極大型網路多媒體服務之研究

(1)

Abstract

目前已有許多具有動態群組與服務品質以支援及時網際網路群播應用之研究。各種現存之貪婪式服務品質感知的選徑通信協定,如YAM與QoSMIC,仍有不具良好的擴充性之問題。另外,大部份現存之動態通信協定對群組密度有不同之適用性。在本報告中,我們提出一分散式路徑選擇之通信協定(稱為DSDMR)。它對群組密度具有自我調適能力且能在極低控制信息負荷與低加入時間延遲之情況下建立近乎最佳貪婪式方法所產生之低成本樹。

關鍵詞: 群播、動態選徑、品質感知選徑 Much research has been done on support- ing real-time IP multicast applications with dynamic group member joining and QoS requirements. Various existing greedy QoS- aware routing protocols, such as YAM and QoSMIC still suffer from the problem of poor scalability. In addition, most existing dyna- mic protocols have different group density adaptability. In this report, we introduce a distributed candidates selection protocol, named DSDMR, that is capable of self- adaptation depending on sensed group densi- ties and create low cost tree close to optimal greedy strategy with very low control over- head and join latency.

Keywords: multicast, dynamic routing, QoS routing

1. Motivation and Goal

Much existing protocols [7] have devoted to address the dynamic nature of the net- work. There exists two major categories of

極大型網路多媒體服務之研究(I)

Research on Very Large Scale Multimedia Services (I)

計畫編號：NSC 89-2213-E-011-020

執行期限：88年8月1日至89年7月31日

主持人：陳秋華博士國立台灣科技大學電子所

technology to handle network dynamics.

The first is tree reconstruction that puts the major concern on rebuilding optimal delive- ry tree. The second major category is tree maintenance that emphasize fast join, but the result may be suboptimal tree temporal- ly. Many protocols have been proposed using this kind of mechanism [1], [2], [3], [4], [5], [6] and it is the major category we focus on in this report.

Dynamic IP multicast routing protocols can be classified into QoS-unaware ones including DVMRP, PIM, and CBT or QoS- aware ones including YAM [1] and QoSMIC [2]. The later is capable of finding the most suitable path for a new member to attach to existing multicast tree while meeting the user's QoS requirement. Ways to find the path can be further separated into two approaches: from new member side, named local search, or from existing multicast tree side, named multicast tree search. Both YAM and QoSMIC are greedy based multiple paths protocols, where YAM uses local search only but QoSMIC finds candidates from both sides concurrently.

The major concerns for a good QoS routing strategy used in real Internet include routing efficiency, scalability, and distribut- ed computation. Besides, we are particularly interested on the aspect of group density adaptability in this report since most proto- cols have biased density proclivity. We propose a density-sensitive dynamic routing protocol, named DSDMR, to achieve effici- ent management of network resources under constrained and unconstrained delay. We also evaluate our group density adaptability algorithm against other QoS-aware proto- cols on all major metrics.

1

(2)

2. Related works

Research on routing algorithms to address network dynamics started from 1988 by Waxman [5], [6]. Thereafter, four well- known IP multicast routing protocols — PIM, DVMRP, MOSPF, and CBT have been proposed [7]. Later greedy proposals YAM [1] and QoSMIC [2] that examine multiple paths from receiver to tree.

In YAM's algorithm, when absent group state in egress node, it starts to flood request bidding “reqbid” packets. Upon reaching an on-tree node, the spanning procedure will be terminated and a “bid” packet will be replied toward the initiator accompanied by install- ing a temporary state and collecting QoS properties along the path. Finally, the initiat- or selects the best one among multiple candi- dates and joins group by sending a “join”

packet toward the chosen on-tree node.

The QoSMIC including both centralized and distributed version (named QoSMIC- centr and QoSMIC-distr respectively) was proposed to overcome the control overhead problem of YAM. The new router uses local search procedure to flood “reqbid” packets with a TTL value equal to 2. On-tree routers receiving the packets become candidates and respond “bid” packets back toward the initiator. At the same time, a multicast tree procedure proceeds by the manager selects some on-tree routers as candidates and then forward "bid" packets to new router as YAM does. Finally, the candidate selection is performed also the same as YAM does.

Some problems of previous works had been pointed out in [3], [4]. YAM will produce large number of control messages from flooding of spanning join in the case of sparse group. QoSMIC-centr has the draw- back of un-robustness and needing priori information for group manager. QoSMIC- distr floods the multicast tree thus suffers from the problem of large overhead for dense groups. In this report, those problems will be solved by our new proposed DSDMR strate- gy as depicted in section 4 which is based on a distributed candidate selection method as stated in next section.

