当前位置:首页 >> >> A new distributed route selection approach for channel establishment in real-time networks

A new distributed route selection approach for channel establishment in real-time networks

A New Distributed Route Selection Approach for Channel Establishment in Real-time Networks
Hariharan Shankar Rahul G. Manimaran Dept. of Electrical and Computer Engineering Laboratory for Computer Science Massachusetts Institute of Technology Iowa State University Ames, IA 50011 USA Cambridge, MA 02139, USA (rahul@lcs.mit.edu) (gmani@iastate.edu) C. Siva Ram Murthy Dept. of Computer Science and Engineering, Indian Institute of Technology Madras 600 036, INDIA (murthy@iitm.ernet.in)

Abstract
In this paper, we propose a new distributed route selection approach, called parallel probing, for realtime channel establishment in a point-to-point network. The existing distributed routing algorithms fall into two major categories, namely, preferred neighbour based or ooding based. The preferred neighbour approach o ers better call acceptance rate, whereas the ooding approach is better in terms of call setup time and routing distance. The proposed approach attempts to combine the bene ts of both preferred neighbour and ooding approaches in a way to improve all the three performance metrics simultaneously. This is achieved by probing k di erent paths in parallel, for a channel, by employing di erent heuristics on each path. Also, the proposed approach uses a notion called intermediate destinations (IDs), which are subset of nodes along the least-cost path between source and destination of a call, in order to reduce the excessive resource reservations while probing for a channel by releasing unused resources between IDs and initiating parallel probes at every ID. Further, it has the exibility of adapting to di erent load conditions by its nature of using di erent heuristics in parallel, and hence a path found for a channel would have di erent segments (a segment is a path between two successive IDs) and each of these segments would very well be selected by di erent heuristics. The e ectiveness of the proposed approach has been studied through simulation for well known network topologies for a wide range of QoS and tra c parameters. The simulation results reveal that the average call acceptance rate o ered by the proposed route selection approach is better than that of both the ooding and preferred neighbour approaches, and the average call setup time and routing distance o ered by it are very close to that of the ooding approach.

Service.

Keywords: Real-time Networks, Channel Establishment, Distributed Routing, Heuristics, Quality of

This work was done when the authors were at the Dept. of Computer Science and Engineering, Indian Institute of Technology, Madras, India.

1

1 Introduction
Packet switched data networks are increasingly being utilised for carrying multimedia tra c such as video and audio which often require stringent quality of service (QoS) requirements in terms of end-to-end delay, delay jitter, and loss. For a network to provide performance guarantees for such multimedia applications, real-time channels 7] are to be established with speci ed tra c characteristics and QoS requirements. The tra c characteristics include parameters such as maximum message rate, maximum message size, and maximum burst size. There are two phases involved in handling a real-time channel: channel establishment 9] and run-time scheduling 9, 19] of packets. The channel establishment phase involves the selection of a quali ed route for the channel satisfying tra c characteristics and QoS requirements of the call request, without compromising the guarantees of the existing channels. Although several channel establishment schemes have been proposed, very few of them have explicitly addressed the issue of route selection despite its importance in the channel establishment phase. There are two basic approaches to the route selection problem: centralised and distributed. In a centralised route selection approach, the existence of a global network manager is assumed which maintains information about all the established real-time channels and the network topology, and can thus select an appropriate route for each real-time channel request. In such a centralised approach, every real-time channel request has to be approved by the network manager. Although this is better in terms of selecting a quali ed route by employing an e cient algorithm for network management, it su ers both in terms of performance and reliability due to the use of a centralised network manager. In contrast with the centralised approach, the distributed route selection approach o ers better performance and is more reliable. Since the number of routes between a source and destination is very high, choosing a quali ed route is not an easy task. The objective of any routing algorithm is to nd a quali ed path with minimal operational overheads. The distributed routing problem, in the context of real-time channel establishment, has been studied by some researchers 3, 8, 10, 15, 17] in the recent past. The primary objective of these algorithms has been to improve the call acceptance rate without considering the response time taken to set up a connection and the routing distance (hops) taken by the connection. In this paper, we propose a new distributed routing approach whose objective is not only improving the call acceptance rate, but also minimising the call setup time and the routing distance. There are two broad schemes for real-time channel establishment, namely, single pass and two pass schemes 14]. In a two pass scheme 7], there are two phases, namely, reservation phase (forward pass) and relaxation phase (reverse pass). The forward pass proceeds from source to destination of the channel request (call), during which at each intermediate node a call admission test (call admission test depends on the scheduling algorithm used) is performed to check whether the resources such as bandwidth, bu ers, and delay guarantee required by the call can be satis ed, and these resources are reserved if the call admission test is successful. Once the forward pass is successful, the destination initiates the reverse pass. During the reverse pass, the resources which were allocated excessively during the forward pass are relaxed so that these excess resources can be allocated to some other calls. If the call is rejected, the resources that were reserved along the path found so far are released. Since the two pass scheme does 2

not reserve resources more than what is required by a call, the average call acceptance rate o ered by a routing algorithm using the two pass scheme is higher than its single pass counterpart. On the other hand, single pass scheme o ers low average call setup time. We believe that call acceptance rate is more important than call setup time. Therefore, in this paper, we consider the two pass scheme for real-time channel establishment. In a two pass scheme, it is necessary to reserve resources during forward pass so that other calls cannot reserve the same resources during their forward pass. This avoids the possibility of two or more calls attempting to con rm the same resource during their reverse pass. The rest of the paper is organised as follows. In Section 2, we discuss the existing routing approaches, their weaknesses, and motivations for our work. The proposed distributed routing approach is presented in Section 3. In Section 4, the performance of the proposed routing approach is compared with that of the existing routing algorithms through simulation studies. Finally, in Section 5, some concluding remarks are made.

2 Background and Motivation
The network, in which the routing problem is addressed here, is an arbitrary point-to-point topology. In a real-time network, at any point in time, many channel establishment requests are active, whose objective is to nd a quali ed routing path from their respective sources to destinations. Path selection within routing is typically formulated as a shortest path optimization problem. The objective function may be number of hops, cost, delay, or some other metric that corresponds to a numeric sum of the individual link parameters along a selected path 10]. E cient algorithms (Dijkstra and Bellman-Ford) exist for computing shortest paths in communication networks. However, within the context of satisfying diverse QoS requirements, the algorithms become more complex as constraints are introduced in the optimization problem. These constraints typically fall into two categories: link constraints and path constraints 3, 10]. A link constraint is a restriction on the use of links on a path based on the available capacity (such as bandwidth) on the links which must be greater than or equal to that required by the call, whereas path constraint is a bound on the combined value of a QoS parameter along a selected path (such as end-toend delay o ered along the path must not exceed what the call can tolerate). Path constraints make a routing problem intractable. A shortest path problem even with a single path constraint is known to be NP-complete 6]. Therefore, heuristics have to be employed to solve the QoS routing problem.

