PANDA, SINGAL AND KESAVAN. Multidestination message passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. In IEEE Transactions on Parallel and Distributed Systems (1999).  Home/Search   Document Details and Download   Summary   Related Articles   Check

....difficult to support multicast. Multicast is an important collective communication in scalable parallel computers, in which the source node sends the same data to an arbitrary number of destination nodes [2] Many multicast routing algorithms have been studied for systems with regular topologies [2, 3, 4, 5]. Usually, the proposed multicast algorithms This work was supported in part by National Science Council under grants NSC 86 2213 E 007 043 and NCHC86 08 024. are based on either one of two schemes: unicast [2] and multidestination messaging [3] In the unicast based approach, multicast is ....

D. K. Panda, S. Singal, and R. Kesavan, "Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths," Technical Report, OSU-CISRC-12/94TR54, Dep. of Computer Science, Ohio-state University.

....by other alternatives; these alternatives include using multiple unicast messages for multicast communication and using a broadcast message to accomplish multicasting. There are several studies on the performance of multicast communication in multicomputer networks [20] 15] 16] 17] 24] [23], 9] 18] 10] Lin et al. 17] 16] present two multicast algorithms, dual path and multipath, based on the Hamiltonian paths in the interconnection network. These are especially suited for wormhole switching in low dimension mesh and torus networks. Wormhole switching is a form of ....

....channels in each node. Multiple consumption channels can be implemented using a single physical consumption channel on which several virtual consumption channels are time multiplexed. Our results presented here expand on our earlier work in [4] 3] Liu and Duato [18] and Panda et al. 24] [23] have independently discovered the consumption channel deadlock problem and have given solutions to specific algorithms. In this paper, we give more general results on this problem and provide sufficiency conditions for the minimum required consumption channels to avoid deadlocks. Another issue ....

[Article contains additional citation context not shown here]

# D.K. Panda, S. Singhal, and R. Keshavan, "Multidestination Message Passing in Wormhole k-Ary n-Cube Networks with Base Routing Conformed Paths," manuscript, Dec. 1995.

....the communication steps. In order to support TBM, the message format provides efficient multi address encoding to be easily implemented at routers. By detailed analysis and simulation, the results show that TBM is more preferable than traditional Umesh[1] Hamiltonian Path[2] and BRCP HL(C,R)[3] multicast schemes, which indicates that current and future massively parallel systems can take advantage of this scheme to implement fast and scalable collective communication operations. 1 Introduction Multicast communication, i.e. the same message is delivered from a source node to an ....

....node by containing n destinations in the header of message. However, in order to avoid deadlock, the message propagation is restricted to a Hamiltonian Path, which leads to visit all the destination nodes serially. Therefore, the long path length results in high latency in multicast operation. In [3], D.K.Panda et al. introduced many grouping schemes for multidestination message passing, in which the Hierarchical Leader based(HL) scheme is implemented under the Base Routing ConformedPath model. With this scheme, all destination nodes are grouped into different level leaders. Taking the case ....

[Article contains additional citation context not shown here]

D.K.Panda, S.Singal and R.Kesavan, "Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths," IEEE Transactions on Parallel and Distributed Systems, Vol. 10(1), pp. 76-96, 1999.

....less than 2 dlog2 (n 1)e Gamma1 acknowledgement messages are required in one cache invalidation transaction when n processor nodes have the copies of the cache block. Detailed analysis and simulation in 2D mesh show that TBM is preferable to traditional Umesh[1] Hamiltonian Path[2] and BRCP HL[3] multicast schemes, which indicates that current and future DSM systems can take advantage of this scheme to deliver better performance. 1 Introduction Many commercial and research systems adopt the directory based write invalidate cache coherence protocol, such as Stanford DASH [4] FLASH[5] ....

....message and forwarding the invalidation message to the next destination node, can be completed by the intermidate nodes. However, the Hamiltonian path based scheme will visit all the destination nodes serially, therefore results in long latency. We use Hampath to denoted this scheme. In [3], D.K.Panda introduced many grouping schemes for multidestination message passing, in which the Base Routing Conformed Path modelbased Hierarchical Leader(BRCP HL) scheme is implemented in 2D mesh. In such scheme, the invalidation messages will be sent to the level Gamma2 leaders by the home node ....

