Tutorial: Maximizing Performance and Network Utility with a Science DMZ

National Science Foundation Award #1826994
Jason Zurawski
zurawski@es.net
ESnet / Lawrence Berkeley National Laboratory
http://bit.ly/2vx2HiQ
Maximizing Performance and Network
Utility with a Science DMZ
Globusworld 2019
Chicago, IL
May 2nd, 2019
https://epoc.global

Outline
• Motivation
• TCP Basics
• DMZ Design Patterns
• Integration w/ Science
• Conclusions
2 – ESnet Science Engagement (engage@es.net) - 5/3/19

The Science DMZ in 1 Slide
Consists of four key components, all required:
• “Friction free” network path
• Highly capable network devices (wire-speed, deep queues)
• Virtual circuit connectivity option
• Security policy and enforcement specific to science workflows
• Located at or near site perimeter if possible
• Dedicated, high-performance Data Transfer Nodes (DTNs)
• Hardware, operating system, libraries all optimized for transfer
• Includes optimized data transfer tools such as Globus and GridFTP
• Performance measurement/test node
• perfSONAR
• Once it’s up, users often need training – Engagement is key
Details at http://fasterdata.es.net/science-dmz/
© 2013 Wikipedia
© 2013 Globus

Michael Sinatra (ESnet)
The Science DMZ has two main purposes:
1.A security architecture which allows for better segmentation of risks, and
more granular application of controls to those segmented risks. In the case
of the Science DMZ, the goal is to limit the segmented risk profile to that of
a dedicated data transfer host.
2.An attempt to streamline tuning and troubleshooting by removing
degrees-of-freedom (from both a reliability and performance perspective)
from the active data transfer path.

The Science DMZ in 1 Picture

Outline
• Motivation
• TCP Basics
• Conclusions

Science DMZ Background
• The data mobility performance requirements for data intensive science are beyond
what can typically be achieved using traditional methods
• Default host configurations (TCP, filesystems, NICs)
• Converged network architectures designed for commodity traffic
• Conventional security tools and policies
• Legacy data transfer tools (e.g. SCP)
• Wait-for-trouble-ticket operational models for network performance

Science DMZ Background
• The Science DMZ model describes a performance-based approach
• Dedicated infrastructure for wide-area data transfer
• Well-configured data transfer hosts with modern tools
• Capable network devices
• High-performance data path which does not traverse commodity LAN
• Proactive operational models that enable performance
• Well-deployed test and measurement tools (perfSONAR)
• Periodic testing to locate issues instead of waiting for users to complain
• Security posture well-matched to high-performance science applications

TCP – Ubiquitous and Fragile
• Networks provide connectivity between hosts – how do hosts
see the network?
• From an application’s perspective, the interface to “the other end” is a socket
• Communication is between applications – mostly over TCP
• TCP – the fragile workhorse
• TCP is (for very good reasons) timid – packet loss is interpreted as congestion
• Packet loss in conjunction with latency is a performance killer
• Like it or not, TCP is used for the vast majority of data transfer applications (more
than 95% of ESnet traffic is TCP)

Packet/Data Loss?!
• “Wait a minute! I thought TCP was a reliable protocol? What do you mean ‘packet loss’,
where is the data going!?”
• The data isn’t loss forever, its dropped somewhere on the path.
• Usually by a device without enough buffer space to accept it, or by someone who thinks the data is
corrupted and they won’t send it
• Once its dropped, we have a way of knowing its been dropped.
• TCP is reliable, each end is keeping track of what was sent, and what was received.
• If something goes missing, its resent. Resending is what takes the time, and causes the slowdown.
• TCP is able to reliably and transparently recover from packet loss by retransmitting any/all lost
packets
• This is how it provides a reliable data transfer services to the applications
• The reliability mechanisms dramatically reduce performance when they are exercised
• We want to eliminate the causes of packet loss – so that we don’t need to test out the (slow)
way that TCP can recover.
• What is the impact of that recovery?

