FAST v3 - California Institute of Technology
Download
Report
Transcript FAST v3 - California Institute of Technology
WAN in Lab
NSF Site Visit
John Doyle,
CDS/EE/BE
Steven Low (PI),
CS/EE
Harvey Newman, Physics
Demetri Psaltis, EE/CNS
Steven Yip, Cisco
March 5, 2003
Reviewer concerns
Narrow focus on TCP/AQM
A range of IST research at Caltech
Spanning theory, implementation, experiment, deployment
WAN in Lab a critical component
Alternatives not discussed
Use spectrum of tools at different stages
How to manage and share WAN in Lab
Part of Federated Emulab
Both demand and excellent support for global sharing
Experience in global collaboration, e.g. Newman’s VRVS
netlab.caltech.edu
Agenda
EAS, IST Initiative, Theory program, FAST
Intellectual environment in which WAN in Lab fits (Murray,
Doyle, Low)
WAN in Lab
Design, capabilities, alternatives, management (Low)
Cisco example & collaboration (Yip)
Education, outreach, poster session
Research talks
Projects that will use WAN in Lab
International collaboration, leverage & impact on HENP &
Grids
netlab.caltech.edu
WAN in Lab
Steven Low
netlab.CALTECH.edu
NSF Site Visit
March 5, 2003
Why Testbed in IST
“A lack of wide-area testbeds would contribute to a
growing tendency towards paper solutions to thesisfactory problems, leaving the real networking world
short of new ideas and technologies”
“Prototypes & testbeds are required to gain
acceptance of new concepts with potential user
communities”
A value of testbeds is “… building and maintaining
research collaborations and communities”
- NSF Workshop on Network Research Testbeds (Nov 2002)
Outline
Proposal summary
Basic design, equipment, costs
Unique features
Alternatives
Spectrum of tools
Emulated delay
Community resource
Demand
Management software
Why Caltech
Leverage on Abilene, HENP, CalREN, TeraGrid
Summary
Reviewer concerns
Review criteria
netlab.caltech.edu
Goal
State-of-the-art WAN
High speed
2.5G 10G
Large distance
50 – 200ms
Controlled & repeatable experiments
Reconfigurable & evolvable
netlab.caltech.edu
S
R
l1
S
OPM
l1
fiber spool
R
S
S
l20
S
EDFA
EDFA
R
l20
S
R
S
S
500 km
H : server
R : router
Max path length = 10,000 km
Max one-way delay = 50ms
electronic
crossconnect
Equipment
26 Servers
GbE cards ( 10GbE cards)
12 routers
10 Cisco 15454 with router blades
2-port GbE, 8-channel OC48
2 Force10 E600
24-port GbE, 2-port OC48
DWDM gears
500km fiber
6 EDFA
2 Dispersion compensation modules
2 optical mux/demux
Tektronix TDS7404 Oscilloscope
Integration with global network
netlab.caltech.edu
Costs
26 Servers: $104K
12 routers: $1.03M
2 Force10 E600: $280K ($340K if OC192)
10 Cisco 15454 with router blades: $750K ($810K if OC192)
DWDM gears: $148K
500km fiber: $8K
6 EDFA: $60K
2 Dispersion compensation modules: $40K
2 optical mux/demux: $40K
Tektronix TDS7404 Oscilloscope: $50K
Integration with global network: $110K
Personnel, software, service & maintenance
Total: $2M (NSF) + $0.67M (cost sharing)
netlab.caltech.edu
Yearly costs
Year 1: $1.128K
10 servers, 5 routers, 2.5Gbps
Year 2: $564K
20 servers, 8 routers, 2.5Gbps
Year 3: $124K
Software development
Year 4: $733K
26 servers, 10 routers, 2.5Gbps
Year 5: $120K
Software development
Total: $2M (NSF) + $0.67M (cost sharing)
netlab.caltech.edu
Networking Lab
NetLab