2.1 Performance Metrics
In a traditional computer network, most of the routing algorithms tend to optimise the message delay or the route distance for a single connection 8, 15]. To optimise the global performance of a real-time network, the following metrics are to be optimised. For an accepted call request "R" let us de ne the functions: accepted(R) = 1. setup(R) = number of nodes visited by the call setup packet. 3

dist(R) = length of the path (in terms of hop count) chosen for R. For a call request R that is rejected, all the functions return a value of 0. Let "N" be the total number of call-requests generated.

Average Call Acceptance Rate: the probability of accepting a real-time channel establishment
(call) request, de ned as the ratio of number of calls accepted to the number of calls arrived in the system. PN accepted(R) ACAR = i=1 N

Average Call Setup Time: the average time required to setup a real-time channel measured in
terms of number of nodes visited by the call setup packet.

PN=1 setup(R) ACST = PN i i=1 accepted(R)
ARD = PN
i=1 accepted(R)

Average Routing Distance: the average hop count of the established channels.

PN=1 dist(R) i

The rst metric is very important as it is a measure of call throughput. The second metric is crucial in the context of real-time and interactive multimedia applications which require fast channel setup. The third metric is essential because a short route is less costly. In the context of real-time communication, traditional metrics such as average message delay and message throughput used in datagram networks are meaningless since they do not necessarily indicate anything about the timeliness of messages.

2.2 Earlier Work on Route Selection
Several heuristic routing algorithms have been proposed for the real-time channel establishment problem 8, 10, 15, 17]. The existing distributed routing algorithms fall into two major categories, namely, ooding based and preferred neighbour based. In a ooding based approach, a packet is forwarded to all (or some) of its neighbours, except the node from which the packet has come, to nd a quali ed route. Whereas in a preferred neighbour approach, a packet is forwarded to a preferred neighbour which is chosen based on a heuristic such as shortest path rst (SPF) or lightly loaded link rst (LLF). The ooding based approach is superior in terms of ACST and ARD at the cost of ACAR. The lower ACST and ARD of the ooding approach is due to its non-backtracking nature since all the paths are probed simultaneously, whereas its poor call acceptance is due to excessive reservation of resources, such as bandwidth and bu ers, along many paths from source to destination during the forward pass of a call setup. In comparison, the preferred neighbour approach o ers higher ACAR at the cost of ACST and ARD. Its higher ACAR is due to reservation of resources along only one path as opposed to multiple paths in the ooding based approach, and its higher (poor) ACST and ARD are due to its backtracking 4

nature (backtracking happens when there is no quali ed path from the current node to the destination through all or some of the preferred neighbours of the current node). A distributed route selection scheme based on ooding approach has been proposed in 15] for route selection during real-time channel establishment. In this approach, the number of messages used for establishing a call is at most 2K , where K is the number of links in the network. This is very expensive and results in tentatively reserving resources in many nodes and links, thereby reducing the ACAR. Sending multiple copies of a message to the destination in order to meet the deadline (end-to-end delay) of the message has been proposed in 11]. The intend of this approach is to send a message as quickly as possible, and it is not meant for real-time channel establishment. If this approach is extended to real-time channel establishment, it will be a variation of ooding, which is not e cient due to its poor ACAR. A distributed routing algorithm, called selective probing, was proposed in 4] wherein probes (call establishment packets) are ooded selectively along those paths which satisfy the QoS and optimization requirements. An improvement to selective probing was proposed in 5]. These algorithms do not aim at maximizing all the three performance metrics. There are two types of preferred neighbour heuristics: (i) local/static knowledge based and (ii) dynamic non-local knowledge based. Algorithms such as SPF and LLF are examples for the rst type. The overhead associated with these algorithms is almost the same since these heuristics either use local link information or relatively static global information. For example, the preferred neighbour table of LLF is based on local link information, whereas that of SPF is based on relatively static topology of the network. Heuristics such as least-delay rst 13, 16] is an example for the second type. Here, the preferred neighbour table for a destination is updated based on the delays o ered by other nodes along the paths from the current node to the given destination. Since the delay is dynamically varying, it has to be propagated from one node to another, thus resulting in more overheads compared to the local/static knowledge based heuristics. The propagation of delay information can be achieved by executing distributed routing algorithms such as Bellman-Ford. A distributed route selection algorithm based on preferred neighbour approach has been proposed in 8]. This paper studied the existing routing heuristics such as SPF and LLF, and found that the SPF performs better in terms of ACAR and ARD under uniform tra c and is poor when the call requests are focused on some hot nodes or links. On the other hand, the LLF tries to balance the load on each link by selecting the preferred neighbour nodes in a round-robin fashion. For unbalanced tra c, it indeed increases the ACAR. But under the uniform tra c, it tries to balance the load on each preferred node. A slightly unbalanced load between two preferred nodes will probably cause the LLF to select a route with a longer distance. Even under light load, the LLF scheme still changes the route dynamically which is quite unnecessary. To overcome the problems associated with the SPF and LLF, a routing algorithm, called two level shortest path (TSPF), has been proposed in 8]. In this algorithm, the links of a node are grouped into two groups, namely, heavy group and light group based on a threshold value of load. The SPF heuristic is applied rst within the light group and then within the heavy group. These algorithms 5

(SPF, LLF, and TSPF) are poor in terms of ACST since they encounter excessive backtracks under heavy loads. Two adaptive algorithms, namely, hot spot avoidance (HSA) routing and virtual path (VP) routing have been proposed recently in 2] for massively parallel systems. The HSA algorithm is more suitable for dynamic environment, whereas the VP algorithm is more suitable for static environment. Most importantly, these algorithms are inadequate for satisfying QoS parameters which is typically the requirement in a real-time network.

2.3 Motivation for Our Work
It is clear from the preceding discussion that there is no single routing algorithm which is suitable for improving all the three performance metrics under di erent load conditions. This has motivated us to come up with a routing approach whose objective is to improve all the metrics (ACAR, ACST, and ARD) simultaneously and cater to di erent kinds of work loads. As an attempt to satisfy this objective, we propose a distributed routing approach which is a generalisation of both ooding and preferred neighbour based approaches by o ering a wide spectrum of solutions ranging from the one provided by the ooding based approach to that provided by the preferred neighbour approach.