A small amount of packet loss makes a huge
difference in TCP performance
Metro Area
Local
(LAN)
Regional
Continental
International
Measured (TCP Reno) Measured (HTCP) Theoretical (TCP Reno) Measured (no loss)
With loss, high performance
beyond metro distances is
essentially impossible

How Do We Accommodate TCP?
• High-performance wide area TCP flows must get loss-
free service
• Sufficient bandwidth to avoid congestion
• Deep enough buffers in routers and switches to handle
bursts
• Especially true for long-distance flows due to packet behavior
• No, this isn’t buffer bloat
• Equally important – the infrastructure must be verifiable so that clean service can be provided
– Stuff breaks
• Hardware, software, optics, bugs, …
• How do we deal with it in a production environment?
– Must be able to prove a network device or path is functioning correctly
• Regular active tests should be run - perfSONAR
– Small footprint is a huge win
• Fewer the number of devices = easier to locate the source of packet loss
© 2013 icanhascheezburger.com

Outline
• Motivation
• TCP Basics
• Conclusions

Science DMZ Takes Many Forms
• There are a lot of ways to combine these things – it all depends on
what you need to do
• Small installation for a project or two
• Facility inside a larger institution
• Institutional capability serving multiple departments/divisions
• Science capability that consumes a majority of the infrastructure
• Some of these are straightforward, others are less obvious
• Key point of concentration: eliminate sources of packet loss / packet
friction

Ad Hoc DTN Deployment
10GE10G
Site Border
Router
WAN
Building or Wiring
Closet Switch/Router
Perimeter Firewall
Site / Campus
LAN
High performance
Data Transfer Node
with high-speed storage
Global security policy
mixes rules for science
and business traffic
DTN traffic subject to firewall
limitations
perfSONAR
Test and measurement
not aligned with data
resource placement
DTN traffic subject to limitations of
general-purpose networking
equipment/config
Note: Site border
router and perimeter
firewall are often the
same device
Conflicting requirements
result in performance
compromises

Legacy Method: Ad Hoc DTN Deployment
• This is often what gets tried first
• Data transfer node deployed where the owner has space
• This is often the easiest thing to do at the time
• Straightforward to turn on, hard to achieve performance
• If lucky, perfSONAR is at the border
– This is a good start
– Need a second one next to the DTN
• Entire LAN path has to be sized for data flows
• Entire LAN path is part of any troubleshooting exercise
• This usually fails to provide the necessary performance.

A better approach: simple Science DMZ
10GE
10GE
10GE
10GE
10G
Border Router
WAN
Science DMZ
Switch/Router
Enterprise Border
Router/Firewall
Site / Campus
LAN
High performance
Data Transfer Node
Per-service
security policy
control points
Clean,
High-bandwidth
WAN path
Site / Campus
access to Science
DMZ resources
perfSONAR
perfSONAR
perfSONAR

Small-scale Science DMZ Deployment
• Add-on to existing network infrastructure
• All that is required is a port on the border router
• Small footprint, pre-production commitment
• Easy to experiment with components and technologies
• DTN prototyping
• perfSONAR testing
• Limited scope makes security policy exceptions easy
• Only allow traffic from partners
• Add-on to production infrastructure – lower risk

Prototype Science DMZ Data Path
10GE
10GE
10GE
10GE
10G
Border Router
WAN
Science DMZ
Switch/Router
Enterprise Border
Router/Firewall
Site / Campus
LAN
High performance
Data Transfer Node
Per-service
security policy
control points
Clean,
High-bandwidth
WAN path
Site / Campus
access to Science
DMZ resources
perfSONAR
perfSONAR
High Latency WAN Path
Low Latency LAN Path

Support For Multiple Projects
• Science DMZ architecture allows multiple projects to put DTNs in place
• Modular architecture
• Centralized location for data servers
• This may or may not work well depending on institutional politics
• Issues such as physical security can make this a non-starter
• On the other hand, some shops already have service models in place
• On balance, this can provide a cost savings – it depends
• Central support for data servers vs. carrying data flows
• How far do the data flows have to go?

