Overload Control in Star Networks: Comparison of Percent Blocking Throttle and Last-In-First-Out Queue Discipline.

Overload control is investigated for a star network that has a bottleneck at the central node and where jobs probabilistically turn "bad" as a function of sojourn time. We compare two qualitatively different control schemes: (1) a dynamic, percent blocking throttle at the peripheral nodes and a standard FIFO discipline at the central node, versus (2) no throttling at the peripheral nodes and a static LIFO discipline at the central node. The major advantage of (2) over (1) is that it gives higher goodput (throughput times the probability a job is "good") for arrival rates three to four times above capacity, even when the work done by the central processor to load a job into the queue is 10% of the nominal service time, and even assuming an ideal monitor for the throttle that always maintains an optimal control setting. The key advantage of (1) over (2) is that it can maintain high goodput in cases where the system would otherwise be overwhelmed. The results suggest a combination of the two schemes: use a LIFO discipline at the central node to attain high goodput, and use a dynamic throttle at the peripheral nodes to regulate the arrivals to a level optimal for the LIFO discipline. More generally, we propose the following rule of thumb: Rule of Thumb: For overload control in star networks, use both a dynamic throttle (e.g., percent blocking, call gapping or rate control) at the peripheral nodes and an other-than- standard-FIFO queue discipline (e.g., push-out, time-out, priorities, blocking, and/or LIFO ordering) at the central node.