12.97 [%] in conventional P2PSIP to 0.4 [%]. In addition, the proposed methods mitigated the processing burden of the core by roughly 70 [%] in comparison with the current centralized architecture. It was also confirmed that the two proposed approaches could handle low peer join/leave rates. The authors will conduct detailed evaluations on a large-scale network. In addition, there are plans to implement the proposed mechanisms into actual SBSs and conduct system evaluations with a core system prototype.

Acknowledgments
This work is supported by the National Institute of Information and Communications Technology (NICT) of Japan.

References
[1] Ministry of Internal Affairs and Communications, Japan, Information and Communications in Japan, White Paper 2008, July, 2008. Available: http://www.johotsusintokei.soumu.go.jp/whitepaper/eng/WP2008/2008-index.html [2] T. Warabino, et al., Session Control Cooperating Core and Overlay Networks for Minimum Core Architecture IEEE Global Communications Conference (GLOBECOM 2009), Dec. 2009. [3] D. Bryan et al., dSIP: A P2P Approach to SIP Registration and Resource Location, draft-bryan-p2psip-dsip-00, Feb. 2007. [4] S. Baset et.al., Peer-to-Peer Protocol (P2PP), draft-baset-p2psip-p2pcommon-01, Feb. 2007. [5] IETF P2PSIP WG. Available: http://www.ietf.org/html.charters/p2psip-charter.html [6] Skype. Available: http://www.skype.com [7] I. Stoica, et al., Chord: A Scalable Peer-to-peer Lookup Service for Internet Applications, Proc. ACM SIGCOMM'01, Sept. 2001. [8] S. Rhea et al., Handling Churn in a DHT, Proceedings of the USENIX Annual Technical Conference, Jun. 2004. [9] B. Y. Zhao et al., Tapestry: A Resilient Global-Scale Overlay for Service Deployment, IEEE Journal on Selected Areas in Communications, 2004. [10] The OverSim P2P simulator. Available: http:// www.oversim.org/ [11] Zegura, et al., A quantitative comparison of graph-based models for Internet topology, IEEE/ACM Transactions on Networking, Dec. 1997. [12] S. Ratnasamy, et al., A Scalable Content-Addressable Network, Proc. ACM SIGCOMM'01, Sept. 2001. [13] Ministry of Internal Affairs and Communications (in Japanese). Available: http://law.egov.go.jp/htmldata/S60/S60F04001000030.html

Copyright 2010 The authors

www.FutureNetworkSummit.eu/2010

Page 9 of 9