What about running the benchmark multiple times on instances of the same type? I get that it would be noisy but lots of workloads run on shared instances so it’s a useful measuring stick in that way.