Multiple Projects
10GE
10GE
10GE
10G
Border Router
WAN
Science DMZ
Switch/Router
Enterprise Border
Router/Firewall
Site / Campus
LAN
Project A DTN
Per-project
security policy
control points
Clean,
High-bandwidth
WAN path
Site / Campus
access to Science
DMZ resources
perfSONAR
perfSONAR
Project B DTN
Project C DTN

Supercomputer Center Deployment
• High-performance networking is assumed in this environment
• Data flows between systems, between systems and storage, wide area, etc.
• Global filesystem often ties resources together
• Portions of this may not run over Ethernet (e.g. IB)
• Implications for Data Transfer Nodes
• “Science DMZ” may not look like a discrete entity here
• By the time you get through interconnecting all the resources, you end up with most of the
network in the Science DMZ
• This is as it should be – the point is appropriate deployment of tools, configuration, policy
control, etc.
• Office networks can look like an afterthought, but they aren’t
• Deployed with appropriate security controls
• Office infrastructure need not be sized for science traffic

Supercomputer Center Data Path
Virtual
Circuit
Routed
Border Router
WAN
Core
Switch/Router
Firewall
Offices
perfSONAR
perfSONAR
perfSONAR
Supercomputer
Parallel Filesystem
Front end
switch
Data Transfer
Nodes
Front end
switch
High Latency VC Path

Distributed Science DMZ
• Fiber-rich environment enables distributed Science DMZ
• No need to accommodate all equipment in one location
• Allows the deployment of institutional science service
• WAN services arrive at the site in the normal way
• Dark fiber distributes connectivity to Science DMZ services throughout the site
• Departments with their own networking groups can manage their own local Science DMZ
infrastructure
• Facilities or buildings can be served without building up the business network to support
those flows
• Security is potentially more complex
• Remote infrastructure must be monitored
• Solutions depend on relationships with security groups

Distributed Science DMZ – Dark Fiber
Dark
Fiber
Dark
Fiber
10GE
Dark
Fiber
10GE
10GE
10G
Border Router
WAN
Science DMZ
Switch/Router
Enterprise Border
Router/Firewall
Site / Campus
LAN
Per-project
security policy
control points
Clean,
High-bandwidth
WAN path
Site / Campus
access to Science
DMZ resources
perfSONAR
perfSONAR
Project A DTN
(remote)
Project B DTN
(remote)
Project C DTN
(remote)

Common Threads• Two common threads exist in all these examples
• Accommodation of TCP
• Wide area portion of data transfers traverses purpose-built path
• High performance devices that don’t drop packets
• Ability to test and verify
• When problems arise (and they always will), they can be solved if the infrastructure is
built correctly
• Small device count makes it easier to find issues
• Multiple test and measurement hosts provide multiple views of the data path
• perfSONAR nodes at the site and in the WAN
• perfSONAR nodes at the remote site

Equipment – Routers and Switches
• Requirements for Science DMZ gear are different
• No need to go for the kitchen sink list of services
• A Science DMZ box only needs to do a few things, but do them well
• Support for the latest LAN integration magic with your Windows Active Directory
environment is probably not super-important
• A clean architecture is important
• How fast can a single flow go?
• Are there any components that go slower than interface wire speed?
• There is a temptation to go cheap
• Hey, it only needs to do a few things, right?
• You typically don’t get what you don’t pay for
• (You sometimes don’t get what you pay for either)

Common Circumstance:
Multiple Ingress Data Flows, Common Egress
Background
traffic or
competing bursts
DTN traffic with
wire-speed
bursts
10GE
10GE
10GE
Hosts will typically send packets at the speed of their
interface (1G, 10G, etc.)
• Instantaneous rate, not average rate
• If TCP has window available and data to send, host
sends until there is either no data or no window
Hosts moving big data (e.g. DTNs) can send large bursts
of back-to-back packets
• This is true even if the average rate as measured
over seconds is slower (e.g. 4Gbps)
• On microsecond time scales, there is often
congestion
• Router or switch must queue packets or drop them

