You need to know about the network topology

The topology of the cluster and its performance for computational tasks, of course, things are connected. The smaller the time delay in the transaction data between nodes – the higher the speed. topology The best from this point of view is the mesh (mesh). In this topology, all nodes are connected to each other without any switching brokers. And so every transfer of data from node to node is a one hop (singlehop). This topology requires the N-1 communication adapters for N nodes. Obviously, the total number of cards is proportional to the square of the number of nodes in the cluster. build such a cluster is very expensive, so actually building clusters with communication intermediaries – switch (switch). The simplest solution in this case – is a star topology. One switch for all nodes. In this case, each node is connected to each one through an intermediary and therefore the transmission of bits of information is always a 2-hop (doublehop). This topology is actually the case in clusters with a small number of nodes. The main limitation is the number of ports in a commercially available switch. The author knows of switches with 48 ports. If the number of cluster nodes is so large that a single switch is not enough, you have to go to increase the number of hop by increasing intermediary switches connected to each other. The classic solution to this is a thick tree topology (fattree). In provides an analysis of the effectiveness of this topology. Lately topology build clusters appeared innovations associated with the blade server technology combined in a single 7U chassis sizes for 19 “rack. The structure of the chassis may already include a switch ( 10 Gbit Ethernet or InfiniBand), which allows not only for the inside of the chassis to connect with each other, but also have multiple connections out to create a topology connection blade nodes. Here we have a blade assembly – a computing power in a single chassis 7U. Besides classical solutions thick tree being considered alternative solutions, which are called hypercube. This solution benefits significantly in reducing the cost of switching equipment. Experimental study on the subject conducted by Spanish experts in HPC. They were built experimental clusters on SuperMicro blade production for different topologies and studies the performance of various algorithms. The result was that it all depends on the algorithm. For some algorithms, the performance of the hypercube and thick tree is practically the same, and for some thick tree has a 4-fold gainC. Minkenberg, Design of HPC systems, Springer Proceedings in Physics: Use Cases of Discrete Event Simulation … 2012Fabrizio Petrini, Marco Vanneschi. k k-Ary n-trees: High Performance Networks for Massively Parallel ArchitecturesJavier Navaridas, Jos? Miguel-Alonso. Indirect cube topology for small-and medium-scale clusters, Dep. de Arquitectura y Tecnolog?a de Computadores Universidad del Pa?s Vasco – UPV / EHU P. Manuel de Lardizabal 1, 20080 San Sebasti?n

Related articles
  • Network Topologies. Bus TopologyNetwork Topologies. Bus Topology
    All computers on the network are connected by lines of communication. Geometric arrangement of lines relative to hosts and physical connection to the network nodes called the physical topology. Dep...
  • Topology of “star”Topology of “star”
    The network was built on the topology of "star" each workstation connected cable (twisted pair) to the hub or hub (Hub) . Concentrator provides parallel computer and, thus, all the computers connec...
  • Token Ring topologyToken Ring topology
    This topology is based on the topology of the "physical ring connecting a star." In this topology, all workstations are connected to a central hub (Token Ring) in a physical star topology.  Centra...
  • The topology of the “ring”The topology of the “ring”
    In a network with a ring topology all the nodes are connected by links in unbroken ring (not necessarily a circle) at which data is transmitted. The output of a PC connected to the input of another...