L. Ni, Y. Gui, and S. Moore. Performance evaluation of switch-based worm hole networks. Journal of Parallel and Distributed Computing, 8(5):462-474, May 199L  Home/Search   Document Details and Download   Summary   Related Articles   Check

....allow the control cmnmunication pattern to be implemented without blocking the network must be partitionable into contention free and channelbalanced disjoint k ary cubes, as described in 3] The processors that must communicate in each step of the distributed procednre form a k ary t cube. From [3], a cube banyan network (and by symmetry, the reverse cube network of Figure 1) has the reqnired property. Thus, the communication required for each step of the distributed algorithm can be accomplished with a single network state. Control communication therefore requires n network states ....

L. Ni, Y. Gui, and S. Moore. Performance evaluation of switch-based worm hole networks. Journal of Parallel and Distributed Computing, 8(5):462-474, May 199L

....of the largest parallel systems in the world, mainly connecting Compaq Alpha based servers, but increasingly other compute platforms too. The performance of interconnection networks and, in particular, switch based wormhole networks has been extensively analyzed by simulation in the literature [2,11]. Since # Corresponding author. performance is strongly influenced by the load, it is very important to evaluate the QsNET under more varied traffic patterns to get a complete view of the network behavior. The patterns considered in this work are representative of real scientific applications in ....

L.M. Ni, Y. Gui and S. Moore, Performance evaluation of switch-based wormhole networks, IEEE Transactions on Parallel and Distributed Systems 8(5) (1997) 462--474.

....some of the largest parallel systems in the world, mainly connecting Compaq Alpha based servers, but increasingly other compute engines too. The performance of interconnection networks and, in particular, switch based wormhole networks has been extensively analyzed by simulation in the literature [12] [3] Since performance is strongly influenced by the load, it is very important to evaluate the QsNET under more varied traffic patterns to get a complete view of the network behavior. The patterns considered in this work are representative of real scientific applications in use at Los Alamos. ....

Lionel M. Ni, Yadong Gui, and Sherry Moore. Performance Evaluation of Switch-Based Wormhole Networks. IEEE Transactions on Parallel and Distributed Systems, 8(5):462--474, May 1997.

....how control cycles can alter the states used for data communication. 4.1. Allocating network bandwidth. To allow the control communication pattern to be implemented without blocking the network must be partitionable into contention free and channelbalanced disjoint k ary cubes, as described in [3]. The processors that must communicate in each step of the distributed procedure form a k ary 1 cube. From [3] a cube banyan network (and by symmetry, the reverse cube network of Figure 1) has the required property. Thus, the communication required for each step of the distributed algorithm can ....

....allow the control communication pattern to be implemented without blocking the network must be partitionable into contention free and channelbalanced disjoint k ary cubes, as described in [3] The processors that must communicate in each step of the distributed procedure form a k ary 1 cube. From [3], a cube banyan network (and by symmetry, the reverse cube network of Figure 1) has the required property. Thus, the communication required for each step of the distributed algorithm can be accomplished with a single network state. Control communication therefore requires n network states ....

L. Ni, Y. Gui, and S. Moore. Performance evaluation of switch-based wormhole networks. Journal of Parallel and Distributed Computing, 8(5):462--474, May 1997.

....can therefore be developed in a control cycle of n steps, proceeding from stage 0 to stage n Gamma 1. To allow the control communication pattern to be implemented without blocking the network must be partitionable into contention free and channel balanced disjoint k ary cubes, as described in [8]. The processors that must communicate in each step of the distributed procedure form a k ary 1 cube. From [8] a cube banyan network (and by symmetry, the reverse cube network of Figure 1) has the required property. Thus, the control communication required for each step of the distributed ....

....allow the control communication pattern to be implemented without blocking the network must be partitionable into contention free and channel balanced disjoint k ary cubes, as described in [8] The processors that must communicate in each step of the distributed procedure form a k ary 1 cube. From [8], a cube banyan network (and by symmetry, the reverse cube network of Figure 1) has the required property. Thus, the control communication required for each step of the distributed algorithm can be accomplished with a single network state. 3. Dynamic determination of the network state In this ....

L. Ni, Y. Gui, and S. Moore. Performance evaluation of switch-based wormhole networks. Journal of Parallel and Distributed Computing, 8(5):462#474, May 1997.