WAN in Lab
3 racks, 2 consoles
Networking Lab
424 sq ft
Next to CACR
Easy connection to
global network
Renovation (cost
sharing)
New IST Building
Jorgensen Lab
netlab.caltech.edu
Unique capabilities
WAN in Lab
Capacity: 2.5 – 10 Gbps
Delay: 0 – 100 ms round trip
Configurable & evolvable
Topology, rate, delays, routing
Always at cutting edge
Risky research
l1
l1
R1
MPLS, AQM,
routing,
…
l2
l2
Integral
part
l3 of R&A networks
l3
R2
l4
Transition from theory, implementation,
demonstration, deployment
l18from lab to marketplace
Transition
R10 resource
l19
l19
Global
l20
l20
(a) Physical network
netlab.caltech.edu
Unique capabilities
WAN in Lab
Capacity: 2.5 – 10 Gbps
Delay: 0 – 100 ms round trip
Configurable & evolvable
Topology, rate, delays, routing
Always at cutting edge
Risky research
MPLS, AQM, routing, …
l1
l2
lIntegral
part
of
R&A
networks
R1
20
l3
R2
Transition from theory, implementation,
demonstration, deployment
l19 Transition from lab to marketplace
l4
R10
R3
Global resource
(b) Logical network
netlab.caltech.edu
Unique capabilities
WAN in Lab
research & production
WAN in Lab
networks
Caltech
Capacity: 2.5 – 10 Gbps
Chicago
CERN
Delay: 0 – 100 ms round trip
Configurable & evolvable
StarLight
Calren2/Abilene
Topology, rate, delays, routing
Geneva
Multi-Gbps
Always at cutting edge
50-200ms delay
SURFNet
Risky research
Experiment
Amsterdam
Dynamic recovery, AQM, MPLS, routing, …
Integral part of R&A networks
Transition from theory, implementation,
demonstration, deployment
Transition from lab to marketplace
Global resource
Federated Netlab (Emulab)
netlab.caltech.edu
Outline
Proposal summary
Basic design, equipment, costs
Unique features
Alternatives
Spectrum of tools
Emulated delay
Community resource
Demand
Management software
Why Caltech
Leverage on Abilene, HENP, CalREN, TeraGrid
Summary
Reviewer concerns
Review criteria
netlab.caltech.edu
Spectrum of tools
log(cost)
HENP
Abilene
CalREN
WAIL
PlanetLab
CAIRN
NLR
?
DummyNet
EmuLab
ModelNet
NS
WAIL
SSFNet
QualNet
JavaSim
Mathis formula
Optimization
Linear model
Nonlinear model
Stocahstic model
log(abstraction)
live nk
WANiLab
emulation
simulation
math
…we use them all
netlab.caltech.edu
Spectrum of tools
live nk
WANiLab
emulation
Distance
High
High
High
Speed
High
High
Low
Realism
High
High
Low
Traffic
High
Low
Low
Configurable
Low
Medium
High
Monitoring
Low
Medium
High
Cost
High
Medium
Low
netlab.caltech.edu
simulation
math
Critical in
development
e.g. Web100
Emulated delay
S
S
l1
R
S
l20
S
High speed
electronic memory
R
l1
R
S
l20
S
R
S
S
Available technology inadequate
Spirent SX/14 Link Simulator: 1ms (155Mbps) – 10s (100bps)
Adequate technology too expensive
2.5Gbps, 100ms delay: IC expert at least 2 man-years & $200K
Less realistic
netlab.caltech.edu
HENP testbed
CWND: 5801-5815
Ins. RTT
Sylvain Ravot (Caltech/CERN)
netlab.caltech.edu
Example1: end-to-end delay
Y: RTT (us)
CWND: 5801-5815
Ins. RTT
instantaneous RTT
average RTT
netlab.caltech.edu
Delay between Geneva & Chicago
X:Real Time (us)
Example1: end-to-end delay
Y: RTT (us)
CWND: 5801-5815
Ins. RTT
RTT=270ms
12450 pkts!?
1500instantaneous
pkt time without
RTTbuildup!?
netlab.caltech.edu
Delay between Geneva & Chicago
X:Real Time (us)
Example1: end-to-end delay
Y: RTT (us)
CWND:
8700->4000
RTT=980ms!?
