Science DMZ Networks: How Different are They Really?

TL;DR

This study quantitatively compares Science DMZ networks to conventional networks, revealing they generally offer lower latency and higher throughput but require careful tuning to maximize benefits.

Contribution

It provides the first comprehensive two-year measurement analysis comparing Science DMZ and campus networks, highlighting their performance differences and tuning requirements.

Findings

Science DMZ networks have lower latency and higher throughput.

Some cases show Science DMZs have longer routes and higher latency.

Proper tuning is essential to fully realize Science DMZ benefits.

Abstract

The Science Demilitarized Zone (Science DMZ) is a network environment optimized for scientific applications. A Science DMZ provides an environment mostly free from competing traffic flows and complex security middleware such as firewalls or intrusion detection systems that often impede data transfer performance. The Science DMZ model provides a reference set of network design patterns, tuned hosts and protocol stacks dedicated to large data transfers and streamlined security postures that significantly improve data transfer performance, accelerating scientific collaborations and discovery. Over the past decade, many universities and organizations have adopted this model for their research computing. Despite becoming increasingly popular, there is a lack of quantitative studies comparing such a specialized network to conventional production networks regarding network characteristics and data transfer performance. We strive to answer the following research questions in this study: Does a Science DMZ exhibit significantly different behavior than a general-purpose campus network? Does it improve application performance compared to such general-purpose networks? Through a two-year-long quantitative network measurement study, we find that a Science DMZ exhibits lower latency, higher throughput, and lower jitter behaviors. However, we also see several non-intuitive results. For example, a DMZ may take a longer route to external destinations and experience higher latency than the campus network. While the DMZ model benefits researchers, the benefits are not automatic - careful network tuning based on specific use cases is required to realize the full potential of such infrastructure.