[Article contains additional citation context not shown here]

D.K.Panda, S.Singal and R.Kesavan, "Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths," IEEE Transactions on Parallel and Distributed Systems, Vol. 10(1), pp. 76-96, 1999.

....Moreover, the TBM scheme provides more opportunities for further cache optimization and allows host processors to work in parallel with the message transfer. Detailed analysis and simulation in 2D mesh show that TBM is more preferable than traditional Umesh[1] Hamiltonian Path[2] and BRCP HL[3] multicast schemes, which indicates that current and future DSM systems can take advantage of this scheme to deliver better performance. Keywords: DSM systems, cache coherence protocol, write invalidate, TBM, multidestination message passing. 1 Introduction Many commercial and research ....

....message and forwarding the invalidation message to the next destination node, can be completed by the intermidate nodes. However, the Hamiltonian path based scheme will visit all the destination nodes serially, therefore results in long latency. We use Hampath to denoted this scheme. In [3], D.K.Panda introduced many grouping schemes for multidestination message passing, in which the Base Routing Conformed Path model based Hierarchical Leader(BRCP HL) scheme is implemented in 2D mesh. In such scheme, the invalidation messages will be sent to the level Gamma 2 leaders by the home ....

[Article contains additional citation context not shown here]

D.K.Panda, S.Singal, and R.Kesavan. Multidestination message passing in wormhole k-ary n-cube networks with base routing conformed paths. IEEE Transactions on Parallel and Distributed Systems, 10(1):76--96, 1999.

....studied under both store and forward routing (see for instance [18, 24] and wormhole routing (see for instance [2, 10, 19, 21, 22, 23, 27, 30, 32, 34] In this paper, we focus on the multicast problem under wormhole routing. We will make use of a particular hardware facility called path based [15, 23, 29] that allows intermediate destination to get a copy of the multicasted message. We will propose new algorithms for the mesh topology. These algorithms will be shown to perform faster than the previously known algorithms developed mainly by Lin, McKinley and Ni in [22] and [23] The next section ....

D. K. Panda, S. Singal, and R. Kesavan. Multidestination message passing in wormhole k-ary n- cube networks with base routing conformed paths. Technical Report OSU-CISRC-12/9-TR54, Dept. of Computer and Information Science, Ohio State University, 1995. (accepted for publication in IEEE Transactions on Parallel and Distributed Systems).

....supported in part by National Science Council under grants NSC 86 2213 E 007 043 and NCHC 86 08 024. Due to its importance, hardware, software, or hybrid approaches to supporting barrier synchronization have been proposed [2, 3, 4, 5] Among the schemes, the one using multidestination messaging [7, 8] has attracted a lot of attention [5] The basic concept of multidestination messaging is to carry multiple destination addresses in a message so that a single message can be sent to multiple destinations. At each intermediate destination node, the router replicates the message, sends one copy to ....

....synchronization, the scheme proposed in [5] for example, uses two different types of multidestination worms, gather and broadcasting, in the reduction and distribution phase respectively. The destination traversing order specified in the message header must conform to the base routing algorithm [8]. Unfortunately, the scheme proposed in [5] only works for dimension ordered routing algorithms. For other base routing algorithms, such as those derived from the turn model [1] this scheme cannot take full advantage of their adaptivity. Furthermore, this scheme implicitly assumes that nodes ....

[Article contains additional citation context not shown here]

D. K. Panda, S. Singal, and R. Kesavan, "Multidestination Message Passing in Wormhole k-ary ncube Networks with Base Routing Conformed Paths," Technical Report, OSU-CISRC-12-95-TR54, Dep. of Computer Science, Ohio-state University.