Ins. RTT
Ins. RTT
Avg RTT
Passive monitoring in WANiLab can help debug
netlab.caltech.edu
X:Real
X:Real
Time
Time
(us)
Example2: 10G Expt
inst RTT
avg RTT
ms
Losses & retransmissions
Delay between Geneva & Sunnyvale
netlab.caltech.edu
Real time ms
Example2: 10G Expt
inst RTT
avg RTT
ms
Retransmission without loss!?
Losses & retransmissions
Real time ms
netlab.caltech.edu
Example2: 10G Expt
inst RTT
avg RTT
ms
Retransmission without loss!?
Real time ms
Passive monitoring in WANiLab can help debug
netlab.caltech.edu
Network debugging
Performance problems in real network
Simulation will miss
Emulation might miss
Live network hard to debug
Enable or speed up FAST development
10GExpt: 20 people in 8 organizations for 3 months
Complete facility available only for a week
Many mysteries unresolved
WAN in Lab
Passive monitoring inside network
Active debugging possible
netlab.caltech.edu
Passive monitoring
Fiber
splitter
GPS
DAG
Timestamp
Header
RAID
Monitor
No overhead on system
Can capture full info at OC48
UofWaikato’s DAG card captures
at OC48 speed
Can filter if necessary
Disk speed = 2.5Gbps*40/1500
= 66Mbps
Monitors synchronized by GPS
or cheaper alternatives
Data stored for offline
analysis
David Wei (Caltech)
netlab.caltech.edu
Passive monitoring
Web100, FAST monitor
Fiber
splitter
Server
monitor
GPS
router
DAG
Timestamp
Header
monitor
monitor
monitor
RAID
router
Monitor
monitor
David Wei (Caltech)
netlab.caltech.edu
monitor
Server
Outline
Proposal summary
Basic design, equipment, costs
Unique features
Alternatives
Spectrum of tools
Emulated delay
Community resource
Demand
Management software
Why Caltech
Leverage on Abilene, HENP, CalREN, TeraGrid
Summary
Reviewer concerns
Review criteria
netlab.caltech.edu
DataTAG link
CWND: 5801-5815
Ins. RTT
Sylvain Ravot (Caltech/CERN)
netlab.caltech.edu
DataTAG link
Funded by EU (CERN), USA (DoE, NSF, Caltech)
OC48 circuit StarLight-CERN
Upgrade to OC192 by August 2003
Linux farms
StarLight: 20 CPU (P4), 20 Syskonnect
CERN: 12 CPU (P4), 12 Syskonnect
50 users, 13 institutes, 7 countries (Feb 2003)
Heavy utilization
European hours: 100% reservation
US hours: 25% reservation, but busy
netlab.caltech.edu
Netbed (Emulab)
Funded by NSF with Cisco donations
Integrates simulation, emulation, live Internet
Emulab Classic
University Utah: 168 PC, 5 100M Ethernet
cards
Connected by 4 Cisco 6409
Testbed backplane limited to 2Gbps
University of Kentucky: 48 PC, similar setup
Netbed: Federated Emulab
University of Utah
netlab.caltech.edu
Dummynet & VLAN
32 nodes in 25 sites
Heavy utilization
July2002
65 user accounts (40 external)
54 projects
Feb 2003
400 user accounts
94 projects (10 Utah, 78 US, 6 Int’l)
Management software
Part of Federated Emulab
Tailor Emulab management software
Jay Lepreau’s team consult on setup
Complementary to existing federated Emulab
WANiLab
High speed large distance (Gbps WAN)
Small network (30 nodes)
Emulab
Low speed (100Mbps LAN, 10M WAN)
Large network (200+ nodes)
Instantly available to Emulab community
Web accessible anywhere any time
Virtual machine for network experimentation
netlab.caltech.edu
Experiment life cycle
(White et al)
Experiment creation
Web based sign-up form by project lead
Approved by Emulab team
Experiment specification
ns script or Java GUI
Can download own OS, host algorithms, etc
Links emulated by Dummynet nodes with specified rate,
delay, loss
Experiment realization
Map target configuration to physical resources
Reserve resources for each experiment
Oversubscription
dynamic reallocation, swap in, swap out
netlab.caltech.edu
Why Caltech: synergies
Caltech networking research
FAST project: the missing experimental facility (Doyle, Low)