Rant Ahead
N.B. You are entering into rant territory on the matter of switch buffering. If you are going to take
away anything from the next section:
1. Under buffered network devices are the single greatest threat to data intensive use of the
network. You can make hosts, operating systems, and application choices perform better
for ‘free’, it will cost $$$ to fix a crappy switch or router
2. You will be steered toward non-optimal choices when you talk with the vendor community
because they don’t understand simple math (but by the end of this, you will).
3. A 1U/2U data center/racklan network device should never be in the path of your data
intensive network use case.
4. Non-passive (e.g. stateful) security devices are the same for buffering, and are actually
worse due to the processing overhead.
5. Anytime you jump around the OSI stack – add friction (e.g. routing when you don’t need to,
application layer inspection, etc.)

All About That Buffer (No Cut Through)

• Data arrives
from multiple
sources
• Buffers have a finite amount of memory
• Some have this per interface
• Others may have access to a shared
memory region with other interfaces
• The processing engine will:
• Extract each packet/frame from the
queues
• Pull off header information to see where
the destination should be
• Move the packet/frame to the correct
output queue
• Additional delay is possible as the
queues physically write the packet
to the transport medium (e.g.
optical interface, copper interface)

• The Bandwidth Delay Product
• The amount of “in flight” data for a TCP connection (BDP = bandwidth * round
trip time)
• Example: 10Gb/s cross country, ~100ms
• 10,000,000,000 b/s * .1 s = 1,000,000,000 bits
• 1,000,000,000 / 8 = 125,000,000 bytes
• 125,000,000 bytes / (1024*1024) ~ 125MB
• Ignore the math aspect: its making sure there is memory to catch and send
packets
• As the speed increases, there are more packets.
• If there is not memory, we drop them, and that makes TCP sad.

• Buffering isn’t as important on the LAN (this is why you are normally
pressured to buy ‘cut through’ devices)
• Change the math to make the Latency 1ms = 1.25MB
• ‘Cut through’ and low latency switches are designed for the data center, and
can handle typical data center loads that don’t require buffering (e.g. same to
same speeds, destinations within the broadcast domain)
• Buffering *MATTERS* for WAN Transfers
• Placing something with inadequate buffering in the path reduces the buffer
for the entire path. E.g. if you have an expectation of 10Gbps over 100ms –
don’t place a 12MB buffer anywhere in there – your reality is now ~10x less
than it was before (e.g. 10Gbps @ 10ms, or 1Gbps @ 100ms)
• Ignore the math aspect, its really just about making sure there is memory to catch
packets. As the speed increases, there are more packets. If there is not memory,
we drop them, and that makes TCP sad.
• Memory on hosts, and network gear

[ ID] Interval Transfer Bandwidth Retr Cwnd
[ 14] 0.00-1.00 sec 524 KBytes 4.29 Mbits/sec 0 157 KBytes
[ 14] 1.00-2.00 sec 3.31 MBytes 27.8 Mbits/sec 0 979 KBytes
[ 14] 2.00-3.00 sec 17.7 MBytes 148 Mbits/sec 0 5.36 MBytes
[ 14] 4.00-5.00 sec 11.2 MBytes 94.4 Mbits/sec 0 1.88 MBytes
• What does this “look” like to a data transfer? Consider the test of iperf below
– See TCP ‘ramp up’ and slowly increase the window
– When something in the path has no more space for packets – a drop occurs. TCP will eventually react to the lost packet,
and ‘back off’
– In the example, this first occurs when we reach a buffer of around 6-8MB. Then after backoff the window is halved a
couple of times
– This happens again later – at a slightly higher buffer limit. This could be because there was cross traffic the first time, etc.

