Overview

So you want to operate a Merge testbed facility?

You’ve come to the right place! This page overviews the core components and network architecture of a typical Merge facility. Once you understand the basic design, head on over to these pages for detailed instructions on:

High Level Design

A testbed facility houses the equipment that materializes experiments. A facility typically consists of at least the following assets:

  • A set of testbed nodes that host experiment nodes. Testbed nodes either operate as hypervisors (supporting many materialization nodes concurrently through virtual machine technology) or as bare-metal machines (supporting one materialization node at a time)
  • Infrastructure servers that host testbed services such as DHCP, DNS, and node imaging for experiment materializations
  • Storage servers that host data stores for the testbed’s core services as well as provide mass storage for experiment use
  • An operations server that functions as central point of command and control for a human operator
  • A set of physical networks:
    • a “management” network through which an operator can connect to and control facility assets
    • an “infrastructure” network which supports the aforementioned services (DHCP, DNS, etc.)
    • an “experiment” network on which virtual networks for experiment materializations are embedded

Canonical Facility Architecture

The following is a generalized view of how a Merge testbed facility is typically configured:

Testbed Nodes & Hypervisors

These machines host experiment nodes for materializations as described by user experiment models. Experiment nodes can either be deployed as virtual machines on top of the Merge hypervisor stack, or can be deployed on a bare-metal server. Each individual testbed node in the facility can alternate between bare-metal operation and hypervisor operation throughout its lifetime. Any transitioning of modes that occurs is entirely automated by Merge software.

Network Emulators

These are special purpose machines that precisely control the performance characteristics of network links in user materializations. When a user requests precise link emulation in a materialization, the embedding for that link is done such that the link traverses a network emulation server in the facility. Currently, emulation servers run a modified version of the Fastclick system to control capacity, delay, and loss rates on emulated links. Because link emulation can be a resource intensive workload, a recommended architectural model is to include a number of dedicated emulation servers that do not host any user experiment nodes or other testbed services.

Infrastructure Servers

Infrastructure servers host “infrapods”, which are per-materialization “pods” that provide services to the materialization, including DHCP, DNS, and an experiment VPN access point for secure external connectivity. A separate infrapod is allocated for each materialization, which means that these services are confined to the namespace of a single materialization.

Storage Servers

Storage servers host the etcd and MinIO data stores that support the core services of the facility, as well as mass storage for experiment use.

Networks

The canonical facility architecture includes a distinct infrastructure network (“infranet”) and experiment network (“xpnet”). The infranet and xpnet serve conceptually distinct purposes and are often deployed in entirely disjoint network fabrics, though it is possible for these networks to share switches.

The infranet supports testbed services such as DHCP, DNS, node imaging, and mass storage. These services support efficient testbed operation but are not designed to transit user experiment traffic, which is instead supported by the xpnet. The xpnet is the underlying substrate over which network links are embedded to provide the desired topological connectivity and link performance characteristics requested by the materialization.

Data Stores

Etcd MinIO


Last modified June 10, 2022: add a bunch of facility content (438cf43)