3. Distributed candidate selection

In QoSMIC, both centralized and distri- buted version use a group manager and employ both local search and multicast tree search in parallel. The distributed version is more practical due to its low overhead and needing information from neighbor only.

Here, we propose a tree join with distributed decision algorithm (abbreviated as TJDD) in Algorithm-I based on proxy rule. That is similar to QoSMIC-distr multicast tree search, but without using of group manager and local search. The algorithm serves as a center piece in later protocol development.

Algorithm-I: Tree Join with Distributed Decision s1. New member node sends “mjoin” packet to group

source (or core root) and starts “Bid-timer”.

s2. Source multicasts “order-bid” packet to all on- tree nodes.

s3. Each on-tree node uses proxy procedure to decide whether to become temporary-candidate or not, then sends “bid” packet toward new member node if it is a temporary candidate.

s4. The “bid” packet can be 'taken-over' by other on- tree node.

s5. Upon reaching new member node, the temporary- candidate implied by “bid” packet becomes real- candidate.

s6. Upon “Bid-timer” timeout, new member node selects a best candidate among real-candidates, then sends “join” packet toward the candidate and sets up state along path.

The proxy rule is given as follows, for each on-tree node that receives “order-bid”

packet, if its upstream neighbor is at a less or equal distance to new node or at least one of its downstream neighbors is at a less distance to new node, then its candidate right is replaced by the neighbor, otherwise it become the temporary candidate. The rule plays the central role for the TJDD algori- thm to select on-tree candidates.

Figure 1 shows that a new member 16 attempts to join an existing multicast tree shown by bold edges with a source 8. After receiving "mjoin" packet from 16, root 8 starts to multicast "order-bid" packets to all on-tree nodes. After proxy steps, four temporary candidates numbered 7,9,10, and 11 are left. They may be 'taken-over' while

"bid' packets are sent toward 16. Candidate right of 7 is taken over by 10. Finally, the best one will be chosen by 16 among final candidate 9, 10, and 11.

2

(3)

4. Density sensitive routing

The proposed distributed TJDD mechani- sm, as well as QoSMIC-distr, can create low cost tree, but it still incurs high control over- head when members are densely populated.

The tendency exists for other protocols since each has different group density adaptability.

Two major joining ways may be used to find suitable candidates on existing tree. One is local join using spanning join from new member side and the other is multicast tree join using group manager or distributed candidate selection heuristic from multicast tree side. All previous approaches are non- adaptive for group densities. YAM's local join by flooding cannot be well suitable for sparse group and TJDD's multicast tree join using flooding multicast tree cannot perform well for dense group. QoSMIC mechanism is non-adaptive for its concurrent two-direction join because of the redundant searching by multicast tree join for dense group.

The Density Sensitive Dynamic Multicast Routing (DSDMR) protocol in Algorithm-II proceeds joining by local join first and then switches to multicast tree join if needed. For example, in figure 1, assume that a TTL value is predefined to 2. For new member node 16, the protocol can find candidates 9,10,11,15 in local join phase and chooses the best candidate after Bid1 timer timeout. No redundant multicast tree search will be triggered. For new member node 20, the DSDMR will automatically trigger tree join phase since no candidate is found by local search after Bid1 timer timeout. Candidate nodes numbered 9,10, and 11 will be found and the best candidate will be chosen in tree join phase after Bid2 timer timeout.

Algorithm-II: DSDMR routing protocol

s1. (local join phase) New member node starts

“Bid1-timer” and floods “req-bid” packets with limited TTL value toward group source to search on-tree nodes, thereupon reply “bid” packets by reached on-tree nodes.

s2. Upon “Bid1-timer” timeout, if new member node can select a best candidate, then goto s4, else trigger tree-join procedure by enter s3.

s3. (tree join phase) New member node starts “Bid2- timer” and sends “mjoin” packet to group source.

The source thereupon proceeds distributed candidate selection by multicasting “order-bid”

packets to on-tree nodes and replied “bid” packets toward new member node by temporar candidates.

The "bid" packet will be 'taken-over' if need.

s4. Upon “Bid2-timer” timeout, the new member node selects a best candidate, then sends “join”

packet toward it and set up the state along path.

The DSDMR is a greedy heuristic that belongs to dynamic QoS-aware routing protocol using multiple-paths mechanism.

Based on density-sensitive strategy of adaptive two-direction join, it is adaptable for both dense and sparse group by using self-adaptability trigger.

5. Simulation

In this report, two network models based on the real Internet structure [10] are used to evaluate various routing protocols by a network simulator ns-2 [9]. The first model, named 'mbone84', is based on 94's major MBone routers and links [8] that includes 84 major routers. The second model is random transit-stub network, e.g. 'random50' for 50 nodes random topology.

A. Group Density Effect Experiments. The group density represented by percentage of member nodes in the group with relative to total nodes will change from 5% to 100%.