Decoding Specifications
• “The buffering behaviors of the switches and their
operating system, such as behavior under memory stress,
are typically proprietary information and not well
documented” http://www.measurementlab.net/blog/traffic-microbursts-and-their-effect-on-internet-measurement/
• “Even if you know how much packet buffer is in the
switch, assumptions on how it is deployed that are not
backed up by testing can lead to unhappiness. What we
like to say is that is the job of the network engineers to
move bottlenecks around.”
• Jim Warner
• http://people.ucsc.edu/~warner/buffer.html

• So lets say the spec sheet says this:
• What does ‘module’ mean?
• Typically this means the amount of memory for the entire switch
(if it’s a single unit) or a blade (if the chassis supports more than
one.
• BUT … this memory can be allocated in a number of different
ways:
• Shared between all ports
• Dedicated (smaller) amounts per-port
• Shared between ASICS, which control a bank of ports

•Consider this architecture
• 48 Ports
• 12 ASICS
• 4 Ports per ASIC
• 72MB total
• 6MB per ASIC
• If all ports are in use – expect that each
port has access to 1.5MB. If only one is
in use, it can use 6MB
• Additional memory is often available
in a ‘burst buffer’ in the fabric
ASIC = application-specific integrated circuit, think 'small routing engine'

• Recall:
https://www.switch.ch/network/tools/tcp_throughput/
• What does 6MB get you?
• 1Gbps @ <= 48ms (e.g. ½ needed for coast-to-coast)
• 10Gbps @ <= 4.8ms (e.g. metro area)
• What does 1.5MB get you?
• 1Gbps @ <= 12ms (e.g. regional area)
• 10Gbps @ <= 1.2ms (e.g. data center [or more accurately, rack or
row)]
• In either case – remember this assumes you are the only thing using that memory
… congestion is a more likely reality

Outline
• Motivation
• TCP Basics
• Conclusions

NCAR RDA Data Portal
• Let’s say I have a nice compute allocation at NERSC – climate science
• Let’s say I need some data from NCAR for my project
• https://rda.ucar.edu/
• Data sets (there are many more, but these are two):
• https://rda.ucar.edu/datasets/ds199.1/ (1.5TB)
• https://rda.ucar.edu/datasets/ds313.0/ (430GB)
• Download to NERSC (could also do ALCF or NCSA or OLCF)

NCAR RDA Performance to DOE HPC Facilities
13.9 Gbps 16.6 Gbps 11.9 Gbps
DTN
nersc#dtn
NERSC
DTN
olcf#dtn_atlas
OLCF
DTN
alcf#dtn_mira
ALCF
DTN
NCAR RDA
rda#datashare
• 1.5TB data set
• 1121 files

Next Steps – Building On The Science DMZ
• Enhanced cyberinfrastructure substrate now exists
• Wide area networks (ESnet, GEANT, Internet2, Regionals)
• Science DMZs connected to those networks
• DTNs in the Science DMZs
• What does the scientist see?
• Scientist sees a science application
• Data transfer
• Data portal
• Data analysis
• Science applications are the user interface to networks and DMZs
• Large-scale data-intensive science requires that we build larger structures
on top of those components

HPC Facilities: The Petascale DTN Project
• Built on top of the Science DMZ model
• Effort to improve data transfer performance between the DOE ASCR HPC
facilities at ANL, LBNL, and ORNL, and also NCSA.
• Multiple current and future science projects need to transfer data between HPC facilities
• Performance goal is 15 gigabits per second (equivalent to 1PB/week)
• Realize performance goal for routine Globus transfers without special tuning
• Reference data set is 4.4TB of cosmology simulation data

