Slides - Cenic

Report
Christopher Paolini
Computational Science Research Center
San Diego State University
100G and Beyond Workshop: Ultra High Performance Networking in California
Calit2 Auditorium • First floor, Atkinson Hall • UC San Diego • La Jolla, CA
Tuesday, February 26, 2013 · Campus and Lab Strategies Panel · 11:00AM – 12:00PM







University network operations centers support multiple,
conflicting missions.
Network Security or Network Performance: which is more
important?
NOCs typically accountable to university business
divisions and contend with legal and public relations
pressures → security wins always.
NOCs not usually accountable to research groups
(often never communicate with faculty).
University enterprise (e.g. general
purpose/financial/personal) computing:
security > performance
Computational and “Big Data” research:
performance > security
What can we do to ensure efficient scientific data
transfer between universities and national labs?
vs.




A network optimized for business is not designed or capable
of supporting data intensive science.
Universities will always need to
support security features that
protect organizational financial
and personnel data.
Solution: create separate data
intensive science network,
external to university enterprise
network
Design formalized by ESnet,
based on traditional network
DMZ paradigm

Science DMZ: (1) dedicated access to high-performance WAN,
(2) high-performance switching infrastructure (large buffer
memory), (3) dedicated data transfer nodes

Science
DMZ
using
CENIC
California
Research
and
Education
Network
resources
NSF Office of CyberInfrastructure
CC-NIE Grant 1245312
• Alcatel-Lucent 10
and 40 Gbps
switching devices,
per CSU policy
• DMZ spans four
campus buildings:
Administration,
Life Sciences
(CSRC Data
Center), Education
& Business
Administration
(UCO Data
Center), and
Chemical Sciences
(VizCenter)
• Primary users:
CSRC affiliated
faculty and
students
• AL OmniVista
2500 for network
management
• Computational
science network
connects to the
DMZ
• Funded in 2009
through NSF MRI
award 0922702
• 8 Cisco 10 Gbps
Catalyst 4900M
switching devices
• CSRCnet spans five
campus buildings:
Administration, Life
Sciences (CSRC Data
Center), Education
& Business
Administration
(UCO Data Center),
Physics, and
Engineering
• Sole users: CSRC
affiliated faculty
and students
• 10G access to SDSC










Facilitate high-performance data transfer for scientific applications using
Globus Online GridFTP
Alcatel-Lucent OmniSwitch 10K (core device)
Two Alcatel-Lucent OmniSwitch 6900s (satellite devices)
Dedicated and independent 10GE (maybe 40GE) uplink to Internet2 and ESnet
via CENIC
Optimized network for high-volume bulk transfer of scientific datasets
Unencumbered, high-speed access to online scientific applications and data
generated at SDSU
External access to science resources not impacted by regular “enterprise” or
business class Internet traffic
Focus on “BigData” Intensive Science: earthquake rupture and wave
propagation, parallel 3D unified curvilinear coastal ocean modeling, geologic
sequestration simulation of supercritical CO2, large-scale proteomic data,
bioinformatics of gene promoter analysis, microbial metagenomics, and
high-order PSIC methods for simulation of pulse detonation engines
Network performance measurement based on the PerfSONAR framework
InCommon Federation global federated system for identity management and
authentication to DMZ connected hosts and services

Extension of the standard, two channel FTP protocol

Control Channel
◦ Command/Response
◦ Used to establish data channels
◦ Basic file system operations (e.g. mkdir, delete, etc.)

Data channel: Pathway over which file is transferred
Scheduled transfers using command line interface:

$ scp xsede#lonestar4:~/GO/bigdatafile xsede#trestles:~/GO/bigdatafile
$ scp xsede#trestles:~/GO/bigdatafile paolini#sdsu:~/GO/bigdatafile






Science DMZ performance monitoring accomplished using
perfSONAR tool suite
Server side tools run on designated hosts attached to key
switches
End-to-end testing with collaborating perfSONAR sites
Determine one way latencies and packet loss between hosts
using One-Way Active Measurement Protocol (OWAMP)
owping -c 10000 -i .01 remotedmz
Periodic throughput tests to remote Science DMZs using
Bandwidth Test Controller (BWCTL)
Resource allocation and scheduling daemon for regularlyscheduled Iperf tests
bwctl -s remotedmz -P 4 -t 30 -f M -w 4M -S 32




U.S. education and research identity federation service
Provides common framework for trusted shared management
of access to on-line resources
Provide users single sign-on convenience and privacy
protection – Shibboleth Service Provider Federating software
Site admins can delegate responsibility for administering
service provider (SP) metadata to another admin


Primary SDSU faculty/staff for Science DMZ implementation:
Name
Role
E-Mail
Phone
Christopher Paolini
CSRC Affiliated Faculty, Network
Engineering and Research
[email protected]
(619) 594-7159
Jose Castillo
Director of Computational Science
Research Center
[email protected]
(619) 594-3430
Rich Pickett
Campus CIO
[email protected]
(619) 594-8370
Kent McKelvey
Director of Network Services
[email protected]
(619) 594-3245
Skip Austin
Network Planning and Design
[email protected]
(619) 594-4211
Gene LeDuc
Technology Security Officer (TSO)
[email protected]
(619) 594-0838
Robert Osborn
Infrastructure Installation,
Configuration, and Support
[email protected]
(619) 594-6004
Current and planned DMZ related research:
Development of new transport layer protocols that use compressed sensing
techniques to perform sparse sampling on streaming petabyte sized datasets
originating from remote CO2 sequestration, curvilinear coastal ocean modeling,
and earthquake rupture and wave propagation simulations
Development of a new Alcatel-Lucent SDN/Application Fluent Network based
protocol for the OS10K that bridges Lustre RDMA traffic between 40GE and FDR
InfiniBand

similar documents