Section: New Results

Network Modeling

Netloc is a tool for hwloc [1] to find the topology of a supercomputer. For that, it discovers all the networks by exploring them by using tools specifying to the network type. The exploration gives all the machines and all the switches, with all the links between them. We improved netloc by adding the visualization of the topologies discovered. The visualization is dynamic and the user can interact with it, to get some information about the machines, the switches or the link such as the physical address, the hostname or the speed of the link. In order to be able to do optimizations that can be helping process placement, we started to class the different topologies. For now, we only handle Clos networks  [21] and we are able to transform them into fat trees. The categorization in classes permits to have a clean graph and then interact with graph partitioners.

To have a complete tool, we need to handle all major classes of topologies such as meshes, torus or hypercubes. When the graph partitioning will be integrated with tools such as Scotch , we will be able to find a good mapping for the processes of a job. It could also helps the resource scheduling to optimize the resource sharing between jobs. The visualization can be improved by showing the architecture information retrieved by hwloc for each machine. We can complete the visualization by giving more information especially when the original graph was transformed to simplify it, as we did to Clos networks to obtain fat trees.