....developed for clusters to support broadcast multicast with minimal NIC assistance, while delivering good performance. In this paper we take on such a challenge. We introduce a multidestination message passing mechanism. Such a mechanism has been developed earlier for router based parallel systems [11, 10, 13] to support ecient collective communication. In this paper we design and implement a multi send primitive to support ecient broadcast multicast that requires minimal assistance from the NIC. Our scheme is designed with the idea that as much processing as possible should be done by the host ....

D.K. Panda, S. Singal, and R. Kesavan, \Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths". IEEE Transactions on Parallel and Distributed Systems, Vol. 10, No. 1, pp. 76-96, January 1999.

....of the header decoding logic required at the switches. However, we assume the multidestination worms under either scheme conform to the base up down routing algorithm, i.e. the path followed by a multicast packet does not violate any of the rules for routing unicast packets in the system [28]. These schemes are discussed in greater detail in the following two subsections. 3.4.3 Tree Based Multicasting using Multidestination Worms Tree based multicasting places no restriction on the replication of a worm at a given switch. The basic idea of tree based replication is to have multiple ....

....for performing multicast in wormhole routed MINs using tree based multi head worms with synchronous replication. Similarly, path based multicast was originally proposed in the context of mesh networks by Lin and Ni [17] and was enhanced to a base routing conformed path model by Panda et al. [28]. Extensions of the latter method to support multiple multicast more efficiently have been proposed by Kesavan and Panda [14] None of these papers have evaluated the alternative enhanced multicasting schemes in comparison to one another in the context of irregular networks and in the presence of ....

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. Technical report, The Ohio State University, December 1995. IEEE Transactions on Parallel and Distributed Systems, to appear.

.... that make message communication easier by making message routing simpler, lowering the average distance per communication, and or increasing the bisection bandwidth [11] For such regular cutthrough networks, many multicast broadcast algorithms have been proposed in the literature in recent years [2, 5, 9, 10, 17, 20, 23, 25, 32]. More recently, cut through switching is being applied to switch based interconnects like Myrinet [4] and ServerNet [15] to build networks of workstations, or NOWs (also called workstation clusters) for cost effective parallel computing. In contrast to traditional parallel systems, these ....

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. IEEE Transactions on Parallel and Distributed Systems. In Press.

....like low latency communication and reduced communication hardware overhead [10] These systems use regular network topologies with various deadlock free routing schemes. For such regular wormhole networks, many multicast broadcast algorithms have been proposed in the literature in recent years [2, 5, 6, 8, 13]. This research is supported in part by NSF Grant MIP 9309627 and NSF Career Award MIP 9502294. More recently, wormhole routing is being applied to switch based interconnects like Myrinet [1] and ServerNet [4] to build networks of workstations for cost effective parallel computing. Switch based ....

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. Technical Report OSUCISRC -12/95-TR54, Dec. 1995. IEEE TPDS, under review.

....to multiple destinations with reduced software overhead can be proposed by augmenting the hardware support at the switch. Prior work on traditional regular topology parallel systems has shown how a message can be sent to multiple destinations in a single step (called a multidestination message) [22, 32, 33]. Some modern cut through switches possess the capability of simultaneous replication of an incoming message to multiple output ports [37] Although replication of messages in cut through switches is deadlock prone, deadlock free replication methods have been proposed [43, 48] This replication ....

....using such support. For example, efficient header encoding decoding schemes need to be designed keeping in mind that the new messages must conform to the base (unicast) routing supported by the system. Such a requirement prevents multicast messages from introducing additional deadlock scenarios [32]. Furthermore, the tradeoffs associated with the cost and complexity of the encoding decoding schemes need to be evaluated against the ability of the corresponding multicast messages to cover arbitrary multicast destination sets. For encoding schemes that are unable to capture arbitrary ....

[Article contains additional citation context not shown here]

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. Technical Report OSU-CISRC-12/95-TR54, The Ohio State University, December 1995. IEEE Transactions on Parallel and Distributed Systems. In Press.

