P. Fraigniaud, J.G. Peters, "Structured Communications in Torus Networks ", in Proc. of 28th Hawaii International Conference on System Sciences, Jan. 1995, pp. 584-593.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Umea University, Faculty of Science and Technology, Department of Computing Science, 90187 Umea, Sweden. E mail: jopsi cs.umu.se, WWW: http: www.cs.umu.se jopsi 1 Earlier Work. A substantial amount of research has been carried out on variants of the gossiping problem on meshes and tori [17, 3, 7, 16, 4, 12]. Recently Soch and Tvrdk [22] have analyzed the following variant of the gossiping problem: Packets of size s can be transferred in one step between adjacent PUs (store and forward model) In each step a PU can exchange packets with all its neighbors (full port model) They show that on a ....

Fraigniaud, P., J.G. Peters, `Structured Communication in Torus Networks,' Proc. 28th Hawaii Conference on System Science, pp. 584--593, 1995.

....size. By properly choosing a convenient size and by randomizing the message transmission we can exploit the network performance in a simple, ecient, robust and predictable way. 8 Conclusion Many studies in the literature use a deterministic approach to structure global communication patterns [12] [16] These studies often neglect low level aspects, as deadlock avoidance and ow control, that are very important to achieve good performance. They often are not very robust in practice, e.g. because the processing nodes loose synchronization. We have shown that a simple randomized algorithm ....

Pierre Fraigniaud and Joseph G. Peters. Structured Communication in Torus Networks. In Proceedings of the 28th Hawaii Conference on System Science, 1995.

....based on an innovative interplay between routing algorithm and flow control. It uses a randomized scheduling of the message transmissions, and reaches nearoptimal performance. Most studies in the literature use a complementary deterministic approach to structure global communication patterns [60] [84] These studies often neglect low level aspects, as deadlock avoidance and flow control, that are very important in practice to achieve good performance. They often are not very robust in practice, e.g. because the processing nodes loose synchronization. ffl BSP enthusiasts could be tempted ....

....with four virtual channels. The message size is 64 bytes. CHAPTER 7. COLLECTIVE COMMUNICATION PATTERNS 118 the same hypothesis increases the execution time of less than 1 7.1. 3 Discussion Many studies in the literature use a deterministic approach to structure global communication patterns [60] [84] These studies often neglect low level aspects, as deadlock avoidance and flow control, that are very important in practice to achieve good performance. They often are not very robust in practice, e.g. because the processing nodes loose synchronization. We have shown that a simple ....

Pierre Fraigniaud and Joseph G. Peters. Structured Communication in Torus Networks. In Proceedings of the 28th Hawaii Conference on System Science, 1995.

....the end of correction phase, the corrected values are again needed by all PUs for advancing the computation, and thus needing two gossiping steps in each application of the method. Earlier Work. A substantial amount of research has been performed on (variants of) the gossiping problem (see, e.g. [1, 3, 9, 2, 14]) Some of these papers address the same problem as considered by us: gossiping on meshes with worm hole routing. The algorithms in [9, 2] are very sophisticated and focus on minimizing the number of routing operations. Theoretically, this is an interesting issue, but practically, such an approach ....

Fraignaud, P., J.G. Peters, `Structured Communication in Torus Networks,' Proc. 28th Hawai Conference on System Science, 1995.

....methods [11] In each application of the block method, computations corresponding to the prediction are carried out by different PUs and these values are required by all other PUs. Earlier Work. A substantial amount of research has been carried out on variants of the gossiping problem [10, 3, 6, 9, 4, 7]. Recently Soch and Tvrdk [14] have analyzed the following variant of the gossiping problem: ffl Packets of size s can be transferred in one step between adjacent PUs (store and forward model) 1 ffl In each step a PU can exchange packets with all its neighbors (full port model) The assumption ....

Fraigniaud, P., J.G. Peters, `Structured Communication in Torus Networks,' Proc. 28th Hawai Conference on System Science, pp. 584--593, 1995. 19

....size. By properly choosing a convenient size and by randomizing the message transmission we can exploit the network performance in a simple, e#cient, robust and predictable way. 8 CONCLUSION Many studies in the literature use a deterministic approach to structure global communication patterns [12], 16] These studies often neglect low level aspects, as E#cient Total Exchange in Wormhole Routed Toroidal Cubes 13 deadlock avoidance and flow control, that are very important to achieve good performance. They often are not very robust in practice, e.g. because the processing nodes loose ....

