| Owicki S. S., Karlin A. R.: Factors in the performance of AN1 Computer Network. SRC Research Report 88. June, 1992. |
....avoids deadlocks during the reconfiguration process by stopping application traffic before starting the reconfiguration process. When the reconfiguration finishes, packet transmission is allowed again. As a consequence, performance degradation of the interconnection network is produced. In [4,5], Rodeheffer and Owicki analyzed the reconfiguration effect on average packet latency. Figure 1 shows this effect. 0 50 100 150 200 0 1 2 3 Time (s) Reconfigurations Figure 1. Consequences of the static reconfiguration in Autonet. Under normal conditions, the network latency varies from ....
....routing tables asynchronously without stopping message traffic while guaranteeing the absence of deadlock. The first step in network reconfiguration is detecting the addition and removal (or failure) of network components (links, switches, and or hosts) This issue has been addressed elsewhere [1,4,5] and is beyond the scope of this paper. Once a change in the topology has been detected, it is necessary to update the routing tables at one or more switches and or hosts. Our primary focus is on achieving a distributed update of the routing tables without stopping message traffic and without ....
Owicki S. S., Karlin A. R.: Factors in the performance of AN1 Computer Network. SRC Research Report 88. June, 1992.
....problem has two parts. First, to figure out a graph that represents the topology of the network and second, given a topology, to compute routes. We refer to the first problem as the mapping problem. There has been a lot of work in computing routes with specific properties from network topologies [73, 67, 22, 38]. One of the most important properties is that routes should be deadlock free. In this section we present a new method for solving the mapping problem, and we discuss how to solve the problem of deadlock free routes. In the following paragraphs we discuss some related work in solving the mapping ....
....The algorithm has to be able to detect cycles and multiple paths to the same node. After the graph that represents the network topology has been determined, it is sent to all nodes in the system. Each node then computes deadlock free routes to all other nodes using the UP DOWN algorithm [67] 2 . The Myricom mapping algorithm runs periodically and maps the full network, in the presence of arbitrary tra#c in the network. A second approach is described in [62] where the topology of the network is determined with a di#erent algorithm. Routes are again computed using the UP DOWN 2 ....
[Article contains additional citation context not shown here]
S. S. Owicki and A. R. Karlin. Factors in the performance of the AN1 computer network. Technical Report 88, DEC System Research Center, 130 Lytton Ave., Palo Alto, CA 94301, June 1992.
....study only considers 2 levels of priority. However, the scheme presented can be extended to support multiple priorities. Deadlock is a problem in wormhole routing networks. To avoid it, a deadlock avoidance routing scheme is used. Up down routing is used in networks such as Myrinet and Autonet [OK92] to avoid deadlock and is implemented in the simulator. Up down routing avoids deadlock by restricting the paths of worms to a set of deadlock free routes. Performance metrics of interest are message network latency and throughput. We define message latency as the time between the moment a ....
Susan S. Owicki and Anna R. Karlin. Factors in the performance of the an1 computer network. Performance Evaluation Review, 20(1), June 1992.
....mapper added to A To appear in the proceedings of the 9th Annual Symposium on Parallel Algorithms and Architectures (SPAA 97) Once the master or elected leader generates a network map, it derives mutually deadlock free routes from it and distributes them throughout the system. We use UP DOWN [5] routing to impose an edge ordering such that all valid routes are paths that follow zero or more edges in the up direction, followed by zero or more edges in the down direction. A valid route never turns from a down edge onto an up edge. To compute the edge orderings, the algorithm picks a switch ....
....[6] all pairs shortest paths algorithm to compute compliant paths between all hosts. Where multiple edges are available between two switches, the algorithm has the option of randomly choosing among them for load balance. The goodness of UP DOWN routes is known to be highly topology dependant [5]. Two common effects are increased congestion about the root and the creation of locally dominant switches in hypercubic networks. The BFS numbering of these switches is such that all edges lead away from them. Consequently, no route will ever use them because doing so would introduce a turn from ....
[Article contains additional citation context not shown here]
S. Owicki and A. Karlin. Factors in the Performance of the AN1 Computer Network. Digital Equipment Corporation Systems Research Center, Palo Alto, CA, 1992.
....that few cells are dropped with a uniform workload [Giacopelli et al. 91] unfortunately local area network traffic is rarely uniform. Instead, a common pattern is client server communication, where a large fraction of incoming cells tend to be destined for the same output port, as described by Owicki and Karlin [1992]. Unlike previous generation networks, fiber links have very low error rates; the links we are using in AN2, for example, have a bit error rate of less than 10 Gamma12 . Thus, loss induced by the switch architecture will be more noticeable. Another technique, often combined with the previous ....
Owicki, S. and Karlin, A. Factors in the Performance of the AN1 Computer Network. In Proc. 1992 ACM SIGMETRICS and PERFORMANCE '92 Conference on Measurement and Modeling of Computer Systems, pages 167--180, June 1992.
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