If looks like DA's multi-size still needs some work since there are 8 different sizes and your devices are grabbing work from #5. #1 is 16 times smaller than the one you are getting right now. Looks like I need to dig into the code and figure out why it is so stupid that it sends really large files to really slow clients. Could be bad benchmarks, bad code, or a combination. It doesn't help that all the benchmarks use flops and Collatz only uses integers. That makes the estimates off by a factor of 8 to start with but it takes client code changes to fix it. There's nothing I can change on the server since the client sends only floating point data and no integer data.