Unix Memory Allocations are Not Poisson

TL;DR

This paper demonstrates that Unix memory request arrivals are not Poisson over short timescales due to burstiness and high variance, challenging traditional modeling assumptions and indicating the need for alternative models.

Contribution

It provides empirical evidence that Unix page requests are not Poisson, highlighting bursty behavior and the inadequacy of traditional stochastic models for short-term analysis.

Findings

Memory requests are bursty and non-Poisson over short timescales.

Page request distributions are similar regardless of request origin.

Requests are not self-similar despite high variance and burstiness.

Abstract

In multitasking operating systems, requests for free memory are traditionally modeled as a stochastic counting process with independent, exponentially-distributed interarrival times because of the analytic simplicity such Poisson models afford. We analyze the distribution of several million unix page commits to show that although this approach could be valid over relatively long timespans, the behavior of the arrival process over shorter periods is decidedly not Poisson. We find that this result holds regardless of the originator of the request: unlike network packets, there is little difference between system- and user-level page-request distributions. We believe this to be due to the bursty nature of page allocations, which tend to occur in either small or extremely large increments. Burstiness and persistent variance have recently been found in self-similar processes in computer networks, but we show that although page commits are both bursty and possess high variance over long timescales, they are probably not self-similar. These results suggest that altogether different models are needed for fine-grained analysis of memory systems, an important consideration not only for understanding behavior but also for the design of online control systems.