Isolated Trees in Multi-Tenant Fat Tree Datacenters for In-Network Computing

Modern datacenter applications do not only require computational power but also a high amount of communication between computation nodes. In order to improve the running time of these applications, the network community has been seeking for offloading common computation patterns to the network devices, an emerging field known as In-Network Compute. Artificial intelligence (AI) constitutes an important application obtaining a major performance improvement from in-network compute due to its distributive inherent properties. Most of existing in-network compute schemes are based on the employment of aggregation trees providing a logical tree topology on top of the physical infrastructure for the different aggregation nodes. In this work, we study the fundamental question of implementing such logical trees in an efficient manner. In order to avoid interference among tenants, trees serving various tenants should be edge-disjoint. We study designing such trees in common datacenter topologies such as fat trees. We show that the availability of the trees is highly affected by the mapping of the tenants to hosts of the topology. We describe algorithms for computing mapping that allows the existence of such trees as well as for finding such trees for a given tenant mapping. We conduct experiments using real workloads to examine trees availability based on network utilization and particular mappings.