Which topologies used in supercomputer?

Answer 1

Several different topologies are used in supercomputing. A mesh architecture is one of the more common topologies out there. Mesh architectures are particularly well suited to applied simulations. The canonical example is a simulation of ocean weather patterns where each node in the supercomputer will represent (and handle calculations for) a specific patch of ocean, such as a 10x10 square mile sector. In this example, each patch of ocean must communicate information, such as wind speed and temperature to other neighboring nodes and in turn, must update the sector parameters based on information received by its neighbors.

Other example topologies include the master/slave architecture where one node serves to aggregate and process results from all other nodes. There are also ring topologies, tree topologies, and hypercube topologies. By far the fastest communicating, and most expensive topology is the all-to-all topology where every node is connected to every other node. This works fine in superomputers with 20 nodes, but in the largest supercomputers, trying to connect thousands of nodes each to the other thousands of nodes becomes practically impossible. Hence why there are other topologies.