J. Yantchev and C.R. Jesshope. Adaptive, low la- tency, deadlock-free packet routing for networks of processors. IEE Proc., Pt. E, 136(3):178-186, May  Home/Search   Document Not in Database   Summary   Related Articles   Check

....This matches the known upper bound of three queues per node for deadlock free, minimal packet routing on cycle and torus networks. 1 Introduction routing in parallel and distributed architectures. A wide range of packet routing algorithms with differing properties and costs have been proposed [1, 2, 3, 4, 5, 6, 8, 9, 10, 12, 13, 14, 15, 16, 17, 19, 20]. In this paper we Permission to copy without fee all or part of this material is granted provided that the copies are not made or distributed for direct commercial advantage, the ACM copyright notice and the title of the publication and its date appear, and notice is given that copying is by ....

....and implemented. The primary disadvantage of queue reservation tech niques is that they require that each node contain some minimum number of queues. Although a great deal of research has been devoted to the problem of minimiz ing the storage requirements of queue reservation algorithms [2, 4, 5, 6, 8, 10, 12, 15, 17, 18, 19, 20], very little is known in terms of lower bounds on the storage which is required by such algorithms. Our goal in this paper is to characterize the properties which these algo rithms must have in order to be free of deadlock and to use these properties to prove lower bounds on storage ....

[Article contains additional citation context not shown here]

J. Yantchev and C.R. Jesshope. Adaptive, low la- tency, deadlock-free packet routing for networks of processors. IEE Proc., Pt. E, 136(3):178-186, May

.... routing [15, 45] Most contemporary multiprocessors use deterministic routing [1, 5, 16, 44] The network performance of deterministic techniques has reached a limit, under random traffic, of a stable maximum throughput of 45 to 50 of the limit set by the network bandwidth [39] Adaptive routing [14, 39, 53] improves network performance under high contention by allowing messages to travel along alternate paths if the primary path is congested. Under adaptive routing, messages may arrive out of order under high congestion, since later messages may take paths that are significantly faster. In addition ....

J. Yantchev and C.R. Jesshope. Adaptive, low latency, deadlock-free packet routing for networks of processors. IEE Proceedings Part E, 136(3):178--186, 1989.

....significantly more storage than wormhole routing. Many techniques have been developed to reduce the storage requirements of deadlock free store and forward and virtual cut through routing algorithms. These techniques can be divided into two classes, namely those which require only central queues [10, 12, 16, 18, 22, 25, 28, 29] and those which require that each node have queues (often called buffers) that are associated with each edge that is incident to the node [9, 15, 19] Algorithms in the first class often require less storage than those in the second class. However, the central queues can become sequential ....

....only a constant number of queues per node in tori of arbitrary size and dimension. For example, the hops so far scheme [22] requires that each node have more queues than the diameter of the torus. The routing algorithm described by Jesshope, Miller and Yantchev [16] and by Yantchev and Jesshope [29] is minimalfully adaptive, but it requires eight queues per node in a two dimensional torus and more queues per node in higher dimensional tori. The techniques created by Pifarr e, Gravano, Felperin and Sanz [25] and by Felperin, Laffitte, Buranits and Sanz [9] lead to minimal fully adaptive ....

J. Yantchev and C.R. Jesshope. Adaptive, low latency, deadlock-free packet routing for networks of processors. IEE Proc., Pt. E, 136(3):178-- 186, May 1989.

....but messages hold fewer resources when they block. Uncongested message latency for packet switched networks with virtual cut through is equivalent to that of wormhole networks. Mad postman is an unusual optimization of packet routing that is used to avoid the latency of the routing decision [YJ89] When using mad postman routing, a message is immediately forwarded in the same direction in which it is currently routing. If the prediction is incorrect, the message is killed and forwarded out the appropriate direction. This technique only works in directed acyclic networks and is rarely ....

.... as the double y algorithm [GN92b] is equivalent to algorithms that have appeared elsewhere [CK92, LH91] Jesshope, Miller, and Yantchev observed that if networks are partitioned into 17 independent virtual networks, then it is sufficient to provide deadlock freedom within each virtual network [YJ89] Though presented in the context of packet routing, this idea was used by Linder and Harden to create minimal fully adaptive wormhole routing algorithms for k ary n cubes [LH91] For torus networks, the Linder Harden algorithm requires (n 1)2 n Gamma1 virtual channels per channel, to support ....

J. Yantchev and C.R. Jesshope. Adaptive, low latency, deadlockfree packet routing for networks of processors. IEE Proc., Part E, 136(3):178--186, May 1989.

....and Advanced Applications Group, IBM Argentina, Buenos Aires, Argentina 1 Introduction This paper studies the problem of deadlock free packet routing in parallel and distributed architectures. A wide range of packet routing algorithms with differing properties and costs have been proposed [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 14, 16, 17, 18, 19, 20, 21, 23, 24]. In this paper we will focus on a particularly simple and important class of routing algorithms which we will call buffer reservation algorithms. A buffer reservation algorithm consists of rules that specify to which buffers a packet may move based solely on the buffer currently holding the ....

....studied and implemented. The primary disadvantage of buffer reservation techniques is that they require that each node contain some minimum number of buffers. Although a great deal of research has been devoted to the problem of minimizing the storage requirements of buffer reservation algorithms [4, 7, 8, 10, 12, 14, 16, 19, 21, 22, 23, 24], very little is known in terms of lower bounds on the storage which is required by such algorithms. Our goal in this paper is to characterize the properties which these algorithms must have in order to be free of deadlock and to use these properties to prove lower bounds on storage requirements. ....

[Article contains additional citation context not shown here]

J. Yantchev and C.R. Jesshope. Adaptive, low latency, deadlock-free packet routing for networks of processors. IEE Proc., Pt. E, 136(3):178--

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