P. Palnati, M. Gerla and E. Leonardi, "Deadlock-free routing in an optical interconnect for high-speed wormhole routing networks," in Proceedings of 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....The problem is that since there is no single orientation, source destination pairs may need con icting orientations. However, we show how the Train Algorithm can be used in this case such that termination times are still within a reasonable bound. Similar ideas have been used in Up Down Routing [22]. We begin by choosing a node in the above tree, which we refer to as the separator. A good way to choose a separator is discussed following Lemma 5) When the separator is removed, the tree is separated into components where each is a tree; denote by C 1 ; C 2 ; Cm these components with ....

P. Palnati, E. Leonardi, and M. Gerla. Deadlock-free routing in an optical interconnect for high-speed wormhole routing networks. In Proceedings of the 1996 International Conference on Parallel and Distributed Systems, Tokyo, June 1996.

....Blocking lnessage Figure 1: Backpressure flow control. message that is in flight . A critical problem of backpressure flow control is the occurrence of deadlocks. Several deadlock avoidance mechanisms have been proposed, including acyclic routing, timeouts and dynamic detec tion resolution [8, 9]. In Myrinet, a type of acyclic routing called Up Down routing is used [3] The concept of deflection routing is illustrated in Figure 2. When a message arriving at a switch finds its preferred output port busy, it doesn t wait. Rather, it is deflected from its shortest path and immediately ....

P. Palnati, M. Gerla and E. Leonardi, "Deadlock-free routing in an optical interconnect for high-speed wormhole routing networks," in Proceedings of 1996.

.... buffer size must be at least 4 times the one way propagation delay on the upstream link (in bytes) The effectiveness of this backpressure flow control scheme has been reported in [21] One drawback of backpressure is the possibility of creating a cycle of blocked worms which result in deadlock [8, 20]. Upon detection of a deadlock, one or more worms could be dropped to break the cycle. However, in this paper, we focus on reliable networks in which loss of worms inside the network must be avoided. Hence deadlock avoidance must be employed [20] Another technique which can be used to limit ....

....a cycle of blocked worms which result in deadlock [8, 20] Upon detection of a deadlock, one or more worms could be dropped to break the cycle. However, in this paper, we focus on reliable networks in which loss of worms inside the network must be avoided. Hence deadlock avoidance must be employed [20]. Another technique which can be used to limit congestion and prevent packet loss in worm LANs is deflection routing. The scheme works as follows: if the worm arriving at the switch finds its preferred output (as defined by a routing table or by a source routing header) busy, it is deflected to ....

[Article contains additional citation context not shown here]

P. Palnati, M. Gerla, and E. Leonardi. Deadlock-Free Routing in an Optical Interconnect for High-Speed Wormhole Routing Networks. In Proceedings of 1996 International Conference on Parallel and Distributed Systems, June 1996.

....consequence of employing the backpressure flow control mechanism is that bidirectional links are required. Thus, the bidirectional shufflenet is preferable to the traditional shufflenet. Wormhole routing has a potential for deadlocks. The issue of deadlock prevention in SSN is addressed elsewhere [17]. 3.3 WDM optical backbone protocols Optical backbone protocols must extend transparently the Myrinet services end to end while achieving efficient utilization (and reallocation) of expensive backbone resources, efficient scaling, congestion protection and fault tolerance. The WDM optical ....

....at the entry points. g) deadlock prevention. Deadlocks may occur in the transfer of datagrams in the backbone network. These would bring the entire network to a halt, and therefore must be prevented. Several schemes are now being considered, and are discussed and evaluated in more detail in [17]. h) dynamic reallocation of wavelengths to different services. The WDM architecture offers the unique opportunity to reallocate bandwidth resources to different services (in our case, guaranteed bandwith and datagram service) based on user requirements. Furthermore, the multihop virtual ....

[Article contains additional citation context not shown here]

P. Palnati, M. Gerla, and E. Leonardi. Deadlockfree Routing in an Optical Interconnect for HighSpeed Wormhole Routing Networks. Submitted for publication.

....in a (small) buffer associated with the input port called a slack buffer. Using the STOP GO mechanism, the rest of the worm is stored in the slack buffers of upstream nodes. One drawback of backpressure is the possibility of creating a cycle of blocked worms which result in deadlock [DS87, PGL96] Upon detection of a deadlock, one or more worms could be dropped to break the cycle. However, in this paper, we focus on reliable networks in which loss of worms inside the network must be avoided. Hence, deadlock free routing must be employed. We next describe the deadlock free up down routing ....

P. Palnati, M. Gerla, and E. Leonardi. Deadlock-free routing in an optical interconnect for high-speed wormhole routing networks. In Proc. Int'l Conf. on Parallel and Distributed Systems (ICPADS '96), Tokyo, June 1996.

