An Overview on Network Hardware of Cluster Computing
A model, but not necessarily the only possible model, of parallel processing with cluster systems involves each of the cluster nodes performing one or more tasks on local data and then exchanging the computed results with other nodes within the cluster. Networks make this possible.
They provide physical channels between nodes by which data are transported and logical protocols that govern the flow and interpretation of the transferred data. Networks are employed in a broad range of integrated systems from the Internet spanning the globe requiring possibly as much as a hundred milliseconds for a message packet to reach its destination to a data bus internal to a computer integrating its various components supporting data transfers in 100 Nano-sec or less, a ratio of a million in network latency.
Networks for commodity clusters fall in between with the initial use of Ethernet exhibiting on average approximately 100 nanosecond latency falling in the middle (logarithmically speaking). Network technology determines the potential value of cluster computing. Its principal properties are bandwidth, latency, scale, and cost. Bandwidth imposes an upper bound on the amount of data that can be transferred in unit time (e.g., Mbps, Gbps). Latency is the amount of time it takes for a message packet to transit the diameter of a system measured in microseconds. Cost is usually considered as the percentage of the total price of the hardware system. Scale is the largest number of nodes that a network can connect effectively. Together, they establish a cluster’s capability, applicability, and user accessibility.
Different applications exhibit varying global data access patterns that may be suitable for some networks rather than others. Higher bandwidth networks ordinarily will have greater generality of application than those networks of lower bandwidth. Similarly, for applications using short messages or involving frequent global synchronization, lower latency networks will be more general purpose than high-latency networks. But superior behavioral properties often come at additional cost that may preclude their use in many environments, where cost is a significant factor in the choice to employ clusters in the first place. Thus the selection of a specific network is dependent on how the cluster is to be used and by whom.
Network Cluster and Switches:
A cluster network includes NICs that connect the cluster node to the network, transport layer links to carry the data, and switches that route the data through the network. NICs move data from message buffers filled by the node processor to signal packets sent out to the transport layer performing a number of translation functions on the data in the process. The data links may comprise one or more parallel channels and may be implemented with metal coaxial cable or optical fiber (advanced development of free-space optical networks is under way). Switches accept messages at their multiple input ports, determine their required routing, switch as many as possible simultaneously sending them out the appropriate output ports, and arbitrating where contention for shared resources (ports, channels) occurs. The earliest Beowulf-class systems used low-cost hubs, rather than the more expensive switches, but these permitted only one transfer to occur at a time on the entire network. Switches deliver much closer to the peak bi-section bandwidth of the network as they isolate separate disjoint paths from each other.
Charles West is a professional content writer and blogger since last 2 years. I have writing expertise in technology and certification topics specially. I love to share that recently I passed my 350-030 exam from SAS institute and 642-994 exam from VMware. Thanks for reading.