DTN Cluster Performance – HPC Facilities
(2016)
6.8 Gbps
7.6 Gbps
6.9 Gbps
13.3 Gbps
6.0 Gbps
6.7 Gbps
11.1 Gbps
10.5 Gbps
7.3 Gbps
10.0 Gbps
13.4 Gbps
8.2 Gbps
DTN
DTN
DTN
DTN
Data set: L380
Files: 19260
Directories: 211
Other files: 0
Total bytes: 4442781786482 (4.4T bytes)
Smallest file: 0 bytes (0 bytes)
Largest file: 11313896248 bytes (11G bytes)
Size distribution:
1 - 10 bytes: 7 files
100 - 1K bytes: 59 files
1K - 10K bytes: 3170 files
100K - 1M bytes: 2817 files
1M - 10M bytes: 3901 files
100M - 1G bytes: 2295 files
1G - 10G bytes: 1647 files
Petascale DTN Project
March 2016
L380 Data Set
NERSC DTN cluster
Globus endpoint: nersc#dtn
ALCF DTN cluster
Globus endpoint: alcf#dtn_mira
OLCF DTN cluster
Globus endpoint: olcf#dtn_atlas
NCSA DTN cluster
Globus endpoint: ncsa#BlueWaters

DTN Cluster Performance – HPC
Facilities (2017)
21.2/22.6/24.5
Gbps
23.1/33.7/39.7
Gbps
26.7/34.7/39.9
Gbps
33.2/43.4/50.3
Gbps
35.9/39.0/40.7
Gbps
29.9/33.1/35.5
Gbps
34.6/47.5/56.8
Gbps
44.1/46.8/48.4
Gbps
41.0/42.2/43.9
Gbps
33.0/35.0/37.8
Gbps
43.0/50.0/56.3
Gbps
55.4/56.7/57.4
Gbps
DTN
DTN
DTN
DTN
NERSC DTN cluster
Filesystem: /project
Data set: L380
Files: 19260
Directories: 211
Other files: 0
Size distribution:
November 2017
L380 Data Set
Gigabits per second
(min/avg/max), three
transfers
ALCF DTN cluster
Filesystem: /projects
OLCF DTN cluster
Filesystem: atlas2
NCSA DTN cluster
Filesystem: /scratch

Science Data Portals
• Large repositories of scientific data
• Climate data
• Sky surveys (astronomy, cosmology)
• Many others
• Data search, browsing, access
• Many scientific data portals were designed 15+ years ago
• Single-web-server design
• Data browse/search, data access, user awareness all in a single
system
• All the data goes through the portal server
• In many cases by design
• E.g. embargo before publication (enforce access control)

Legacy Portal Design
10GE
Border Router
WAN
Firewall
Enterprise
perfSONAR
perfSONAR
Filesystem
(data store)
10GE
Portal
Server
Browsing path
Query path
Data path
Portal server applications:
· web server
· search
· database
· authentication
· data service
• Very difficult to improve performance
without architectural change
– Software components all tangled
together
– Difficult to put the whole portal in a
Science DMZ because of security
– Even if you could put it in a DMZ, many
components aren’t scalable
• What does architectural change mean?

Architectural Examination of Data Portals
• Common data portal functions (most portals have these)
• Search/query/discovery
• Data download method for data access
• GUI for browsing by humans
• API for machine access – ideally incorporates search/query + download
• Performance pain is primarily in the data handling piece
• Rapid increase in data scale eclipsed legacy software stack capabilities
• Portal servers often stuck in enterprise network
• Can we “disassemble” the portal and put the pieces back together better?
• Use Science DMZ as a platform for the data piece
• Avoid placing complex software in the Science DMZ

Legacy Portal Design
10GE
Border Router
WAN
Firewall
Enterprise
perfSONAR
perfSONAR
Filesystem
(data store)
10GE
Portal
Server
Browsing path
Query path
Data path
· web server
· search
· database
· authentication
· data service

Next-Generation Portal Leverages Science
DMZ
10GE10GE
10GE
10GE
Border Router
WAN
Science DMZ
Switch/Router
Firewall
Enterprise
perfSONAR
perfSONAR
10GE
10GE
10GE
10GE
DTN
DTN
API DTNs
(data access governed
by portal)
DTN
DTN
perfSONAR
Filesystem
(data store)
10GE
Portal
Server
Browsing path
Query path
· web server
· search
· database
· authentication
Data Path
Data Transfer Path
Portal Query/Browse Path

