Page 1 of 25
FABRIC Community Visioning Workshop Report
April 15-16, 2020
Executive Summary 2
Motivation 2
Participation 2
Overarching Themes 3
Participant Feedback on FABRIC Architecture 5
Breakout Sessions: Major Highlights 6
IoT/Edge 6
AI/ML 7
Security 8
Network 9
Big Data 9
Lightning Talks 10
Major Takeaways 10
Future Workshops 10
Upcoming Roundtables 11
Appendices 11
Agenda 12
Articles, Papers and Resources Suggested by Workshop Participants 14
Questions and Answers 15
Submitted White Papers 22
FABRIC is supported in part by a Mid-Scale RI-1 NSF award under Grant No. 1935966
Page 2 of 25
Executive Summary
The FABRIC Community Visioning Workshop brought together over 200 active participants to
listen to updates on the FABRIC architecture and schedule as well as provide input on the initial
design and use cases. A combination of talks and break out sessions ensured a rich dialog
between the community and the FABRIC team and, despite the virtual format, participants were
actively engaged. There were a multitude of insightful take-aways that the FABRIC team is
using to finalize the technical architecture, refine our community engagement strategy to include
shorter format sessions, and reach out to additional users. Given the virtual format and robust
online interaction, a substantial amount of Questions and Answers are included as an Appendix.
Motivation
Building a new testbed without the broader CISE and domain science community input on use
cases and technical features prior to production would be counterproductive, and we don’t
believe in a “build it and they will come” approach. We purposefully began our series of FABRIC
events with a Community Visioning Workshop to gather early input and feedback from
experimenters on specific use cases and to obtain partial validation of the proposed initial
design prior to final decisions. Though we began our workshop series with those who are by
and large active testbed users and advocates, our vision for future workshops is to include
experimenters who have not previously relied upon testbeds, but who could benefit from them.
Our goals are to be inclusive of the CISE and domain science communities, and remain
open-minded as to potential use cases while focusing on the initial Science Drivers.
Participation
Participants were originally required to submit a short white paper to attend the in-person
workshop, which was held in conjunction with the NSF Huge Data workshop. There were 51
papers submitted. The final lightning talks were selected from this initial pool. Talks were chosen
to represent a breadth of topics. There were 56 people registered for the original in-person
workshop, and we had planned an estimated total of 85 people in attendance, including
registrants, the FABRIC leadership team, FABRIC science design drivers, NSF Program
Officer(s), and the FABRIC advisory committee.
Due to COVID, the team decided early to move to a virtual workshop for the safety of all
participants. With the move to an abbreviated, online format, registration was opened up. This
change of format resulted in 328 registered for Day 1 and 308 registered for Day 2. Day 1 had
11 Panelists* and 220 total users attend, with 170 as the max concurrent views**. On Day 2, we
had 14 Panelists*, 152 total users attend, with 130 as the max concurrent views** An estimated
147 institutions across 15 countries were represented at the workshop, including: Brazil,
2
Page 3 of 25
Canada, China, Germany, Ireland, Italy, Japan, Netherlands, New Zealand, Republic of Korea,
Slovakia, South Africa, Spain, United Kingdom and the US.
Participants interacted with the speakers by asking questions via Zoom and interacted with each
other and the FABRIC team via a community Slack channel.
*Panelists were members of the FABRIC leadership team and presenters.
**Maximum number of online viewers at the same time during the webinar, excluding Panelists.
Overarching Themes
Data: Although FABRIC funding is for construction of a testbed, data is crucial to every aspect
of its operation, highlighting the importance of addressing open questions about data.
Discussions revolved around four aspects: 1. the challenge of acquiring data for experiments,
the availability and re-use of it by other experimenters (leading to better reproducibility) as well
as its storage; 2. the type of data and metadata that would be made available from the testbed
infrastructure itself; 3. access to various types of “sensitive” data such as real world
packet-traces, data that reflects the networking characteristics of data coming from other
countries/regions, medical-type data and industry data sets; and 4. FABRIC as a vehicle for
experimentation on moving big data.