....to deadlocks. A deadlock would bring the entire network to a halt, and therefore must be prevented. The main options are to perform up down routing on a spanning tree and to use virtual channels to allow deadlock free shortest paths. These schemes are discussed and evaluated in more detail in [17]. h) dynamic reallocation of wavelengths to different services. The WDM architecture offers the unique opportunity to reallocate bandwidth resources to different services (in our case, guaranteed bandwidth and datagram service) based on user requirements. Furthermore, the multihop virtual ....

....of deflection. A higher number of hops traveled indicates a higher level of congestion. Thus, backpressure flow control and deflection routing are possible alternative solutions to the problem of congestion in the network. Wormhole routing with backpressure flow control can lead to deadlocks [6, 17]. Hence, we use the virtual channel approach or the Up Down routing approach [17] Deflection routing could result in a livelock, thus we use a deflection buffer (DF) to store some of the worm and probabilistically break livelocks (a procedure that we call delayed deflection) 13] With delayed ....

[Article contains additional citation context not shown here]

P. Palnati, M. Gerla, and E. Leonardi. Deadlockfree Routing in an Optical Interconnect for HighSpeed Wormhole Routing Networks. Technical Report CSD-950064, UCLA, 1995.

.... buffer size must be at least 4 times the one way propagation delay on the upstream link (in bytes) The effectiveness of this backpressure flow control scheme has been reported in [20] One drawback of backpressure is the possibility of creating a cycle of blocked worms which result in deadlock [8, 19]. Upon detection of a deadlock, one or more worms could be dropped to break the cycle. However, in this paper, we focus on reliable networks in which loss of worms inside the network must be avoided, hence deadlock avoidance must be employed; this is in contrast with ATM networks, where cells ....

....control, since worms are blocked when the preferred output ports are not available, circular waits are possible leading to deadlocks. Similar problems occur in packet networks with window based flow control. Deadlocks in backpressure wormhole routing networks have been extensively investigated [8, 17, 19]. Two solutions will be briefly reviewed: Up Down routing on a spanning tree as used in Autonet [21] and the use of Virtual Channels [8, 19] See Fig. 2 for an example of the two alternative approaches discussed next. 2 In Myrinet, source routing is used. Thus, there is only one preferred output ....

[Article contains additional citation context not shown here]

P. Palnati, M. Gerla, and E. Leonardi. Deadlock-Free Routing in an Optical Interconnect for High-Speed Wormhole Routing Networks. In Proceedings of 1996 International Conference on Parallel and Distributed Systems (ICPDS), pages 256--264, June 1996.

....These algorithms are difficult to design, and inevitably add latency to the connection establishment phase. Thus, this approach is considered impractical and is not pursued further. 2 Wormhole routing networks are prone to deadlocks [8] thus, some deadlock avoidance routing scheme must be used [21]. 4 An alternative is to dedicate a pool of links from the network topology to support all QoS connections. Several QoS connections could then be multiplexed on these links. The circuits are accepted based upon some criteria such as peak average bandwidth required. Pacing (input rate control at ....

....to a deadlock. Upon detection of a deadlock, the network could be reset (or some worms dropped to break the cycles) However, this option involves worm loss. Thus, we focus on using deadlock free routing either Up Down routing as in Autonet [23] or virtual channel based shortest path routing [8, 21] (see Fig. 3) In the virtual channel approach (which is of interest for QoS support) each buffer associated with the input port of each switch is split into several disjoint buffers. The link between the upstream node and this input port is then treated as a collection of several virtual ....

[Article contains additional citation context not shown here]

P. Palnati, M. Gerla, and E. Leonardi. Deadlock-Free Routing in an Optical Interconnect for High-Speed Wormhole Routing Networks. In Proceedings of 1996 International Conference on Parallel and Distributed Systems, June 1996.

....node Blocking message Figure 1: Backpressure flow control. message that is in flight . A critical problem of backpressure flow control is the occurrence of deadlocks. Several deadlock avoidance mechanisms have been proposed, including acyclic routing, timeouts and dynamic detection resolution [8, 9]. In Myrinet, a type of acyclic routing called Up Down routing is used [3] 2.2 Deflection routing The concept of deflection routing is illustrated in Figure 2. When a message arriving at a switch finds its preferred output port busy, it doesn t wait. Rather, it is deflected from its shortest ....

P. Palnati, M. Gerla and E. Leonardi, "Deadlock-free routing in an optical interconnect for high-speed wormhole routing networks," in Proceedings of 1996 International Conference on Parallel and Distributed Systems, Tokyo, June 1996.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC