D.A. Reed, R.M. Fujitomo, Multicomputer networks: Message based parallel processing, MIT Press, 1987.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....these techniques, and implement them on nCUBE 2 parallel computer to compare their performance. Our study indicates that the sampling technique is the better approach. 1 Introduction There are several architectures for designing multiprocessor database computers. However, the multicomputer model [14], is most popularly used for its scalability to support very large databases [1, 2, 4, 7, 12, 17, 18] In this architecture, the processing nodes (PNs) are interconnected through a communication network. Each PN has its own private memory and dedicated disk drives. Parallel database systems ....

D. A. Reed and R. M. Fujimoto. Multicomputer Networks: Message-Based Parallel Processing. Scienti c Computation. MIT Press, 1987.

....results in 2(wk =N) 2(w=k) bits per unit time for openended networks and 4(w=k) for wrapped ones. These parameters are given in Table I for some popular special cases of k ary d cubes. Notice that for the hypercube network only the maximum injection rate does not depend on the network size [7], which makes it very attractive for building a scalable massively parallel system. Another attractive feature of the binary hypercube is that it combines advantages of open ended and wrapped networks. The hypercube is totally symmetric, and so all its channels are loaded evenly; still, we can use ....

Reed, D. A., and Fujimoto, R. M. Multicomputer Networks: MessageBased Parallel Processing. MIT Press, Cambridge, MA, 1987.

....network has a full duplex bandwidth of 175 MBps (M bytes per sec. Myricom s Myrinet network [ 25 ] has a peak duplex bandwidth of 160 MBps and is eminently suitable for large scale embedded applications. Secondly, with the advent of generalized communication methods such as the routing switch [ 26 ] the fat tree [ 27 ] and wormhole routing [ 28 ] communication variance is of diminishing importance. Wormhole communication has reduced the message set up component. Routing congestion can be alleviated either by a randomized routing step [ 29 ] or by a changing pattern of ....

D. A. Reed and R. M. Fujimoto. Multicomputer Networks: Message-Based Parallel Processing. MIT, Cambridge, MA, 1988.

.... One of the most straightforward and the least expensive means of achieving this end is to construct multicomputer networks that consist of nodes with local memory (no shared memory) and a communication controller, where each node is connected by communication links to a number of other nodes [23]. Whenever a node wants to communicate with another node, it communicates through other nodes unless there exists a direct communication link between the two. The interconnection structure for a multicomputer network plays a central role in determining the overall performance of the system [14, ....

....[23] Whenever a node wants to communicate with another node, it communicates through other nodes unless there exists a direct communication link between the two. The interconnection structure for a multicomputer network plays a central role in determining the overall performance of the system [14, 23]. Since the 1960 s, many authors have been concerned with the problems associated with the design and analysis of interconnection structures [1, 2, 7, 10, 21, 23] One of the most popular interconnection structures being used is a hypercube [9, 25] We propose an interconnection structure for ....

[Article contains additional citation context not shown here]

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing, The MIT Press, 1987.

....applications use parallelism primarily to increase the disk I O bandwidth. The level of I O concurrency achievable determines the degree of parallelism can be attained. Several parallel computer architectures are suitable for database applications. Among them, the multicomputer architecture [18] has been recognized as having the most scalable structure for handling very large databases [2, 3, 6, 10, 15, 21, 22] In this architecture, the processing nodes (PNs) are interconnected through a communication network. Each PN has its own private memory and dedicated disk drives. In order to ....

Danial A. Reed and Richard M. Fujimoto. Multicomputer Networks: Message-Based Parallel Processing. Scientific Computation. MIT Press, 1987.

....difference between STM and MD will be even 12 smaller. That the worst case behavior is being considered in the above equations is due to the n factor in equation (7) Given our assumption that destination nodes are uniformly distributed the mean distance between two nodes in a hypercube is n=2 ( [17], p.52) Thus, if the n factor in equation (7) is replaced by n=2, the same derivation results in the mean packet delay equations for the three schemes. All the delays would be reduced by a factor of 1=2. 4 Summary Placing multiple processors on a node allows the introduction of a two level ....

D. Reed and R. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing. Cambridge, Massachusetts: MIT Press, 1987. 15

....for parallel operation. We have described a study which exposed these shortcomings in [6] Figure 1: Data Parallel and Explicit Message Passing models. The other popular programming model for massively parallel machines (and networks of workstations) is explicit message passing and C or Fortran [20]. This can be achieved via a number of message passing substrates such as PVM [12] MPI [10] the Crystalline OS [11] or any number of proprietary messaging systems. The basic idea is that concurrency is expressed as concurrent tasks which communicate and synchronize via message passing. These ....

D. Reed and R. Fujimoto. Multicomputer Networks: Message-based Parallel Processing. MIT Press, 1985. 20

....of the source node to the destination node increases. We assume that newly created tasks diffuse from areas of high processor utilization to areas of lower processor utilization with a bound on the maximum migration distance. Based on these assumptions, we are using the decay routing distribution[31] to characterize the mean internode distance of the SBCH topology. The general form of the mean internode distance is given in [31] l = lmax X l=1 lP (l) 1) where P (l) represents the probability for a packet to reach its destination at distance l. Then following the discussion in [31] the ....

....to areas of lower processor utilization with a bound on the maximum migration distance. Based on these assumptions, we are using the decay routing distribution[31] to characterize the mean internode distance of the SBCH topology. The general form of the mean internode distance is given in [31] l = lmax X l=1 lP (l) 1) where P (l) represents the probability for a packet to reach its destination at distance l. Then following the discussion in [31] the average message distance under the decay routing distribution is estimated: l = d Gamma 1 (d lmax Gamma 1) lmax X l=1 d ....

[Article contains additional citation context not shown here]

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing. Cambridge, MA: MIT Press, 1987.

....Width Network connectivity measures the resiliency of the network and its ability to continue operation despite disabled components. Connectivity can be defined as the minimum number of nodes or links that must be removed in order to partition 70 the network into two or more disjointed subnetworks[47]. The connectivity of a network can be divided in two ways: arc connectivity and node connectivity. The arc connectivity between two nodes i and j is the minimum number of links that must be removed from the network to disconnect nodes i and j. The node connectivity between two nodes i and j is ....

....removed from the network to disconnect nodes i and j. The node connectivity between two nodes i and j is the minimum number of nodes other than i or j that must be removed to disconnect i and j. The arc connectivity of a 2D spanning bus hypercube is 2 while its node connectivity is 2(w Gamma 1) [47]. A n dimensional hypercube has an arc connectivity of n and a node connectivity of n. Thus, in an SBCH(w;n) the arc connectivity is (2 n) or log 2 N Gamma 2log 2 w 2 and the node connectivity is (2(w Gamma 1) n) or 2(w Gamma 1) log 2 N Gamma 2log 2 w. The bisection width of a ....

[Article contains additional citation context not shown here]

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing. Cambridge, MA: MIT Press, 1987.

....bypass a faulty link which is competitive with other networks [12] 33] 3COMPARISONS OF SMLH WITH POPULAR NETWORKS In this section, we compare the SMLH network with existing, well known topologies. These include the Binary Hypercube (BHC) 8] the Generalized Hypercube (GHC) 12] the Torus [34], the Spanning Bus Hypercube (SBH) 21] the Hierarchical Cubic Network (HCN) 9] the Cube Connected Cycle (CCC) 35] the Hyper deBruijn (HdB) 36] the Folded Peterson (FPT) 37] the Hypermesh (HM) 33] and the Optical Multi Mesh Hypercube (OMMH) 17] The comparison parameters include ....

# D.A. Reed and R.M. Fujimoto, Multicomputer Networks: MessageBased Parallel Processing. Cambridge, Mass.: MIT Press, 1987.

....of the source node to the destination node increases. We assume that newly created tasks diffuse from areas of high processor utilization to areas of lower processor utilization with a bound on the maximum migration distance. Based on these assumptions, we are using the decay routing distribution[28] to characterize the mean internode distance of the SBCH topology. The general form of the mean internode distance is given in [28] l = lmax X l=1 lP (l) 1) where P (l) represents the probability for a packet to reach its destination at distance l. Then following the discussion in [28] the ....

....to areas of lower processor utilization with a bound on the maximum migration distance. Based on these assumptions, we are using the decay routing distribution[28] to characterize the mean internode distance of the SBCH topology. The general form of the mean internode distance is given in [28] l = lmax X l=1 lP (l) 1) where P (l) represents the probability for a packet to reach its destination at distance l. Then following the discussion in [28] the average message distance under the decay routing distribution is estimated: l = d Gamma 1 (d lmax Gamma 1) lmax X l=1 ....

[Article contains additional citation context not shown here]

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing. Cambridge, MA: MIT Press, 1987.

....could provide one solution to the implicit discussion on suitable architectures. The development of a modular processor, the transputer, has stimulated the discussion by providing accessible technology. A similar enterprise is that of a modular switch, for instance the Torus routing chip [ 16 ] It is a truism of parallelism that improvements in processing speed will come from being able to combine processors rather than further technological development of the uniprocessor, which had reached a plateau in 1983 [ 17 ] Though since 1983 there has been renewed activity, as [ 18 ] makes ....

D. A. Reed and R. M. Fujimoto. Multicomputer Networks: Message-Based Parallel Processing. MIT, Cambridge, MA, 1987.

....apply to store and forward networks, but we do not expect this to be too restrictive with current parallel and distributed systems. In first generation parallel systems, topology was an important issue [ 13 ] With the advent of generalized communication methods such as the routing switch [ 14 ] the fat tree [ 15 ] and wormhole routing [ 16 ] the situation appears less critical. On second generation machines, wormhole communication has reduced the message set up component. Routing congestion can be alleviated either by a randomized routing step [ 17 ] or by a changing pattern ....

....though it does not in this instance exceed the performance of uniform tasksize scheduling. A simple estimate of the minimum time can be taken from t factor = n p vh; 17) where v is the number of scheduling rounds from factoring. In Table 1, the minimal time estimates arising from equations 14 17 are compared. The two estimates form upper and lower bounds for the timings. An exact result for the exponential cdf (here = 1) is available by solving (3) p = p X i=2 1 i = lnp fl O(p Gamma1 ) 18) with fl being Euler s constant (0:5772157 : 6 . By using (7) with (18) ....

D. A. Reed and R. M. Fujimoto. Multicomputer Networks: Message-Based Parallel Processing. MIT, Cambridge, MA, 1988.

....parallel execution. This is the focus of the thesis. 1.2 Parallel Discrete Event Simulation Here, PDES assumes a Logical Process Modeling Methodology for simulation software development. The system to be simulated is conceptualized as a network of inter 6 acting Physical Processes (PPs) [22, 23, 38, 71]. A distributed simulation model of the physical system is realized by an equivalent system of computational units called Logical Processes (LPs) State transitions in the physical system are modeled by the occurrence of an event. Each LP is responsible for modeling the state transitions in a ....

D. Reed and R. Fujimoto. Multicomputer Networks: Message-Based Parallel Processing. The MIT Press, 1987. MIT Press Series in Scientific Computation, Editor: Dennis Gannon.

....those studied here are examples of what are often referred to as multicomputers, multicomputer networks or microprocessor based multiprocessors. These consist of a number of microprocessors, each of which has some local memory and which can communicate with other such processors over a network [17,18]. The multicomputer considered here is based on the INMOS Transputer. The transputer is a generic name for a family of high performance devices that have been developed specifically for use in multiprocessor systems. At the heart of a transputer is a highperformance (up to 25 MIPS) 16 bit or ....

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing, Cambridge, MA: MIT Press, 1987.

....input queue length q i to be one and specifying a crossbar switch for the switch routing unit. If the output queues are discarded from the model in Figure 2, an input buffered switch box (Figure 3b) 8, 18] can be modeled. Now consider a central memory buffering switch as shown in Figure 3(c) [20]. Without a buffer allocation constraint, incoming packets are stored in the central memory as long as the memory with size M is not full. With buffer allocation limitation, incoming packets are stored in the central memory as long as either the memory is not full, or the number of packets stored ....

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-based Parallel Processing, MIT Press, Cambridge, MA, 1987. 22

....32 bit data and 4 bit reserved field) over a 64x64 grid network. The router chip lacks special interfacing to a dataprocessor, so packets must be injected into and removed from the network by using the normal links. 1 Because of space limitations, other important designs such as described in [15] and [9] are omitted. The CPs listed here clearly illustrate the great variety of possible design choices. 2 The processors are also connected in a 256x256 grid. However, this grid doesn t have hardware support for the routing of messages. DOOM Communication Processor [4] The CP, developed for ....

Daniel A. Reed and Richard M. Fujimoto. Multicomputer Networks Message-Based Parallel Processing. MIT press, 1987.

....increases, the granularity of computation does not need to increase to keep the machine balanced. Thus, the machine can be scaled up to an arbitrary number of processors. Architectural scalability is claimed by many different architectures, including hypercubes, meshes, tori, and rings. But as Reed[7] points out, scalability of an architecture must be claimed with respect to a particular application and the communication patterns that the application exhibits. For example, a two dimensional torus or mesh is scalable with respect to 2 D nearest neighbor stencil applications such as ....

) Daniel A. Reed, Richard M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing, The MIT Press, Cambridge, Massachusetts, 1985.

....topology is its parameter called dimension. The value of this parameter determines the number of vertices and the number of edges of the hypercube. Dimension can be used, in a natural way, to label vertices and to facilitate routing between an arbitrary pair of vertices. See, for example, [4, 8]. It has long been known that broadcasting and gossiping can be easily accomplished in a hypercube, by using dimension. See [2] for references on these problems. However, hypercubes can only be constructed for n = 2 d vertices. In this paper, we show that a set of graphs, constructable for any ....

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing, MIT Press, Cambridge, 1987.

....uniform message routing (i.e. random destinations) is assumed. In many applications, spatial locality exists in communication patterns. For example, a processor may send a message to the processors within certain distance (i.e. number of hops) with a higher probability than to the other processors [33]. The effect of such non uniform message routing on the latency reduction achievable in PM needs to be investigated further. 5 Concluding Remarks With the high bandwidth available in optics, it becomes increasingly important to reduce the communication latency which ultimately limits the ....

D. Reed and R. Fujimoto, Multicomputer networks: message-based parallel processing. MIT Press, 1987.

....to route packets around congested or faulty areas [16, 13, 5] It is important to minimize the addressing and control information that must be sent with each packet as well as to maximize the availability of network resources for active connections. With message transport based on virtual circuits [15, 3], many of the desirable properties described above are realized. ##################### Supported by Hughes Aircraft Company and the State of California MICRO program. Y. Turner is supported by a Hertz Foundation Graduate Fellowship. Packets are sent through pre established logical paths, thus ....

....basic technique of Dynamic Virtual Circuits is described. The hardware support for Dynamic Virtual Circuits in the ComCoBB chip is described in Section IV. II. Static Virtual Circuits The two fundamental approaches to message routing in multicomputers are circuit switching and packet switching [15, 3]. With circuit switching, a static physical path is set up between the sender and receiver before communication takes place. Once the path is set, data can be transmitted quickly at nearly the full bandwidth of the links with almost no redundant control information [5] A disadvantage of circuit ....

[Article contains additional citation context not shown here]

D. A. Reed and R. M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing, The MIT Press (1987).

No context found.

D.A. Reed, R.M. Fujitomo, Multicomputer networks: Message based parallel processing, MIT Press, 1987.

No context found.

Reed, D.A. and Fujimoto, R.M. "Multicomputer Networks: Message-based parallel processing". The Mit Press. London, England, 1987.

No context found.

D.A. Reed, and R.M. Fujimoto, Multicomputer Networks: Message-Based Parallel Processing, MIT Press, 1987.

No context found.

Reed, D. A., Fujimoto, R., "Multicomputer Networks: Message-Based Parallel Processing", MIT Press, 1988.

First 50 documents

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