Priority queues with bursty arrivals of incoming tasks

TL;DR

This paper analyzes how bursty, power-law distributed arrivals of tasks affect priority queue waiting times, deriving a new analytical relationship for the waiting time distribution exponent based on the arrival distribution.

Contribution

It introduces a generating function approach to determine the waiting time exponent for priority queues with power-law task arrivals, extending previous models that assumed Poisson arrivals.

Findings

Derived the exponent  for waiting time distribution based on power-law arrival rates.

Showed that the waiting time exponent is nonuniversal and depends on the power-law exponent b3 of arrivals.

Provided analytical results that differ from previous models assuming Poisson arrivals.

Abstract

Recently increased accessibility of large-scale digital records enables one to monitor human activities such as the interevent time distributions between two consecutive visits to a web portal by a single user, two consecutive emails sent out by a user, two consecutive library loans made by a single individual, etc. Interestingly, those distributions exhibit a universal behavior, $D(\tau)\sim \tau^{-\delta}$, where $\tau$ is the interevent time, and $\delta \simeq 1$ or 3/2. The universal behaviors have been modeled via the waiting-time distribution of a task in the queue operating based on priority; the waiting time follows a power law distribution $P_{\rm w}(\tau)\sim \tau^{-\alpha}$ with either $\alpha=1$ or 3/2 depending on the detail of queuing dynamics. In these models, the number of incoming tasks in a unit time interval has been assumed to follow a Poisson-type distribution. For an email system, however, the number of emails delivered to a mail box in a unit time we measured follows a powerlaw distribution with general exponent $\gamma$. For this case, we obtain analytically the exponent $\alpha$, which is not necessarily 1 or 3/2 and takes nonuniversal values depending on $\gamma$. We develop the generating function formalism to obtain the exponent $\alpha$, which is distinct from the continuous time approximation used in the previous studies.