Fraigniaud, P.---Peters, J. G.: Structured Communication in Torus Networks. In Proceedings of the 28th Hawaii Conference on System Science, 1995.

....methods [9] In each application of the block method, computations corresponding to the prediction are carried out by different PUs and these values are needed by all other PUs. Earlier Work. A substantial amount of research has been performed on (variants of) the gossiping problem [1, 2, 7]. In some sense, we turn back to basics. Rather than to design an even more sophisticated algorithm, along the lines of [7] we present a fairly simple algorithm and show that it actually works in practice. An essential point is that we achieve an optimal trade off between start up and routing ....

Fraignaud, P., J.G. Peters, `Structured Communication in Torus Networks,' Proc. 28th Hawai Conference on System Science, 1995.

....by different PUs, and these values are needed by all PUs for the correction phase, requiring a gossiping of the data. Previous Work. A substantial amount of research has been performed on finding efficient algorithms for collective communication operations on wormhole routed systems (see, e.g. [1, 4, 12, 3, 17]) However, most papers either deal with very small packets or with very large packets. Both these extreme cases require algorithms optimizing only one parameter. If the packets are small, then the number of start ups should be minimized. Peters and Syska [12] considered the broadcasting problem ....

Fraigniaud, P., J.G. Peters, `Structured Communication in Torus Networks,' Proc. 28th Hawai Conference on System Science, Volume 2, pp. 584--593, IEEE, 1995.

.... of size L from x to y is generally modeled by fi d x y ffi L b where fi is the start up time that consists of local initialization processes (hard or soft) ffi is the switching time of the routers, d x y is the length of the route 2 from x to y, and b is the bandwidth of the links (see [11]) In general, we have fi AE ffi and fi AE 1 b . Therefore, as for the store and forward mode, assuming T x y (L) constant is a good approximation when L is not too large. The line and cut through models In this paper, we are interested in the communication complexity of broadcasting and ....

....[18, 19] Unfortunately, most of these variations yield NP complete problems when looking for time optimal broadcasting or multicasting protocols. Farley s theorem In the line model, most of the known results concern particular network architectures as trees [4, 26] cycles [23] meshes or tori [5, 11]. A major result which applies to any topology is due to Farley [4] who has shown that, under the line model, for any network G of n nodes, the broadcasting time from any node of G is dlog 2 ne. This result will be referred as Farley s Theorem in the following. Note that this theorem is optimal ....

[Article contains additional citation context not shown here]

P. Fraigniaud and J. Peters. Structured Communication in torus networks. In IEEE, editor, 28th Annual Hawaii International conference on system sciences, pages 584--593, 1995.

....nodes. In this paper, the line model plus hypothesis 4 is called the cut through model (the routing function can be adaptive or not) Most of the known results about broadcasting under these hypotheses deal with particular network architectures as trees [4, 21, 22] cycles [19] meshes or tori [5, 10]. Many results have been also derived when hypothesis 2 is replaced by the all ports hypothesis, that is, when a node can simultaneously communicate with as many nodes as its number of ports (see for instance the 1 As in [4] but as opposed to [16] a sender or receiver node can be an inner node ....

P. Fraigniaud and J. Peters. Structured Communication in torus networks. In IEEE, editor, 28th Annual Hawaii International conference on system sciences, pages 584--593, 1995.

....and lower bounds are parameterized in terms of the number of virtual channels available, so the effects of virtual channels on performance can be determined directly. To study the effects of multiple ports, we compare our algorithms and lower bounds to similar results for single port models in [6]. We study four structured communication patterns in this paper. In the broadcasting pattern, a single processor sends the same message to all other processors. In the scattering pattern, a single processor sends a personalized message to each other processor. Broadcasting and scattering are ....

....surveys [5, 7, 8, 10] Most of the earlier work on these problems used models of store and forward routing, but there are many recent papers that use circuit switched routing. A good recent survey of these results is [11] Our model of computation is the same as in [13] and almost the same as in [6]. The difference between the two models is that the model in this paper and [13] is an all ports model which permits a node to communicate simultaneously through all of its ports whereas the model in [6] is a one port model. Most of our conclusions about the effects of multiple ports are based on ....

[Article contains additional citation context not shown here]

P. Fraigniaud and J.G. Peters. Structured communication in torus networks. Proc. 28th Annual Hawaii International Conference on System Sciences, 1995, 584--593.

No context found.

P. Fraigniaud, J.G. Peters, "Structured Communications in Torus Networks ", in Proc. of 28th Hawaii International Conference on System Sciences, Jan. 1995, pp. 584-593.

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