Bounded Mean-Delay Throughput and Non-Starvation Conditions in Aloha Network

TL;DR

This paper analyzes how different exponential backoff factors in slotted Aloha networks impact mean delay and starvation, revealing optimal settings for bounded delay and unifying conditions for delay and fairness.

Contribution

It introduces a quantitative definition of starvation, links delay and starvation conditions, and identifies optimal backoff parameters for bounded mean delay in Aloha networks.

Findings

Optimal backoff factor for bounded delay is approximately 1.3757.

Sustainable throughput drops significantly when delay bounds are enforced.

Starvation can cause non-converging performance in simulations.

Abstract

This paper considers the requirements to ensure bounded mean queuing delay and non-starvation in a slotted Aloha network operating the exponential backoff protocol. It is well-known that the maximum possible throughput of a slotted Aloha system with a large number of nodes is 1/e = 0.3679. Indeed, a saturation throughput of 1/e can be achieved with an exponential backoff factor of r = e/(e-1)=1.5820. The binary backoff factor of r = 2 is assumed in the majority of prior work, and in many practical multiple-access networks such as the Ethernet and WiFi. For slotted Aloha, the saturation throughput 0.3466 for r = 2 is reasonably close to the maximum of 1/e, and one could hardly raise objection to adopting r = 2 in the system. However, this paper shows that if mean queuing delay is to be bounded, then the sustainable throughput when r = 2 is only 0.2158, a drastic 41% drop from 1/e . Fortunately, the optimal setting of r = 1.3757 under the bounded mean-delay requirement allows us to achieve sustainable throughput of 0.3545, a penalty of only less than 4% relative to 1/e. A general conclusion is that the value of r may significantly affect the queuing delay performance. Besides analyzing mean queuing delay, this paper also delves into the phenomenon of starvation, wherein some nodes are deprived of service for an extended period of time while other nodes hog the system. Specifically, we propose a quantitative definition for starvation and show that the conditions to guarantee bounded mean delay and non-starved operation are one of the same, thus uniting these two notions. Finally, we show that when mean delay is large and starvation occurs, the performance results obtained from simulation experiments may not converge. A quantitative discussion of this issue is provided in this paper.