In real usage scenarios our performance overhead would be much smaller because a workload’s extra cache allocation does not have to be dropped immediately when a new time window starts and can still provide hits while being gradually replaced by the other workloads,