3 The New Distributed Route Selection Approach
In this section, we propose a distributed route selection approach, called parallel probing, which combines the bene ts of both ooding and preferred neighbour approaches, and also the bene ts of multiple preferred neighbour heuristics that are employed by it in a uni ed way. In our approach, we search for a quali ed path by simultaneously probing at most k di erent paths using k di erent heuristics one for each path. Since searching in parallel (for a quali ed path) reduces the number of backtracks as compared to the sequential searching, the average call setup time is less in the case of parallel search, but at the cost of reserving more resources. For example, when k paths are simultaneously searched by parallel probing, the resources that are tentatively reserved (which are not available for other calls) is approximately k times that of the single path searched by sequential probing. To alleviate this excessive resource reservation without losing the fast call setup capability, the parallel probing uses a concept called intermediate destinations (IDs), which are subset of nodes along the leastcost path between source and destination of a call. The least-cost metric can be in terms of (minimum) number of hops, or based on (least) load, combination of hop count and load, or some other. When a call request arrives, the source node (ID0 ) rst decides the IDs for the call and then initiates probes for a quali ed path to the rst intermediate destination (ID1 ) by sending probe packets (the ID list is appended to each probe packet) in parallel on k di erent paths by employing k di erent heuristics. The probe packet that rst reaches ID1 is considered to be the winner of that segment which originates at source and ends at ID1 . The subsequent probe packets corresponding to the same call reaching at ID1 are rejected and hence the resources reserved by them are released immediately up to the previous intermediate destination (i.e., the source). Now, the parallel probing starts all over again from ID1 to 6

ID2 in the similar manner. This procedure is repeated until the destination is reached or timeout has

occurred. The reverse pass follows the path used during the forward pass either for relaxing/con rming (when the call is accepted) or for releasing (when the call is rejected) the resources along the path. If the resources reserved at a node is not con rmed within a time interval, then the node automatically releases the resources reserved for that call. When a probe packet is sent from the source, it is assigned a unique identi er obtained by concatenating the node number with the local request counter and the heuristic identi er. The heuristic identi er indicates the heuristic based on which a node nds its neighbour.

3.1 Node and Procedure Types
To formally present the parallel probing approach, we rst de ne some node types, and then de ne some procedures which will be invoked when a probe packet is received. The nodes of the network, with respect to a call, are classi ed into four types:

SOURCE: The source node of the call. DESTINATION: The destination node of the call. ID: An intermediate destination (ID) of the call. The call can have more than one ID, and the

search for paths proceed from one ID to another. The selection of IDs and the sequence in which the IDs are to be searched is decided by the least-cost metric between the SOURCE and the DESTINATION of the call. Let ID0 ; ID1 ; :::; IDn?1 ; IDn be the sequence of IDs of a call. Without loss of generality, ID0 is the SOURCE and IDn is the DESTINATION.

IID: An intended intermediate destination (IID) of the call which is an ID to which route is being
probed currently. IDi is IID i IDi?1 was the previous IID and IDi+1 will be the next IID. Without loss of generality, SOURCE (ID0 ) is the rst IID and DESTINATION (IDn ) will be the last IID. These are nodes between IDs along the least-cost path between SOURCE and DESTINATION.

SN: A simple node (SN). The set of nodes which are not members of any of the above categories.
A segment is the path obtained by the parallel probing algorithm between two consecutive IDs of a call. The procedures executed by nodes during call establishment are classi ed into six categories: Reserve(), Forward(), Release(), and Backtrack() - the procedures executed during the forward pass; Relax() and Reject() - the procedures executed during the reverse pass. The IDs of a call are appended to each probe packet, PROBE PKT, at the source node of the call. Each probe packet also has a path eld that captures the sequence of nodes that constitutes the current path.

Reserve(N , IDj , PROBE PKT): Route the probe packet (PROBE PKT) from node N to its

preferred neighbour for reaching node IDj based on a heuristic. This involves performing call admission test on the best preferred link of node N - checking for the availability of bandwidth required by the call; if the admission test is successful, the resources are reserved. If the admission test fails on the best preferred link, the call admission test is performed on the next best preferred 7

link. This is repeated for a xed number of times. If the call admission is successful, it returns "success", it returns "failure", otherwise. Reserve(N , IDj , PROBE PKT)

begin Repeat If (call admission is successful)

end. Forward(IDi , IID, PROBE PKT): This procedure initiates parallel probes on k paths using k

Reserve bandwidth on the preferred link. Send PROBE PKT to the preferred neighbour. return(success). Else Select the next preferred link. Until (maximum neighbours have been tried). return(failure).

di erent heuristics. For each probe, it invokes Reserve() procedure. If all probes fail the admission test, then a Release() procedure is invoked to release the resources reserved between IDi to IDi?1 . Forward(IDi, IID, PROBE PKT)

begin

end. Release(IDi , IDi?1, PROBE PKT): Release the resources along the path between intermediate
destination IDi and intermediate destination IDi?1 using the path stored in the PROBE PKT.

Let H1 , H2 , ..., Hk be the k heuristics. For p = 1 to k do Select the best neighbour based on heuristic Hp . status p] = Reserve(IDi ,IID,PROBE PKT). If (none of status p] is successful, 1 p k) then If (IDi is SOURCE) then call is rejected. Else Release(IDi , IDi?1, PROBE PKT).

Backtrack(N , PROBE PKT): Backtrack from node N to its predecessor (say pred(N )) in the

routing path, which is stored in PROBE PKT. This implicitly releases the resources reserved for the call from pred(N ) to N . The node N could be of type SN or ID. If pred(N ) is SOURCE, the PROBE PKT is dropped.

Relax(DESTINATION, SOURCE, PROBE PKT): This procedure is invoked when the forward

pass is successful. This Relaxes the excess resources such as bandwidth, bu ers, and delay guaranteed (delay is an additive metric) along the path, where the path is stored in PROBE PKT, from DESTINATION to the SOURCE of the call. If cycles are present in the path, they are also removed by releasing the resources reserved in the nodes which form the cycles. 8

Reject(IDi , SOURCE, PROBE PKT): This procedure is invoked when the channel is not possible
to set up. This releases the resources along the path, where the path is stored in PROBE PKT, from node IDi to SOURCE of the call. The IDi is the last ID in the IDs list. This means that no ID can become IID to which probe can be initiated. As a special case, IDi could be the DESTINATION of the call.

3.2 Parallel Probing Algorithm
Figure 1 shows the pseudo code of the parallel probing algorithm which is executed when a call request arrives or a probe packet is received at a node. When a probe packet is received, depending on the node type di erent case statements will be executed. In parallel probing, at a node of type SN and ID, the resources might be reserved for a call more than once due to multiple probe packets passing through that node. If the call is successful and that node is part of the quali ed path, then only one of these reservations will be con rmed by the Relax() procedure and the remaining reservations will be released either by Reject() or Release() procedures.

3.2.1 Properties of Parallel Probing Approach
The parallel probing approach possesses the following properties:

Liveness: The use of IDs ensures the forward movement of probe packets towards the destination.
This eliminates the possibility of a probe packet getting stuck within a group of nodes that form a cycle. heuristics employed (as shown in Figure 2) depending on the load condition on that segment.

Adaptiveness: The di erent segments of a quali ed path could very well be selected by di erent Generality: The parallel probing approach reduces to the ooding approach if the k heuristics
selected for parallel probing are the same, and it reduces to the preferred neighbour approach when the number of paths searched in parallel is one, i.e., k = 1.