Firstly, the results for cost of multicast tree show that almost the same performance for all GRD protocols with difference within +/- 2% relative to OPT-GRD. Where OPT-GRD is an optimal version of greedy strategy which exhaustively replies "bid" packets by all on-tree nodes. Secondly, in figure 2(a), the result for join latency exhibits well performance for DSDMR and YAM.

Thirdly, in figure 2(b), the result of control overhead comparison between GRD protocols exhibits the best performance for DSDMR with about 2 times better than QoSMIC and 2~8 times better than TJDD.

3

represents candidate right of node a (d = 4) replaced by node b (d = 3) by using proxy procedure represents candidate right of node c replaced by node d by using take-over procedure P

P P

P

4

1 2 3

5 6 7

8 9 10

11 12 13

14 15 16 17

18 19 20

3

2

3 2 3

2 3

4 4

3 4

2 P

P P

P

3 P

T

a b

4 3

P

c d

T

Figure 1. An example for TJDD algorithm.

(4)

B. Scalability Experiments. By increasing node count in figure 3(a) and 3(b), which represent 20% and 70% densities respective- ly. For sparse group in figure 3(a), YAM isn't scale well caused by large flooding. For dense group case in figure 3(b), QoSMIC don’t scale well caused by redundant multicast tree joins, while DSDMR still exhibits well scalability for dense group caused by it's self-adaptive capability.

By increasing average node degree in figure 4(a), YAM is sensitive to network degree at sparse case. In figure 4(b), the bad performance of YAM is reduced since flooding effect is alleviated at dense case, while DSDMR still exhibit better scalability of topology degree. Thus, the poor scalability exists for YAM at sparse case and for QoSMIC at dense case, but not for DSDMR.

C. Constrained Delay Experiments. We examine DSDMR's capability by considering delay violation ratio under both constrained and unconstrained delay cases. Constrained delay version of DSDMR clearly can select the best path among multiple paths to meet the constrained delay. Thus it can alleviate delay violation ratio with relative to unconstrained delay version of DSDMR.

6. Contributions and Conclusions

Existing QoS-aware routing protocols such as YAM and QoSMIC outperforms other approaches by being able to select the qualified candidate meeting the desired QoS requirement from multiple paths. The YAM

protocol has the problem of high control overhead in case of sparsely populated group, while QoSMIC protocol has the drawback of relying on centralized group manager and exists poor performance in dense case due to its redundant concurrent multicast tree search. We propose a distribu- ted multiple candidates selection heuristic, named DSDMR, with self-adaptability capability that can adapt its behavior in the face of different group densities. From the simulation results, the DSDMR exhibits good performance while having very low tree cost close to optimal greedy strategy.

References

[1] K. Carlberg and J. Crowcroft, "Building Shared Tree Using a One -to-many Joining Mechanism", ACM Computer Comm. Review, January 1997, pp. 5-11.

[2] M. Faloutsos , A. Banerjea, and R. Pankaj, "QoSMIC:

Quality of Service Sensitive Multicast Intern Protocol

", ACM SIGCOMM, Vancouver, Sept. 2-4, 1998.

[3] S. Chen, K. Nahrstedt, and Y. Shavitt, "A QoS-Aware Multicast Routing Protocol", INFOCOM 2000.

[4] A. Fei and M. Gerla , "Receiver-initiated Multicasting with Multiple QoS Constraints", INFOCOM 2000.

[5] B.M. Waxman, "Routing of Multipoint Connections

", IEEE J. Sel. Areas Comm., 1988, 6(9):1617-1622.

[6] M. Imase and B. M. Waxman, "Dynamic Steiner Tree Problem", SIAM J. Disc. Math., Vol. 4, No. 3, August 1991, pp. 369-384.

[7] J. Lin and R.S. Chang, "A Comparison of The Internet Multicast Routing Protocols", Computer Comm. 22, 1999, pp. 144-155.

[8] S. Casner, "Major MBONE routers and links", from ftp.isi.edu:mbone/mbone-topology. ps, 1994.

[9] UCB/LBNL/VINT, "Network Simulator ns-2", http://www-mash.cs.berkeley.edu/ns/.

[10] K.Calvert, M.Doar, and E.W.Zegura, "Modeling Inter- net Topology", IEEE Comm. Magazine, June 1997.

4 (a) join latency comparison

(b) control message overhead comparison Figure 2. Group density effect for mbone84.

0 0.1 0.2 0.3 0.4 0.5

0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95

group density

ave. join latency (sec)

QoSMIC YAM DSDMR OPT-GRD TJDD

1 10 100 1000

0.1 0.2 0.3 0 . 4 0.5 0.6 0.7 0.8 0.9 1