Put The Data On Dedicated Infrastructure
• We have separated the data handling from the portal logic
• Portal is still its normal self, but enhanced
• Portal GUI, database, search, etc. all function as they did before
• Query returns pointers to data objects in the Science DMZ
• Portal is now freed from ties to the data servers (run it on Amazon if you want!)
• Data handling is separate, and scalable
• High-performance DTNs in the Science DMZ
• Scale as much as you need to without modifying the portal software
• Outsource data handling to computing centers
• Computing centers are set up for large-scale data
• Let them handle the large-scale data, and let the portal do the orchestration of data
placement

JGI Data Portal

Reasons To Scale Data Portals
• Some reasons are obvious
• Increase in size of data objects (MB à GB à 100s of GB)
• Number of data objects (many thousands per data set)
• Other reasons are paradigm shifts
• Modern data analysis on HPC can use a *lot* of data
• Today’s HPC facilities are far more capable than in the past
• Retrofit / rebuilt data portals and data repositories
• Significant wins from increased data analysis

Science at Scale: Genomics

Science at Scale: Climate

They Can Use All The Data
• Groups like these need large data sets
• Much of the data in their field is behind legacy portals
• Significant human effort to retrieve what they need
• Legacy systems perform poorly, especially at scale
• Legacy data portals are a product of their time
• We now live in the future from the perspective of those designs
• Current systems far exceed the capabilities available 15 years ago
• From the perspective of today’s systems, legacy portals are products of a bygone
past
• It is now perfectly reasonable for a scientist to want all the data
• Machine learning + HPC
• But this only works if the scientists can get to the data at scale

IT/Engineering (not the scientists) Have To Do This
• The scientific community cannot do this for themselves
• Individual researchers do not control the resources
• Computing centers
• Data repositories
• Science networks
• Our community owns these – we have to do the work
• Integration, performance engineering, interoperability
• Science Engagement to teach scientists how to use the
better platforms
• This is the path forward
¯_(ツ)_/¯

Networks Cannot Do This Alone
• We need a whole-community effort
• Networks
• HPC facilities
• Data repositories / Data portals
• Experimental facilities
• Science collaborations
• Science programs
• Networks can help, and must be part of the conversation
• Heavy lifting is now at the network edge, in collaboration with the
network core
• Need to get the architecture right – we know how to do this
• Engagement is critical
• We have to teach people to use the new thing
• They aren’t going to magically find out about new services

Long-Term Vision
ESnet
(Big Science facilities,
DOE labs)
Internet2 + Regionals
(US Universities and
affiliated institutions)
International Networks
(Universities and labs in Europe,
Asia, Americas, Australia, etc.)
High-performance feature-rich
science network ecosystem
Commercial Clouds
(Amazon, Google,
Microsoft, etc.)
Agency Networks
(NASA, NOAA, etc.)
Campus HPC
+
Data

It’s All A Bunch Of Science DMZs
DTN
DTN
DTN
DTN
DMZDMZ
DMZDMZ
DTN
DATA
DTN
DMZDMZ
DTN
DTN
DMZDMZ
DATA
DTN DTN
DMZDMZ
Parallel
Filesystem
DTN DTN
DTN
DTNDMZDMZ
DATA
DTN
DTN
DMZDMZ
Experiment
Data Archive
DTN
DTN
DTN
DTN
DTN DMZDMZ
DTN DTN
DMZDMZ
DATA