Cycle Free Path: The quali ed routing path produced by the parallel probing is cycle free. Theorem 1: Any quali ed routing path produced by the parallel probing is cycle free. Proof: There are two types of cycles possible: (i) cycles involving SNs, i.e., the originating and termi-

nating nodes of a cycle is an SN, and (ii) cycles involving IDs, i.e., the originating and terminating nodes of a cycle is an ID. The rst type of cycles cannot exist in a path produced by the parallel probing algorithm since the heuristics employed for parallel probes ensure cycle freeness in each segment of the path. Even if the heuristics allow cycles, such cycles can be removed during the reverse pass. The second type of cycles could arise in two ways as discussed below. The rst possibility is when IID is IDk and the current node is IDi such that i < (k ? 1). This means that the probe packet has reached an ID which had become IID in the past, which implies a cycle of the 9

Parallel Probing(SOURCE, DESTINATION, k)
begin 1.

When a new call request arrives do the following. SOURCE: Assemble probe packet (PROBE PKT) with intermediate destinations ( 0 1 n?1 n ), without loss of generality, 0 is SOURCE and n is DESTINATION = 1 ; Forward(SOURCE, , PROBE PKT) /* Initiates parallel probing on paths */
I D ; I D ; :::; I D ; ID ID ID IID ID IID k

2.

switch

When a probe packet arrives do the following. (current node type) /* node type is with respect to the probe packet */
case

IID: Let the current node be i . /* Intended intermediate destination */ If (PROBE PKT is already seen by i ) then Release( i, i?1 , PROBE PKT). /* for loser packet */ Else = i+1 ; Forward( i , , PROBE PKT). /* for winner packet - the rst PROBE PKT */
ID ID ID ID IID ID ID IID

case

ID: Let the current node be i . /* Intermediate destination */ If (IID is ( ? 1)) then /* reached a past ID; cycle is encountered */ k such that Release( i , k?1 , PROBE PKT). /* cycle removal */ Else If (IID is probe packets have returned) then /* IID is not reachable */ i+1 and all the Delete i+1 from IDs list. /* this is to keep track of the actual IDs list */ If ( i is the last element in the IDs list) then Reject( i ,SOURCE, PROBE PKT). /* call is rejected. */ Else = i+2 ; Forward( i , , PROBE PKT). /* parallel probe to new IID */
ID ID i < k ID ID ID k ID ID ID IID ID ID IID

Else

If

status = Reserve( i, IID, PROBE PKT). /* already known IID */ (status is failure) then Backtrack( i , PROBE PKT).
ID ID N N

case

SN: Let the current node be . /* Simple node */ status = Reserve( , IID, PROBE PKT). /* already known IID */ If (status is failure) then Backtrack( , PROBE PKT).
N

case

DESTINATION: If (PROBE PKT is already seen) then /* for loser packet */ Release(DESTINATION, n?1 , PROBE PKT). Else if (call is acceptable) then /* for winner packet and call is acceptable */ Analyse for cycles - mark the nodes in the path that form the cycles. Relax(DESTINATION, SOURCE, PROBE PKT). /* call is accepted - reverse pass */ Else Reject(DESTINATION, SOURCE, PROBE PKT). /* call is rejected - reverse pass */
ID

end.

Figure 1: Parallel probing algorithm

10

form ID0 ; ID1 ; :::; IDi ; :::; IDk?1 , ..., IDi . The portion of path from IDk?1 to IDi has to be removed. The rst step of case ID of the algorithm exactly does this. The second possibility arises when the following sequence of events takes place. (a) The IID is IDk and the current node is IDi such that i > k; This means that a probe packet has reached an ID which will become IID in the future; In such cases, the ID node forwards the probe packet just like an SN node (the last Else part of case ID of the algorithm); and (b) IDi becomes IID. The sequence of events (a) and (b) results in creation of a cycle of the form ID0 ; ID1 ; :::; IDi ; :::; IDk , ..., IDi . The portion of the path from IDi to IDk to IDi has to be removed. Such types of cycles are removed during reverse pass of the algorithm either by the Relax() procedure or by the Reject() procedure.

Theorem 2: Dangling resources are unusable resources that are reserved on some links (and nodes)
such that they are neither part of any successfully established real-time channel nor part of any ongoing channel setup attempt. The parallel probing algorithm does not leave dangling resources in the network.

and also the node types which execute these procedures. Resource reservation is done by procedures Reserve() and Forward(), and either or both of these are executed by nodes of type SOURCE, IID, ID, and SN. Reserve() reserves on a single link and Forward() reserves on k links. Resource release is done by procedures Release() and Backtrack(), and either or both of these are executed by nodes of type IID, ID, SN, and DESTINATION. The forms of unnecessary resource reservations, which would become dangling, and the proper release of such resources are listed below. Since the algorithm ensures the release of resources between IID (say IDi ) and IDi?1 for all the nonrst probe packets reaching IID, by executing Release(), there are no dangling resources between two consecutive IDs. When an ID, say IDi , becomes the last element in the IDs list, the resources reserved from IDi to SOURCE are released by Reject(). As a special case, IDi could be the DESTINATION. The algorithm detects some cycles, involving IDs, in the forward pass and removes the unnecessary resources in the cycles (refer case ID of the algorithm). Similarly, it removes other cycles during reverse pass by Relax(). When Backtrack() takes place at an ID or SN, the resources reserved on a link are released. Thus, all the reserved resources are either used by the successful channels or properly released or part of ongoing channel setup attempts.

Proof: To prove this, we identify the procedures which reserve resources and which release resources

Theorem 3: A call setup initiated at a node is either set up or rejected in a nite time. Proof: A call setup is initiated by executing Forward() at the SOURCE node of the call. Since the
11

algorithm has liveness property, i.e., forward movement of probe packets from one ID to another in the

order of ID0 , ID1 , ..., IDn?1 , IDn , it completes the forward pass in a nite time. It is obvious that the reverse pass also takes nite time by executing either Relax() (for successful setup) or Reject() (for call rejection) procedure.

3.3 Example for Parallel Probing
SN H1 H1 PROC5 ID1 PROC2 H2 PROC3 SN SN PROC4 DESTINATION
(ID2)

SN

SOURCE PROC1
(ID0)

H2

SN - Simple Node ID - Intermediate Destination H1 - Heuristic 1 H2 - Heuristic 2

PROC1 : PROC2 : PROC3 : PROC4 : PROC5 :

Forward(SOURCE,ID1,PROBE_PKT) Forward(ID1,DESTINATION,PROBE_PKT) Release(ID1,SOURCE,PROBE_PKT) Relax(DESTINATION,SOURCE,PROBE_PKT) Release(DESTINATION,ID1,PROBE_PKT)
Invokation order (in time) of procedures Proc1 Proc2 Proc3 Proc4 Proc5

Probe packet which reaches first (successful) Probe packet which reaches second (unsuccessful) Path release Path confirm (reverse pass)

