A.S. Vaidya, A. Sivasubramaniam and C.R. Das, "Performance benefits of virtual channels and adaptive routing: An application-driven study," International Conference on Supercomputing, pp. 140-147, July, 1997 10  Home/Search   Document Not in Database   Summary   Related Articles   Check

....However, virtual channels and adaptive routing have been shown to increment router delay [4] thus increasing the execution time of latencysensitive parallel applications. For those applications, it has been suggested that routers should implement neither virtual channels nor adaptive routing [23]. Therefore, including virtual channels and adaptive routing in a wormhole router is a design tradeoff. In this paper a new virtual cut through (VCT) router is proposed. We show that router delay can be reduced with respect to a wormhole (WH) router while keeping the same functionality. We have ....

....306.1 Table 5. Base latency (ns) As we have previously mentioned, our main goal is to reduce the latency values in the most frequently used scenarios of network operation because many parallel applications for DSM machines are latency limited, thus benefiting from reductions in message latency [23]. As expected, the deterministic Bubble router, which has the lower through delay, exhibits the lowest network base latency. Even when adaptivity is supported, the adaptive Bubble router exhibits latency values better than those obtained by deterministic wormhole routers. On the contrary, ....

A.S. Vaidya, A. Sivasubramaniam and C.R. Das, "Performance benefits of virtual channels and adaptive routing: An application-driven study," International Conference on Supercomputing, pp. 140-147, July, 1997 10

....n flies and k ary n trees. To the best of our knowledge, this is the first study that deals with the relation between wormhole networks and the communication patterns generated by SPMD applications. In fact most of the existing literature is limited to directory based cache coherency protocols [16] [17] 18] 19] We argue that not much is known about the relations between the programming model and the interconnection network and that commonly used communication benchmarks do not necessarily reflect the behavior of the interconnection network during the execution of a real parallel ....

A. S. Vaidya, A. Sivasubramaniam, and C. R. Das. Performance Benefits of Virtual Channels and Adaptive Routing: An Application-Driven Study. In Proceedings of the 11th ACM International Conference on Supercomputing, Vienna, Austria, July 1997. 28

....only two studies that have evaluated the impact of VCs and or adaptive routing on the performance of CCNUMA systems through execution driven simulation. The first one, by Kumar and Bhuyan [10] concluded that the use of VCs on a torus was a winning choice, whereas the other one, by Vaidya et al. [18], pointed out a lack of scalability in the experiments conducted in the former work, and claimed no advantage in using extra VCs. The authors also concluded that adaptivity was not beneficial. We comment on these works in Section 5. Overall, it is a common belief that VCs and adaptivity have ....

....the bars indicating execution times for the other routing algorithms. The figure also shows that, in 1 way network configurations, neither extra VCs nor adaptivity are able to reduce the execution time much. At most, they reduce it by 5 . This is in agreement with results obtained by Vaidya et al. [18], although it contradicts the results of the same authors in [17] for pipelined designs. When network ASICs and wires are fast enough, it is seldom the case 1 2 4 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 Radix Bristling Factor 1 2 4 0.9 1 1.1 1.2 1.3 1.4 1.5 ....

[Article contains additional citation context not shown here]

A. S. Vaidya, A. Sivasubramaniam, and C. R. Das. "Performance Benefits of Virtual Channels and Adaptive Routing: An Application-Driven Study". Proc. 11th ACM Intl. Conf. on Supercomputing, July 1997.

....network parameters: number of VCs, number of flit buffers per channel and number of links between the compute node and the router. Their results show that four VCs per input port offers the best performance in most cases and that 2 to 4 flit buffers per channel are usually enough. Vaidya et al. [18] analyze the performance benefits of virtual channels and adaptive routing in 2 D meshes. They find that these enhancements reduce the network latency to varying degrees depending on the application and that the improvement on execution time is most of the time negligible. They conclude that the ....

A. S. Vaidya, A. Sivasubramaniam, and C. R. Das. Performance benefits of virtual channels and adaptive routing: An application-driven study. In Procs. of the 11th Int'l Conf. on Supercomputing (ICS-97), pages 140--147, July 1997.

....of buffer space. We model our MIN connected CC NUMA multiprocessor in a detailed execution driven simulator to study the impact of the advancements in switch design on the execution of several applications. Executiondriven evaluations for mesh interconnection networks have been reported recently [38, 76] to test the effectiveness of virtual channels and adaptive routing. These studies evaluate the multiprocessors at the network system level. We perform our measurements at each crossbar switch within the MIN to analyze network interference and waiting delays accurately. C. Memory Management ....

....the effect of memory management. Since memory management dictates the placement of data, it determines the number of messages that are transferred over the network. It also affects dynamic changes in traffic behavior over the execution time of applications. Recently, execution driven simulations [38, 76] reported contradictory results on the performance benefits of virtual channels. The primary reason for their contradiction was the use of widely different memory management policies. As a result, a study on the effect of memory management is necessary to accurately judge the benefits of network ....

[Article contains additional citation context not shown here]

A. Vaidya, A. Sivasubramaniam and C. Das, "Performance Benefits of Virtual Channels and Adaptive Routing: An Application-Driven Study," Proc. 11th International Conference on Supercomputing, Jul. 1997.

....the execution of applications. The aim of this paper is to evaluate the performance of an IN with realistic application data accesses and memory management policies governing the placement of the data blocks. Execution based evaluations for mesh interconnection networks have been reported recently [9, 10] to test the effectiveness of virtual channels and adaptive routing. While these studies provide useful data, they do not explore different memory management techniques as we do. Also, several cache coherence studies [11, 12] do not model the IN switches explicitly, thus ignoring possible inter1 ....

