1. PRPv1 Discussion topics
Eli Dart, Network Engineer
ESnet Science Engagement
Lawrence Berkeley National Laboratory
National Research Platform Workshop
Bozeman, MT
August 7, 2017

2. Outline PRPv1 as a routed layer3 network
Layer2 vs. Layer3 Science DMZs
Notes on the theme of Scalability
11/24/2018

3. PRPv1 Diagram Primary substrate is CENIC HPR
Layer3 connectivity between campus DMZs
Connectivity to other sites via Layer3 peering
Usually with CENIC
But, if there is other policy, traffic follows that policy

4. Layer2 vs. Layer3
Connecting two DTNs to each other at Layer2 is quite different from connecting them at Layer3
Layer2 connection:
Determine whether layer2 services exist between DMZs
Scientist typically can’t do this, and so needs an engineer to help
Determine which IP addresses will be used for the connection
Scientist typically can’t do this either – need engineering assistance
Then configure the DTNs with the addresses, VLANs, etc.
Resolve security policy issues between DTNs, if any
Set up the circuit
Can be done by the scientist if they have permission – typically done by network engineers
Debug the circuit if the DTNs can’t talk
Often requires detailed knowledge of the network path and devices
Layer2 circuit often must be modified in order to troubleshoot it
No way for a scientist to do this
11/24/2018

5. Layer2 vs. Layer3, Continued
Connectivity at Layer3 is straightforward
Scientist reasons about their data
Scientist uses the data transfer application
Scientist involves helpdesk/engineering only in case of problems
The difference between the Layer2 and Layer3 case is stark
In the Layer2 case, scientist must involve engineering at the beginning
In the Layer3 case, engineering is only involved in case of problems
The choice of Layer3 for PRPv1 was a wise choice
11/24/2018

6. Scalability
There are many different ways to think about scalability – some are:
Technology, architecture, complexity
People
People scale terribly – primary driver for other scalability needs
Technology - hardware
DTNs, networks, storage, detectors, simulations, etc.
Much of this is governed by the same underlying physics
Semiconductors, Moore’s Law, etc.
Storage currently appears to be more difficult
Technology – software
Command-line tools vs. GUI tools vs. API-driven tools
More data, more files  error handling, automation
11/24/2018

7. Scalability, Continued
Architecture
Science DMZ model makes it easier to get good results more often
Less work for people
Ability to more easily deploy better tools and systems
Complexity
Nobody’s friend
People are bad at keeping track of complex things
Complex processes are hard to get right every time
Simplifying and standardizing allows tasks to be automated
Remember the larger context
Science is changing dramatically
Scientists need a way to keep up, and we need a way to help them
11/24/2018

8. PRPv1 – Good Set Of Choices
The PRP collaboration made a good set of decisions
Routed Layer3 architecture leverages many advantages
Self-service by scientists (just go use the DTNs without reasoning about the network)
Standard troubleshooting tools work
Etc. etc. etc.
Allows the collaboration to focus on the harder stuff
Fixing performance problems
Science engagement
Etc.
As we continue to think about scaling in the NRP context, let’s not forget our early successes
11/24/2018

9. Thanks! Eli Dart Energy Sciences Network (ESnet)
Lawrence Berkeley National Laboratory