Figure 2: Typical channel establishment by parallel probing Figure 2 depicts how the parallel probing algorithm establishes a channel. Here, k = 2. Note that the established path has two segments, in which the rst segment is selected by heuristic H1 and the second is by heuristic H2. The sequence of events taking place in Figure 2 is given below. 1. To start with IID = ID1 . SOURCE executes procedure Forward(SOURCE,IID,PROBE PKT). This means sending two probe packets, one using heuristic H1 (on upper link) and the other using heuristic H2 (on lower link). 2. Probe packet corresponding to heuristic H1 reaches ID1 (which is the current IID) rst. (This is the winner packet.) 3. Now, IID = DESTINATION. ID1 executes procedure Forward(ID1 ,DESTINATION,PROBE PKT). This means sending two probe packets, one using H1 (on upper link) and the other using H2 (on lower link). 4. Heuristic H2's probe packet, which was sent from SOURCE, reaches ID1 . (This is a loser packet.) 5. ID1 executes procedure Release(ID1 ,SOURCE,PROBE PKT). This is to immediately release the resources in the segment between SOURCE and ID1 obtained using heuristic H2. 12

6. DESTINATION (which is the current IID) receives heuristic H2's probe packet which was sent from ID1 . (This is the winner packet). 7. DESTINATION accepts the call and executes procedure Relax(DESTINATION,SOURCE,PROBE PKT). This is to con rm the channel from the DESTINATION to the SOURCE. 8. DESTINATION receives heuristic H1's probe packet which was sent from ID1 . (This is a loser packet.) 9. DESTINATION executes procedure Release(DESTINATION,ID1 ,PROBE PKT). This is to release the resources in the segment between ID1 and DESTINATION obtained using heuristic H1.

4 Simulation Studies
We have conducted extensive simulation studies to evaluate the performance of our parallel probing approach in terms of average call acceptance rate, average call setup time, and average route distance for a wide range of tra c and QoS parameters for di erent network topologies. Before presenting the results, we describe the simulation model and the parameters used in the simulation.

4.1 Simulation Model
To study the e ectiveness of the proposed parallel probing approach, in terms of all the three metrics, we have compared its performance with that of the ooding algorithm 15] and the two level shortest path (TSPF) algorithm 8]. Since the objective of any routing algorithm is to nd a quali ed path with minimal operational overheads, we chose local/static knowledge based heuristics for parallel probes. The probe packets are assumed to use control channels for path search. The number (k) of parallel paths searched simultaneously by the parallel probing approach is taken as 2 and the heuristics employed are SPF and LLF. The IDs for the parallel probing approach are chosen based on the shortest path between the source and destination of the call. For example, let the length of the shortest path be n and the number of IDs be m excluding source and destination. Then, the IDi of a call is the (i mn )-th node in the shortest path. +1 In our simulation, ooding algorithm sends a message to all its neighbours except the node from which the packet has come. The call admission test used, at each node, is the admission test of Hierarchical Round Robin (HRR) 18, 19] scheduling algorithm. HRR is a rate based scheduling discipline which has a simple admission test: to admit a new channel on a link, it checks whether the sum of the utilisation (utilisation of a channel is the ratio of its maximum message length to the period of its frame size) of all the channels passing through the link (including the new channel) is less than or equal to one. Two di erent network topologies have been considered for evaluating the performance of the di erent route selection algorithms. For performance study in a wide area network, the ARPA network shown in Figure 3 (21 nodes, 26 links) and the USA network shown in Figure 4 (26 nodes, 39 links) have been taken as representative topologies 8]. In our simulation, the delay and bandwidth of the links of the 13

1 3 0 4 2

5

8

11

12 13

16

18

6

9

20 14

7

10

15

17

19

Figure 3: The ARPA network (21 nodes, 26 links)
1 0 3 2 4 12 15 16 13 14

11 5 6 7 8 9 10 21 23 25 17 19

18 20 22 24

Figure 4: The USA network (26 nodes, 39 links) networks are taken to be 1 and 100, respectively. Each point in the simulation curves is an average of 5 simulation runs. For each simulation run, 5000 call requests are generated. The simulation parameters are given in Figure 5. The call requests are generated according to the following two distributions: 1. Uniform distribution: the source-destination pair of a call is uniformly chosen from the node set. 2. Hot communication pair (HCP) distribution: p% of all call requests are set to a particular hot source and destination pair, the other calls follow uniform distribution. The other parameters of a call are generated as follows: The duration, bandwidth, and end-to-end delay requirements of a call are uniformly distributed between their respective minimum and maximum values. The inter-arrival time of call establishment requests follow exponential distribution with mean 1= for each node. For studies in Figures 6-10 and Figure 12, there is no hot communication pair, and for studies in Figures 6-11, the number of intermediate destinations used by the parallel probing approach is taken as 14

parameter explanation
del-min del-max bw-min bw-max dur-min dur-max num-ID hp-pair stime

value taken when varied xed call arrival rate 0.6,0.7,...,1.0 0.5,0.65] minimum end-to-end delay of a call 10 maximum end-to-end delay of a call 10,20,...,50 20 minimum bandwidth of a call 2 maximum bandwidth of a call 2,4,...,12 4 minimum duration of a call 200 maximum duration of a call 200,400,...,1200 400 number of intermediate destinations 0,1,2,...,5 2 excluding source and destination % of hot-pair communication 10,20,...,60 0 maximum call setup time 3,4,...,7 Figure 5: Simulation parameters

2. To recall, the main objective of this paper is to propose a route selection approach which attempts to improve all the three performance metrics simultaneously, since all of them are important in a real-time network, and also to adapt to di erent load conditions. To this e ect, the experiments are so designed to study the e ect of di erent QoS and tra c parameters for both uniform and hot-pair (non-uniform) communication patterns. Here, we present the results for the ARPA network (Figures 6-12) only. The results for the USA network exhibit similar behaviour that of the ARPA network.

4.2 Simulation Results
From Figures 6(a)-11(a), it can be observed that the parallel probing approach o ers higher ACAR than the other two algorithms for all the parameters varied. This is due to less call setup overhead (in terms of resource reservation while path probing), due to the use of intermediate destinations, and high adaptiveness of the parallel probing in terms of using multiple heuristics for path probing, i.e., it allows a routing path to have di erent segments selected by di erent heuristics. Whereas, the ooding su ers due to tentative reservation of resources on multiple paths simultaneously which increases blocking of new calls. On the other hand, the TSPF su ers due to its less adaptiveness, and is still better than the ooding. From Figures 6(b,c)-11(b,c), it can be seen that the ACST and ARD o ered by the ooding are smaller than that of the other two algorithms. This is due to simultaneous probing for paths on all links of a node. Also, note that the ACST and ARD for the ooding are the same because it does not encounter backtrack. The ACST and ARD of the parallel probing approach are very close to that of ooding and better than that of TSPF due to its controlled backtracking nature. 15

