Benchmarking the Ethernet-Federated Datacenter

Currently, datacenters (DC) are the largest closed-loop systems in IT, growing toward multi-million node Clouds. Benchmarking, optimizing and managing such large systems is an active area of research and an open challenge, driven by the trend toward stricter service level agreements (higher performance) and lower power. To evolve the DC design and management from today’s ad-hoc solutions to a rigorous discipline, we argue the need for benchmarking based on advanced monitoring and modeling methods. In addition to a positional report of work in progress—our primary target—here we contribute a three-pronged DC benchmarking proposal. Our approach combines resource monitoring and workload characterization with DC performance modeling, both by simulation and by analytical means. While the monitoring and transaction tracing module is still at an early stage, our modeling methods have already been successfully tested in the two study cases. They build on the recent progress made in key independent fields: Ethernet DCB monitoring on L2, ARM-based transaction tracing, largescale High-Performance Computing (HPC) simulations, and analytical modeling of dynamic distributed systems.