....how control cycles can alter the states used for data communication. 4.1. Allocating network bandwidth. To allow the control communication pattern to be implemented without blocking the network must be partitionable into contention free and channel balanced disjoint k ary cubes, as described in [3]. The processors that must communicate in each step of the distributed procedure form a k ary 1 cube. From [3] a cube banyan network (and by symmetry, the reverse cube network of Figure 1) has the required property. Thus, the communication required for each step of the distributed algorithm can ....

....the control communication pattern to be implemented without blocking the network must be partitionable into contention free and channel balanced disjoint k ary cubes, as described in [3] The processors that must communicate in each step of the distributed procedure form a k ary 1 cube. From [3], a cube banyan network (and by symmetry, the reverse cube network of Figure 1) has the required property. Thus, the communication required for each step of the distributed algorithm can be accomplished with a single network state. Control communication therefore requires n network states ....

L. Ni, Y. Gui, and S. Moore. Performance evaluation of switch-based wormhole networks. Journal of Parallel and Distributed Computing, 8(5):462--474, May 1997.

....[52] Efficient hardware communication support can be implemented for these operations to reduce their latency further. Additional work in this area is needed. ffl Research efforts on wormhole routing in switch based networks should continue because of the renewed interests in such architectures [54]. ffl Finally, application specific routing algorithms are needed for specialized systems, probably related to the routing of real time traffic or multiple classes of traffic [60] This report has concentrated on hardware routing algorithms. However, some functionality can be implemented in ....

L. M. Ni, Y. Gui, and S. Moore, "Performance Evaluation of Switch-Based Wormhole Networks, " Int. Conf. on Parallel Processing, 1995.

....In contemporary multicomputer systems, the wormhole switching mechanism [2, 3] is used to improve the communication latency. The wormhole switching technique has been studied extensively for direct networks [3] However, studies on wormhole switching for indirect networks have been limited [4]. Multicomputer systems increase their computing speed by performing several computations concurrently. These activities require coordination and synchronization between processing elements through interprocessor communication which can be either one to one (unicast communication) or within a ....

....flit in a pipelined fashion. The pipelined flow control dramatically reduces the communication latency and is insensitive to the distance in a contention free network. Only buffer size of one flit is required at each input channel. Detailed treatments of wormhole routing in MINs can be found in [4]. Switching elements are the main components of a MIN. As shown in Figure 2, a switch comprises of input channels with the associated buffer(s) crossbar interconnection, and a set of output channels. There are a few alternatives for the location of buffers in the switch. In this work, we consider ....

L. M. Ni, Y. Gui, and S. Moore, "Performance evaluation of switch-based wormhole networks," in International Conference on Parallel Processing, August 1995.

....in switch based networks, each channel in the network should not be utilized by processors from different processor partitions. Our recent work has shown that a cube unidirectional MIN with N = k n nodes can be partitioned into contention free and channel balanced 2 disjoint binary cubes [20]. Such cube based allocation restricts that the binary addresses of those allocated processors in a processor partition must form a subcube 3 . A constraint of this policy is that each processor partition always has to be a subcube. This 2 In channel balanced partition, the number of channels ....

L. M. Ni, Y. Gui, and S. Moore, "Performance evaluation of switch-based wormhole networks," in Proc. of the 1995 International Conference on Parallel Processing, vol. I, Aug. 1995.

....in Switch Based Networks Switch based networks, or indirect networks, have emerged or resurged as another promising network architecture for constructing scalable parallel computers. Most of the switch based networks are based on some variations of multistage interconnection networks (MINs) [19]. ffl Difficult to avoid deadlock with path based approach ffl An example of deadlock for a path based multicast: a path from node 0 to nodes 4 and 5 (Fig. 8) 20] 6.1 Asynchronous Multicast Trees An example of two deadlocked asynchronous multihead worms: one rooted at 0 and the other rooted ....

L. M. Ni, Y. Gui, and S. Moore, "Performance evaluation of switch-based wormhole networks," in Proc. of the 1995 International Conference on Parallel Processing, vol. I, Aug. 1995.

....technique is not restricted to these two networks. This concept can be applied to any network as long as the underlying networks can be partitioned into contention free processor clusters. In some networks, such as a butterfly unidirectional MIN, this partitioning may not be possible [4]. In this case, the best one can do is to minimize network contention as much as possible. Instead of preventing a common communication channel used by different senders at any time, some channels are allowed to be shared. However, the senders who share the same communication channels are ordered ....

L. M. Ni, Y. Gui, and S. Moore. Performance evaluation of switch-based wormhole networks. In Proc. of the 1995 International Conference on Parallel Processing, volume I, Aug. 1995.

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