There were questions on the capabilities of testbed for supporting features related to data
integrity and privacy, especially as experimenters include new types of data for testing
algorithms and test new types of IoT/Edge systems with novel data collection capabilities.
Access to sensitive data such as medical imaging for AI/ML experiments and to “normal”,
realistic, network traffic for security testing was brought up. Notions of what application data can
appropriately and safely be used on a testbed vary considerably, and more clarity is needed
from the project in working with community and data experts. Privacy techniques for the
collection and sharing of such data need to be identified and enabled. Despite the privacy
challenges, security experts believed FABRIC could be used as a platform to help inform the
community on data sharing practices for cyber-detection and attack related data sets.
A topic of particular interest was the ability to generate traffic in FABRIC slices based on real
traffic traces acquired from providers or based on modeling realistic traffic. This will require
further input from the community for acquiring, sanitizing and curating traffic traces from
providers and building mechanisms that would allow replay or generation based on this data
inside the experiments, likely as part of tooling provided to FABRIC by respective experimenter
communities.
Measurement: The need for “fine grained measurements capabilities” came up quite a bit, and
the varying interpretations of the term “measurement” came to light. For example, some want to
gather telemetry data at the optical layer from transponders to expose micro events that could
3
Page 4 of 25
serve as a feedback loop for improving measurement tools. Others asked about the
measurement of the health of the FABRIC infrastructure itself - specifically, the collection of
server statistics such as cores, CPU and memory stats as well as network measurements such
as Netflow statistics and “packet by packet monitoring where needed”. For IoT and real time
sensors at the edge, measurements of jitter and latencies were seen as critical. Varying
opinions on public measurement/visibility into experiments were expressed.
Storage: Storage continues to be a challenge for both experimenters and cyberinfrastructure
providers. While FABRIC plans to have a fair amount of scratch disk and short term storage,
open questions remain about how much infrastructure data and metadata to store, and how to
work with experimenters on obtaining access to data sets (if at all). One participant commented
that experiment reproducibility becomes a concern when the data volumes get large enough
that storage is not easy or possible, as well as when compute/storage becomes transient and/or
distributed. The need for a distributed resilient storage system (both FABRIC-related and in
general) came up, as questions arose about the amount of persistent, permanent storage
provided by FABRIC and the possibility of enabling some storage at a cloud or external provider.
Security and Privacy: Security and privacy feedback fell into one of two categories: 1.
concerns about FABRIC safely interacting with the real world since many experiments won’t
work in a closed network and need access to real traffic; and 2. security of the FABRIC
infrastructure and software, since security assurances are critical to support many use cases
(e.g., medical use cases). European participants surfaced the GDPR aspects of data, with
University of Amsterdam offering to share their document on guidelines around data that may
and may not enter their own testbed and how and where it’s stored.
Collaboration & Community: A myriad of projects expressed interest in working with FABRIC.
These ranged from recent NSF CC* and MidScale awards (such as SAGE NSF#1935984), to
existing testbeds, to up-and-coming International efforts, as well as stand-alone projects. A few
asked the community to think about testbed co-development along with collaborative
experimentation. The recent NSF IRNC solicitation and its testbed area mentioning FABRIC was
discussed as a vehicle for such collaboration. A comment was made that the networking and
systems research community has typically conducted experiments on only one (or perhaps
two/three) testbeds and that experiments (software, scripts to run experiment trials, analysis
software, etc.) are often not portable. Changing this paradigm and being able to conduct
experiments across testbeds (i.e., FABRIC as a testbed of testbeds) was deemed important.
Engaging with other large NSF-funded projects, both in CISE (e.g. the software institutes,
centers of excellence, ...) as well as in the science directorates, especially the Large Facilities,
was encouraged, as was working with projects like SENSE, AutoGOLE, IRNC testbeds,
SEARRCH and a new global testbed under development by the Global Network Advancement
Group (GNA-G). Some referred to FABRIC as a “testbed of testbeds”.
4