4.2.1 E ect of Call Tra c Characteristics
The e ect of tra c characteristics of a call, namely, call arrival rate, call setup time, and call duration are studied in Figures 6,7, and 8, respectively. From Figure 6(a), the ACAR decreases for increasing values of call arrival rate. This is because of more calls arrive at higher loads, which in turn consume more resources, thus resulting in more calls getting dropped. That is, at higher loads, more and more calls attempt to use the xed available resources. Also, it can be observed that the ACST and ARD increase with load. This is because at higher loads, nding a quali ed path involves more search in the network as most of the resources are already in use by the existing channels. In Figure 7, the e ect of varying call setup time is studied. The increase in call setup time introduces a trade-o between improvement in ACAR due to more backtracks and deterioration in ACAR due to excess resource reservation for a longer time. From Figure 7(a), as the call setup time increases, the ACAR improvement phenomenon is e ective for all the three algorithms (not very signi cant for ooding). This is because the route selection algorithms are allowed to probe for channels for a longer time (more backtracks), which increases the chances of establishing a channel. Note that, the ACAR curve corresponding to ooding is almost at since ooding does not backtrack. Also, note that, as the call setup time increases, the ARD also increases for all the three algorithms. This is also because the algorithms are allowed to probe for paths for longer times resulting in longer paths. The ACAR deterioration phenomenon will be prominent when the ratio of call duration to call setup time is very less. In our experiments, we have chosen this ratio to be approximately 100 ( 300/3.5) which is a large value and hence the excess resource reservation e ect is outperformed by gain obtained by more backtracks. From Figure 8(a), it can be observed that the ACAR decreases with increasing call duration. This is because for higher values of call duration, the resources are blocked (not available for the new calls) by the currently active calls for longer duration. From Figures 8(b) and (c), the increasing call duration increases the call setup time for parallel probing and TSPF because of blocking of resources. The setup time of ooding does not increase much because of its ooding (and hence no backtracking) nature.

4.2.2 E ect of Call QoS Requirements
The e ect of QoS requirements of a call, namely, the bandwidth and end-to-end delay requirements are studied in Figures 9 and 10, respectively. From Figure 9(a), the ACAR o ered by all the three algorithms starts decreasing for increasing values of bandwidth requirement. The reason is due to reserving of more bandwidth for the currently active calls, when call bandwidth is more, which results in blocking of new calls. Also for the same reason, the ACST and ARD increase with increasing bandwidth requirement. The e ect of varying end-to-end delay constraint imposed by a call is studied in Figure 10. As the end-to-end delay increases, the ACAR also increases. This is because the chances of meeting the delay 16

70
Average call acceptance rate

7.5 TSPF Parallel probing Flooding 7
Average call setup time

60 50 40 30 20 10

6 5.5 5 4.5 4 3.5 3 2.5 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 Call arrival rate

Average route distance

6.5

TSPF Parallel probing Flooding

7

TSPF Parallel probing Flooding

6

5

4

3 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 Call arrival rate

0 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 Call arrival rate

Figure 6: E ect of call arrival rate on (a) ACAR (b) ACST (c) ARD
70
Average call acceptance rate

5.5 TSPF Parallel probing Flooding 5
Average call setup time

5.5 TSPF Parallel probing Flooding 5
Average route distance

60 50 40 30 20 10 0 3 3.5

4.5 4 3.5 3 2.5 2 1.5 1

4.5 4 3.5 3 2.5 2 1.5 1

TSPF Parallel probing Flooding

4 4.5 5 5.5 6 Maximum call setup time

6.5

7

3

3.5

4 4.5 5 5.5 6 Maximum call setup time

6.5

7

3

3.5

4 4.5 5 5.5 6 Maximum call setup time

6.5

7

Figure 7: E ect of call setup time on (a) ACAR (b) ACST (c) ARD
100
Average call acceptance rate

10 TSPF Parallel probing Flooding 9
Average call setup time

10 TSPF Parallel probing Flooding 9
Average route distance

90 80 70 60 50 40 30 20 10 0 200 400

TSPF Parallel probing Flooding

8 7 6 5 4 3

8 7 6 5 4 3

600 800 1000 Maximum call duration

1200

2 200

400

600 800 1000 Maximum call duration

1200

2 200

400

600 800 1000 Maximum call duration

1200

Figure 8: E ect of call duration on (a) ACAR (b) ACST (c) ARD 17

100
Average call acceptance rate

6.5 TSPF Parallel probing Flooding 6
Average call setup time

6.5 TSPF Parallel probing Flooding 6
Average route distance

90 80 70 60 50 40 30 20 10 0 2

TSPF Parallel probing Flooding

5.5 5 4.5 4 3.5 3

5.5 5 4.5 4 3.5 3

4 6 8 10 12 Maximum bandwidth requirement by a call

2

4 6 8 10 12 Maximum bandwidth requirement by a call

2

4 6 8 10 12 Maximum bandwidth requirement by a call

Figure 9: E ect of maximum bandwidth requirement by a call on (a) ACAR (b) ACST (c) ARD
90
Average call acceptance rate

7 TSPF Parallel probing Flooding 6.5
Average call setup time

7 TSPF Parallel probing Flooding 6.5
Average route distance

80 70 60 50 40 30 20 10 0 10

TSPF Parallel probing Flooding

6 5.5 5 4.5 4 3.5 3

6 5.5 5 4.5 4 3.5 3

15 20 25 30 35 40 45 Maximum end-to-end delay of a call

50

10

15 20 25 30 35 40 45 Maximum end-to-end delay of a call

50

10

15 20 25 30 35 40 45 Maximum end-to-end delay of a call

50

Figure 10: E ect of maximum end-to-end delay of a call on (a) ACAR (b) ACST (c) ARD constraint of a call is more when the end-to-end delay is large, i.e., the nodes along the path of a call can assign a higher node deadline (delay) for the messages of the call. The ACST metric decreases with increasing end-to-end delay for the same reason.

4.2.3 E ect of Hot Pair Communication
For this study, nodes 3 and 14 of the ARPA network (Figure 3) are chosen as Hot Communication Pair (HCP). From Figure 11, the ACAR decreases with increasing HCP, and the ACST and ARD increase with increasing hot communication percentage. This is due to the fact that the increase in hot communication saturates the paths between the HCP nodes after reserving resources for some number of calls, and hence rejects the subsequent calls until some of the resources are released due to call tear down. This is applicable to all the three algorithms. In conclusion, the parallel probing approach o ers higher call acceptance rate than that of the TSPF and ooding, and its setup time and routing distance are more close to that of the ooding. The parallel probing is better than ooding and TSPF in improving all the three metrics simultaneously for di erent 18

60
Average call acceptance rate

7 TSPF Parallel probing Flooding 6.5
Average call setup time

7 TSPF Parallel probing Flooding 6.5
Average route distance

50 40 30 20 10

6 5.5 5 4.5 4 3.5 3

6 5.5 5 4.5 4 3.5 3 2.5

TSPF Parallel probing Flooding

0 10 20 30 40 50 Hot pair communication percentage 60

2.5 10 20 30 40 50 Hot pair communication percentage 60

10

20 30 40 50 Hot pair communication percentage

60

Figure 11: E ect of hot pair communication on (a) ACAR (b) ACST (c) ARD simulated load conditions. This conclusion is applicable for both the APRA and USA network topologies.

