what you are missing is that there are two sets of graphs, since there are two ways of measuring performance: number of calls (scales as n^2 with no backoff) an

what you are missing is that there are two sets of graphs, since there are two ways of measuring performance: number of calls (scales as n^2 with no backoff) and time to completion (scales linearly with no backoff) and an unjittered backoff will improve the number of calls while making the time to completion hugely worse (exponential?) - but jittered backoff looks to drop the number of calls substantially and also scales almost linearly, while still not doing too badly on time to completion.