It’s All A Bunch Of Science DMZs
DTN
DTN
DTN
DTN
DMZDMZ
DMZDMZ
DTN
DATA
DTN
DMZDMZ
DTN
DTN
DMZDMZ
DATA
DTN DTN
DMZDMZ
Parallel
Filesystem
DTN DTN
DTN
DTNDMZDMZ
DATA
DTN
DTN
DMZDMZ
Experiment
Data Archive
DTN
DTN
DTN
DTN
DTN DMZDMZ
DTN DTN
DMZDMZ
DATA
HPC Facilities
Single Lab
Experiments
Data
Portal
LHC
Experiments
University
Computing

Outline
• Motivation
• TCP Basics
• Conclusions

ESnet’s Science DMZ – Enabling Big Data Science
The Science DMZ design pattern supports secure, high-performance data movement between and among
facilities, labs, universities, data portals, and clouds
DevelopmentDeploymentPlatform
10GE10GE
10GE
10GE
Border Router
WAN
Science DMZ
Switch/Router
Firewall
Enterprise
perfSONAR
perfSONAR
10GE
10GE
10GE
10GE
DTN
DTN
API DTNs
(data access governed
by portal)
DTN
DTN
perfSONAR
Filesystem
(data store)
10GE
Portal
Server
Browsing path
Query path
· web server
· search
· database
· authentication
Data Path
Data Transfer Path
Portal Query/Browse Path
Long Term Vision
Development Support
§ Science DMZ supports development,
prototyping, and incremental
deployment of advanced network and
data services
§ Petascale DTN
Project
§ Ongoing production
operation across
science complex
2017-2018
§ Modern Research
Data Portal design
pattern
§ Pacific Research
Platform
§ Science DMZs deployed
at DOE labs and facilities
§ NSF CC-NIE and
successor programs fund
over 100 Science DMZs
at universities
2011-20162005-2010
§ Requirements analysis of DOE Office of
Science facilities and projects from network
and data perspective
§ Performance engineering, troubleshooting,
experience
§ Creation of Science DMZ design pattern
Science Networks:
Campus, Regional,
National, International
Science DMZ
Science DMZ
Science DMZ Science DMZ
DTN
DTN
DTN
DTN
DTN
DTN
DTN
DTN
Data Platform
Data Platform
§ Science Engagement model for science
support
§ OIN workshop series trains CI engineers
throughout the community
21.2/22.6/24.5
Gbps
23.1/33.7/39.7
Gbps
26.7/34.7/39.9
Gbps
33.2/43.4/50.3
Gbps
35.9/39.0/40.7
Gbps
29.9/33.1/35.5
Gbps
34.6/47.5/56.8
Gbps
44.1/46.8/48.4
Gbps
41.0/42.2/43.9
Gbps
33.0/35.0/37.8
Gbps
43.0/50.0/56.3
Gbps
55.4/56.7/57.4
Gbps
DTN
DTN
DTN
DTN
NERSC DTN cluster
Filesystem: /project
Data set: L380
Files: 19260
Directories: 211
Other files: 0
Size distribution:
November 2017
L380 Data Set
Gigabits per second
(min/avg/max), three
transfers
ALCF DTN cluster
Filesystem: /projects
OLCF DTN cluster
Filesystem: atlas2
NCSA DTN cluster
Filesystem: /scratch
10GE
10GE
10GE
10GE
10G
Border Router
WAN
Science DMZ
Switch/Router
Enterprise Border
Router/Firewall
Site / Campus
LAN
High performance
Data Transfer Node
Per-service
security policy
control points
Clean,
High-bandwidth
WAN path
Site / Campus
access to Science
DMZ resources
perfSONAR
perfSONAR
perfSONAR
Production Data Movement
§ Science DMZs support large scale data
movement and data placement across
the science complex (DOE and
collaborators)
§ Performant security
Scalable Data Platform
§ Science DMZs provide
foundation for data-driven
science globally

Tutorial: Maximizing Performance and Network Utility with a Science DMZ

More Related Content

What's hot (20)

Similar to Tutorial: Maximizing Performance and Network Utility with a Science DMZ (20)

More from Globus (20)

Recently uploaded (20)

Tutorial: Maximizing Performance and Network Utility with a Science DMZ