4.2.4 E ect of Number of Intermediate Destinations
The e ect of number of Intermediate Destinations (IDs) used in the parallel probing approach is studied in Figure 12. For this, 0%, 20%, and 40% HCP is considered, nodes 3 and 14 of the ARPA network are chosen as HCP. Note that the concept of routing to IDs has been introduced with the purpose of controlling the excessive resource reservation on parallel paths. The choice of number of IDs is very crucial in deciding the overall performance of the network. When the number of IDs is very low, excessive resource reservations along the path occur which in turn reduces the ACAR. As a special case, when num-ID is 0, the parallel probing reduces to k-path ooding, which reserves resources on k paths from source up to the destination (in the worst case) during forward pass. On the other hand, when the number of IDs is very large, the routing path is dictated by the least-cost metric which is used to identify the IDs. As a special case, when num-ID is equal to the length of the least-cost path, then the parallel probing reduces to the leastcost path heuristic. The proper choice of num-ID depends on the topology of the network. When the topology is dense, higher value of num-ID is preferable in order to reduce the excessive reservation by probe packets which arises due to the existence of more disjoint paths between any source-destination pair. Any reasonable value of num-ID should be much lesser than the diameter of the network. From Figure 12 (for ARPA), the overall peak performance (by simultaneously considering all the three metrics) occurs when the number of IDs (num-ID) is 1. The peak performance for the USA network topology was found to occur when num-ID is 2 with nodes 3 and 19 are chosen to be HCP. Similarly, the choice of k is also crucial in deciding the overall performance of the network. When the topology is sparse, lower value of k is preferable because higher value of k will make the probe packets to visit many common nodes and hence reserving resources multiple times in such nodes, thereby reducing ACAR. In parallel probing, apart from performance point of view, extra processing overhead comes in 19

70
Average call acceptance rate

5 0% hot pair 20% hot pair 40% hot pair 4.5
Average call setup time

5 0% hot pair 20% hot pair 40% hot pair 4.5
Average route distance

60 50 40 30 20 10 0 0

0% hot pair 20% hot pair 40% hot pair

4 3.5 3 2.5 2 1.5 1

4 3.5 3 2.5 2 1.5 1

1 2 3 4 Number of intermediate destinations

5

0

1 2 3 4 Number of intermediate destinations

5

0

1 2 3 4 Number of intermediate destinations

5

Figure 12: E ect of number of intermediate destinations on (a) ACAR (b) ACST (c) ARD the form of sending k probe packets at every ID as opposed to one probe packet in conventional preferred neighbour approaches such as SPF, LLF, and TSPF. This extra overhead would be very less as the number of IDs is very less compared to the diameter of the network.

5 Conclusions
In this paper, we have proposed a new distributed routing approach, called parallel probing, for real-time channel establishment in a point-to-point network, which is a generalisation of preferred neighbour based and ooding based routing approaches. The parallel probing attempts to improve all the three performance metrics (call acceptance rate, call setup time, call routing distance) simultaneously by combining the bene ts of preferred neighbour approach (better call acceptance) and ooding based approach (minimum call setup time and routing distance). Further, it has the exibility of accommodating any routing heuristic as one of the heuristics for parallel search in order to adapt to di erent load conditions. We have demonstrated the e ectiveness of the proposed parallel probing approach through simulation for a wide variety of QoS and tra c parameters for di erent load conditions (uniform and hot-pair communication) for two well known network topologies. Our simulation studies reveal the following: The parallel probing always o ers higher call acceptance rate than one of the known preferred neighbour algorithms such as the two level shortest path algorithm. The call setup time and route distance o ered by the parallel probing are close to the ooding, and are always lower (better) than that of the two level shortest path algorithm. The proper values of number of intermediate destinations (IDs) and k are crucial in deciding the performance of the parallel probing and they depend on the connectivity of the network. In conclusion, the parallel probing approach is better in terms of improving multiple performance metrics simultaneously for a wide range of QoS and tra c parameters for diverse communication (uniform and hot-pair) tra c. Currently, we are extending the parallel probing approach to the multicast routing. 20

Acknowledgments
The authors would like to thank R. Sriram and anonymous reviewers for their suggestions to improve the presentation of the paper.

References
1] C.M. Aras, J.F. Kurose, D.S. Reeves, and H. Schulzrinne, \Real-time communication in packetswitched networks," Proc. IEEE, vol. 82, no. 1, pp. 122-139, Jan. 1994. 2] M. Boari, A. Corradi, and C. Stefanelli, \Adaptive routing for dynamic applications in massively parallel architectures," IEEE Parallel and Distributed Technology, pp. 61-74, Spring 1995. 3] S. Chen and K. Nahrstedt, \An overview of quality of service routing for next-generation high-speed networks: Problems and solutions," IEEE Network, pp.64-79, Nov./Dec. 1998. 4] S. Chen and K. Nahrstedt, \Distributed Quality-of-Service routing in high-speed networks based on selective probing, in Proc. IEEE Local Computer Networks (LCN), 1998. 5] S. Chen and K. Nahrstedt, \Distributed QoS routing with imprecise state information," in Proc. IEEE Intl. Conf. on Computer Communications and Networking (ICCCN), pp.614-621, 1998. 6] M.R. Garey and D.S. Johnson, \Computers and intractability: A guide to the theory of NPcompleteness," W.H. Freeman, 1979. 7] D. Ferrari and D.C. Verma, \A scheme for real-time channel establishment in wide-area networks", IEEE JSAC, vol. 8, no. 3, pp. 368-379, Apr. 1990. 8] N. Haung, C. Wu, and Y. Wu, \Some routing problems on broadband ISDN," Computer Networks and ISDN Systems, no. 17, pp. 101-116, 1994. 9] D.D. Kandlur, K.G. Shin, and D. Ferrari, \Real-time communication in multihop networks," IEEE Trans. Parallel and Distributed Systems, vol. 5, no. 10, pp. 1044-1056, Oct. 1994. 10] W.C. Lee, M.G. Hluchyi, and P.A. Humblet, \Routing subject to quality of service constraints in integrated communication networks," IEEE Network, July/August, 1995. 11] P. Ramanathan and K.G. Shin \Delivery of time-critical messages using a multiple copy approach," ACM Trans. Computer Systems, vol.10, no.2, pp.144-166, May 1992. 12] S. Rampal, D.S. Reeves, and D.P. Agrawal, \An evaluation of routing and admission control algorithms for real-time tra c in packet-switched networks," in Proc. High Performance Networking, pp. 77-91, 1994. 13] H.F. Salama, D.S. Reeves, and Y. Viniotis, "A distributed algorithm for delay-constrained unicast routing", in Proc. IEEE Infocom, 1997. 21

14] S. Shenker and L. Breslau, \Two issues in reservation establishment," in Proc. ACM SIGCOMM, pp. 14-26, 1995. 15] K.G. Shin and C. Chou, \A distributed route-selection scheme for establishing real-time channels," in Proc. High Performance Networking, pp. 319-330, 1995. 16] R. Sriram, G. Manimaran, and C. Siva Ram Murthy, \Preferred link based delay-constrained least cost routing in wide area networks," Computer Communications, vol.21, no.18, pp.1655-1669, Nov. 1998. 17] Z. Wang and J. Crowcroft, \Quality-of-service routing for supporting multimedia applications," IEEE JSAC, vol. 14, no. 7, Sept. 1996. 18] H. Zhang and S. Keshav, \Comparison of rate-based service disciplines," in Proc. ACM SIGCOMM, pp.113-121, Sept. 1991. 19] H. Zhang, \Service disciplines for guaranteed performance service in packet-switching networks," Proc. IEEE, vol. 83, no. 10, pp. 1374-1396, Oct. 1995.