....message latency tremendously. Advanced switch architectures together with the first touch policy gives the best performance for all the applications. Contrary to current assumptions of constant network delays, we show that the impact of switch architectures is significant. In a recent study [10], the performance benefits of virtual channels were shown to be negligible contradicting earlier work in [9] where a buddy allocation was used. However, their work [10] was based on a hand optimized memory management policy, which underestimates the frequency of remote memory accesses in realistic ....

[Article contains additional citation context not shown here]

A. Vaidya, A. Sivasubramaniam and C. Das, "Performance Benefits of Virtual Channels and Adaptive Routing: An Application-Driven Study," Proc. 11th International Conference on Supercomputing, Vienna, July 1997.

....the introduction of techniques such as virtual channel [3] and adaptive routing [7] The effect of all these advances in the INs has to be judged from the dynamic changes during the execution of applications. Executionbased evaluations for mesh interconnection networks have been reported recently [8, 9] to test the effectiveness of virtual channels and adaptive routing. These studies evaluate the multiprocessors at the network level and do not reflect the effect of design details of switch hardware. Also, the measurements are done at the execution time level instead of a detailed study at the ....

....for all applications differed greatly depending on the memory management policy. Results showed that buddy and round robin memory management techniques greatly improve their performance when virtual channels are employed. The benefit is minimal for the first touch policy. The work reported in [9] assumed a unrealistic hand optimized memory allocation and contradicted earlier results in [8] without realizing that the latter paper used a buddy allocation policy. As we demonstrated in this paper, the degree of performance improvement varies widely with the memory management techniques and ....

A. Vaidya, A. Sivasubramaniam and C. Das, "Performance Benefits of Virtual Channels and Adaptive Routing: An Application-Driven Study," Proc. 11th International Conference on Supercomputing, Vienna, July 1997.

....the introduction of techniques such as virtual channel [5] and adaptive routing [7] The effect of all these advances in the INs has to be judged from the dynamic changes during the execution of applications. Executionbased evaluations for mesh interconnection networks have been reported recently [1, 8, 14] to test the effectiveness of virtual channels and adaptive routing. These studies evaluate the multiprocessors at the network level and do not reflect the effect of design details of switch hardware. Also, the measurements are done at the execution time level instead of a detailed study at the ....

A. Vaidya, A. Sivasubramaniam, and C. Das. Performance benefits of virtual channels and adaptive routing: An application-driven study. Proc. 11th International Conference on Supercomputing, Vienna, July 1997.

....faster interconnection networks than multicomputers because messages are shorter (a typical message consists of a cache line and some control information) and much more frequent. While the interconnection network is not the bottleneck yet, some researchers began to report that a lower latency [21] and a higher network bandwidth [2] may significantly reduce the execution time of several parallel applications. As processor clock frequency is increasing at a faster rate than network bandwidth, the interconnection network may become a bottleneck within the next few years [7] The situation is ....

....Finally, in Section 5 some conclusions are drawn. 3. A Simulation Model Based on Application Message Traces In general, the interconnection network is not the bottleneck in current DSM systems. Therefore, techniques that improve network throughput are of little interest in this environment [21]. Nowadays, DSMs and NOWs are designed by using the same processors. The main architectural differences between these machines are the interconnection network and communication assisting circuitry such as the node controller. In this paper, we focus on the network. In a NOW, links are usually ....

[Article contains additional citation context not shown here]

A. S. Vaidya et al., "Performance benefits of Virtual Channels and Adaptive Routing: An Application Driven Study," in Int. Conf. on Supercomputing, 1997

....interconnection network do not deliver very good performance due to the high latency associated with shared memory accesses. A significant portion of this latency is incurred by messages in the interconnection network [6] and a lot of research is currently taking place to reduce this latency [19, 27, 5, 22]. Virtual channels [7] have been proposed as an attractive mechanism to avoid deadlocks [1, 8] provide adaptivity [4] increase throughput [9] and reduce latency in wormhole routed networks [9] Most of these studies on using virtual channels mechanism in wormhole networks have focused on ....

....for improving network performance. Out of these products, the SGI Spider interconnect [17] is directly geared towards cache coherent DSM systems. However, this study [17] does not clearly outline the performance benefit of virtual channels in DSM systems. Recently, an application driven study [27] has shown that only a negligible performance benefit exists in using virtual channels and adaptive routing in a DSM system. According to this study, the enhancements due to virtual channels and adaptive routing might not be justified considering the associated increase in router complexity. ....

[Article contains additional citation context not shown here]

A. S. Vaidya, A. Sivasubramaniam, and C. R. Das. Performance benefits of virtual channels and adaptive routing: An application-driven study. In Proceedings of ACM International Conference on Supercomputing (ISC'97), July 1997.

....and contention, that an application experiences, depend on a number of factors including the connectivity, the bandwidth of the links in the network, the switching delays, and the length of the message. Of these factors, link bandwidth and connectivity are the most crucial network parameters [48]. In this section, we synthesize the bandwidth requirements of the chosen applications for the binary hypercube platform as a function of the following system parameters: the number of processors, CPU clock speed, and application problem size. We also discuss how these results can be used to ....

A. Vaidya, A. Sivasubramaniam, and C. Das, "Performance Benefits of Virtual Channels and Adaptive Routing: An Application-driven Study," in Proceedings of the ACM 1997 International Conference on Supercomputing, July 1997, pp. 140--147.

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