....at least log 2 d(d 1)e phases to cover d destinations. Such unicast based multicasting schemes have been recently proposed for irregular networks [4] Multidestination routing has been proposed for parallel systems based on regular networks. Originally proposed for direct (router based) networks [5, 7], this scheme has been recently extended to regular switch based parallel systems [12, 14] However, the problem of how multidestination routing can be extended to irregular networks has not yet been addressed. Similarly, the extent to which such a multidestination message passing based multicast ....

....3 Tree Based Multidestination Message Passing Multidestination message passing is a technique that has recently been proposed for routing messages from a single source to multiple destinations. Proposed originally in the context of parallel systems based on direct (router based) networks [5, 7], this work has been extended recently to switch based networks [12, 14] A tree based multidestination worm covers multiple destinations in a switch based network by replicating at the switches on its path. The copies of the multidestination worm formed by replication at a switch are forwarded to ....

[Article contains additional citation context not shown here]

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. Technical Report OSU-CISRC-12/95-TR54, The Ohio State Univeristy, December 1995. IEEE Transactions on Parallel and Distributed Systems, to appear.

....of the header decoding logic required at the switches. However, we assume the multidestination worms under either scheme conform to the base up down routing algorithm, i.e. the path followed by a multicast packet does not violate any of the rules for routing unicast packets in the system [11]. These schemes are discussed in greater detail in the following two subsections. 3.2.3 Tree based Multicasting using Multidestination Worms Tree based multicasting places no restriction on the replication of a worm at a given switch. The basic idea of tree based replication is to have multiple ....

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. IEEE TPDS, to appear.

....primitives for point to point communication as well as various collective communication operations between processes. While considerable research has addressed the development of efficient algorithms for collective communication operations over regular networks such as meshes and hypercubes [1, 3, 4, 5, 8, 10, 11, 12, 14], much less work has been done for optimizing collective communication over irregular switch based clusters [9, 15] In this paper we address the collective communication operation of all to all broadcast (called All Gather in MPI) for switch based clusters with arbitrary topology. The paper is ....

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. Technical Report OSUCISRC -12/95-TR54, The Ohio State University, December 1995. IEEE Transactions on Parallel and Distributed Systems. In Press.

....primitives for point to point communication as well as various collective communication operations between processes. While considerable research has addressed the development of efficient algorithms for collective communication operations over regular networks such as meshes and hypercubes [1, 3, 4, 6, 9, 11, 12, 13, 15], much less work has been done for optimizing collective communication over irregular switch based clusters [10, 16] In this paper we address the collective communication operation of all to all broadcast (called All Gather in MPI) for switch based clusters with arbitrary topology. The paper is ....

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary ncube Networks with Base Routing Conformed Paths. Technical Report OSU-CISRC-12/95-TR54, The Ohio State University, December 1995. IEEE Transactions on Parallel and Distributed Systems. In Press.

....that collective communication operations can be implemented on MPP networks with reduced latency. This concept was futher extended to allow multidestination worms to take row column paths [2] and all paths conforming to the base routing using the Base Routing Conformed Path (BRCP) model [11]. Modern cut through switches possess the capability of simultaneous replication of an incoming worm to multiple output ports [13] Although this facility has been exploited for implementing broadcast by flooding the network, there has been little prior work on using this capability for arbitrary ....

....path based multidestination worms on such networks. 3.1 Multidestination Message Passing and Path based Worms The concept of wormhole message passing with multiple destinations was first introduced by Lin and Ni for [7] Hamiltonian meshes. Then, multidestination message passing was introduced [11] for k ary n cube networks with different base routing schemes. These schemes were proposed for regular direct networks typical of MPPs. In such networks each node consists of a router and a processor. The processor is connected to the router through a set of injection and consumption channels. ....

[Article contains additional citation context not shown here]

D. K. Panda, S. Singal, and R. Kesavan. Multidestination Message Passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. Technical Report OSU-CISRC-12/95-TR54, Dec 1995. IEEE TPDS, to appear.

No context found.

PANDA, SINGAL AND KESAVAN. Multidestination message passing in Wormhole k-ary n-cube Networks with Base Routing Conformed Paths. In IEEE Transactions on Parallel and Distributed Systems (1999).

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