22

from Bharathidasan University, Thiruchirapalli, in 1989, M.Tech in Computer Technology from the Indian Institute of Technology, Delhi, in 1993, and Ph.D in Computer Science and Engineering from the Indian Institute of Technology, Madras, in 1998. Since January 1999, he has been with the Department of Electrical and Computer Engineering, Iowa State University, USA, as an Assistant Professor. His research interests include QoS routing and multicasting, networked multimedia systems, and parallel and distributed real-time systems. the Indian Institute of Technology, Madras, India in 1997 and an S.M. degree in Computer Science from the Massachusetts Institute of Technology, Cambridge, MA in 1999. As a research assistant with the Intelligent Networks and Adaptive Transmission group at the Laboratory for Computer Science at the Massachusetts Institute of Technology, he has worked and published in areas related to the design of congestion control schemes and adaptive applications for the Internet. He has also been a teaching assistant for courses in Computer Systems. Mr. Rahul is a recipient of the President of India gold medal from the Indian Institute of Technology, Madras and the AT&T Leadership Award. His e-mail address is rahul@lcs.mit.edu and his URL on the Web is http://inat.lcs.mit.edu/ rahul. from Regional Engineering College, Warangal, in 1982, an M.Tech in computer engineering from the Indian Institute of Technology (IIT) Kharagpur, in 1984, and a Ph.D in computer science from the Indian Institute of Science (IISc), Bangalore in 1988. From March to September 1988 he worked as a scienti c o cer in the Supercomputer Education and Research Centre at IISc. He subsequently joined IIT Madras as a lecturer in computer science and engineering. He became an assistant professor in August 1989 and is currently an associate professor with the same department. He has held visiting positions at the German National Research Centre for Information Technology (GMD), Sankt Augustin, Germany, the University of Washington, Seattle, and the University of Stuttgart, Germany. He is a recipient of the Seshagiri Kaikini Medal for the best PhD thesis and also of the Indian National Science Academy Medal for Young Scientists. He is a member of the IEEE.

G. Manimaran (M '99, ACM M '99) obtained a B.E degree in Computer Science and Engineering

Hariharan Shankar Rahul received a B.Tech degree in Computer Science and Engineering from

C. Siva Ram Murthy (M '98) obtained a B.Tech in electronics and communications engineering

23


更多相关文档:

Load Profiling for Efficient Route Selection in Multi-Class ....pdf

Load Profiling for Efficient Route Selection in ...application requesting the establishment of the VC....resources in a distributed system are equally ...

Guaranteed Message Delivery Time on Real-Time Transputer ....pdf

2001 149 Guaranteed Message Delivery Time on Real-Time Distributed Systems ...channel establishment is the selection of routes between the source nodes and...

Chapter 5 A New Admission Control Method for Real-Time ....pdf

Sec. 5.2 The Tenet Approach 106 Channel establishment is a distributed ...A new channel's route may be computed by the source (or destination) if...

Dangerous Goods Route Selection Criteria.pdf

and particularly the research related to route selection criteria and methods....routing of dangerous goods in large cities; Establishment of new criteria ...

A New Approach to Local Route Recovery for Multihop TCP in Ad....pdf

A New Approach to Local Route Recovery for Multi...cient use of the precious channel bandwidth and ... the route re-establishment process is more ...

Dynamic multi-path communication for video traffic.pdf

The latter approach admits and reserves/allocates ...an alternative route, called the shadow channel. ...establishment and the reservation of the alternative...

[]A Distributed Model for Next Generation Router Software_....pdf

Time for route establishment and time for ...a new distributed architecture for router of ...This channel shares the bandwidth on SF with ...

Route-Lifetime Assessment Based Routing (RABR) Protocol for ....pdf

Hence, an optimal route selection should be based...The route establishment is done by broadcasting ...connected channels, does not perform as well as ...

Support Industrial Hard Real-Time Traffic with Switched ....pdf

This centralized approach has the disadvantage that...route/switch from the source to the destination....As new real-time requests (channel establishment)...

Experiments with the tenet real-time protocol suite on the ....pdf

the most suitable route for this channel. ...methods for providing both real-time and best-...time channel establishment in wide-area networks. ...

...Wireless Mobile Applications with Dynamic Establishment_....pdf

Channel establishment In order to allocate channels...Here, based on Dynamic Source Routing [7], we...cations in distributed environments, Computer ...

...probability for distributed lightpath establishment in WDM....pdf

A lightpath establishment protocol is responsible for ?nding a route and a...method is one that reserves multiple free wavelength channels on every hop ...

...BGP IDRP Route Server alternative to a full mesh routing_....pdf

Establishment of BGP/IDRP sessions among all ...The routing approach described in this paper ...selection criteria to the routes received from ...

Extensible Real-Time Data Dissemination on Channel-Based ....pdf

Since it is also desirable for distributed real-time system to support real...The route of a channel will be chosen at the time of its establishment....

D. Modeling Route Change in Soft-State Signaling Protocols ....pdf

We believe this modeling approach will be useful ...er a dynamic number of channels connected to a ...6. New route and state establishment. This MSC ...

...TCP Performance through Pre-detection of Route Failure in ....pdf

goal of fast re-establishment of new route[1]....(RSSI) of the wireless communication channel ... An Approach for Preemptive Route Repair in ...

rfc2918 Route Refresh Capability for BGP-4.txt

rfc2918 Route Refresh Capability for BGP-4_IT/...To accomplish this, a commonly used approach, ...session establishment time via capability ...

...Optimal Active Route Timeout for the AODV Routing Protocol....pdf

While the former approach performs well in low ... to avoid the overhead of route establishment. ...exponentially distributed idle time we describe ...

...Aware Routing Exploiting Transient Route Availability in ....pdf

comparable route establishment and maintenance ...2 a load sharing approach for routing decisions ...(MME-DSR) route selection algorithm to select ...

Dynamic Establishment of Protection Paths in WDM Networks, ....pdf

Dynamic Establishment of Protection Paths in WDM ...nonoverlapping wavelengths (or wavelength channels)....(and possibly the route as well) of any ...

更多相关标签:
网站地图

文档资料共享网 nexoncn.com copyright ©right 2010-2020。
文档资料共享网内容来自网络,如有侵犯请联系客服。email:zhit325@126.com