The applicability of DR under a wide range of load conditions allows us to evaluate resource allocation policies not only with respect to mean response time but also with respect to robustness. We have introduced two types of robustness measures: static robustness (robustness against misestimation of load) and dynamic robustness (robustness against fluctuations in load). The study of resource allocation policies for the Beneficiary-Donor model with respect to both mean response time and robustness provides us with lessons that are useful in designing good resource allocation policies.
The study of single-threshold allocation policies, T1 and T2, reveals the tradeoff between low mean response time at an estimated load versus static robustness. Specifically, the T2 policy provides good static robustness, but its mean response time is high. On the other hand, the T1 policy optimized for an estimated load provides low mean response time at the estimated load, but it can lead to instability at higher loads (poor static robustness). While it is possible to tune the T1 policy for the higher load, this would degrade the mean response time at the estimated load.
The tradeoff between low mean response time and static robustness in single-threshold allocation policies motivates us to propose the adaptive dual threshold (ADT) policy. We find that the ADT policy not only provides low mean response time but also excels in static robustness. The ADT policy operates as a T1 policy, but its threshold value is self-adapted to the load. This flexibility allows the ADT policy to provide low mean response time under a range of loads. Hence, when the load is not exactly known, the ADT policy is a better choice than the T1 policy. However, we find that the advantage of the ADT policy over the T1 policy with respect to mean response time (under a given load) is quite small (the improvement is usually less than 5%). This may be surprising, since the optimal policy appears to have infinitely many thresholds.
Surprisingly, our analysis shows that poor static robustness does not necessarily imply poor dynamic robustness. For example, we observe that even when the load is fluctuating, the T1 policy, which lacks static robustness, can often provide low mean response time.