IST Initiatives: testbed tied to rich theory program (Murray, Psaltis)
Combination of theory, implementation, experiment & deployment
Synergy in research
Caltech’s leadership role in IT for global HENP (Newman)
Vibrant research in HENP, astronomy, geological sci, biology, visualization,
CACR
Early testing ground & adopter of FAST (Newman)
Availability of real data for ultrascale networks
Synergy in facility
Integration with HENP networks, Abilene, CalREN XD, TeraGrid (see
Newman’s talk)
Synergy with Cisco
See Yip’s talk
netlab.caltech.edu
Why Caltech: experience
Hardware
Cisco’s testbed
Psaltis, Yip, Hajimiri, DeHon
Software
Netbed management software
Operation
Newman’s group
Testbed driven by networking research
IST, Theory Program, FAST, optics, scientific
computing, network coding, …
netlab.caltech.edu
Team
Hardware
Yip’s team: Doraiswami (Cisco)
Psaltis (EE/CNS), DeHon (CS), Hajimiri (EE)
Caltech Information Tech Services, CACR
Software
Lepreau’s team
Low’s team: Almsberger (CS), Jin (CS), Wei (CS), Hu (CS)
Operation
Newman’s team: Bunn (Physics), Ravot (Physics/CERN),
Suresh (CACR)
Testbed driven by networking research
Caltech IST Institute
netlab.caltech.edu
Global research network
NewYork
ABILEN
E
UK
SuperJANET4
It
GARR-B
STARLIGHT
ESNET
GENEVA
GEANT
NL
SURFnet
STAR-TAP
Fr
Renater
Newman (Caltech)
netlab.caltech.edu
WAN in Lab
Caltech
CALRE
N
Outline
Proposal summary
Basic design, equipment, costs
Unique features
Alternatives
Spectrum of tools
Emulated delay
Community resource
Demand
Management software
Why Caltech
Leverage on Abilene, HENP, CalREN, TeraGrid
Summary
Reviewer concerns
Review criteria
netlab.caltech.edu
Reviewer concerns
Narrow focus on TCP/AQM
A range of IST research at Caltech (Murray, Doyle)
Spanning theory, implementation, experiment, deployment
WAN in Lab a critical component
External projects in HENP, Grid & Emulab communities
Alternatives not discussed
Use spectrum of tools at different stages
Each complementary but not replaceable
DWDM gears more realistic and cheaper
How to manage and share WAN in Lab
Part of Federated Emulab
Both demand and excellent support for global sharing
Experience in global collaboration, e.g. VRVS
How much hardware development needed
Mostly off-the-shelf (Yip)
Sample system & experience from Cisco
Local expertise: Psaltis (Optics), Yip (Cisc), Hajimiri (high speed IC), DeHon
(VLSI)
netlab.caltech.edu
Review Criteria
Intellectual merit
Theory, implementation, experiment, deployment
Must inform and influence each other intimately
Approach validated by pilot project
Experimental facility tied to rich theory program
Broader impacts
HENP’s global collaborations a model for future
corporations & society
FAST protocols enabling technology
Shared by & stimulate external research that need high
speed large distance
HSTCP, Scalable TCP, TCP Westwood, AVQ, REM/PI, …
Internet as simplest complex system
netlab.caltech.edu
Review Criteria
Integration of research & education
Excellent projects for undergraduates and graduates
During & after development
Unique teaching platform for advanced networking, distributed
systems, complex systems, optics course
Bruck, Chandy, Doyle, Hickey, Low, Psaltis
Diversity
33% women grad students in Netlab
50% women postdocs and grad students in Doyle’s group
Synergy among projects
Bring together 4 CISE projects (1 ITR, 1 STI, 2 pending)
Leverage for additional funding and industry collaborations
netlab.caltech.edu
NSF Workshop Criteria
Tested driven by research agenda
Rich and strong networking effort
“A network that can break”
Multi-user experimental facility
With a clear research focus and foreseeable impact
Federated testbed
Leverage on Netbed’s management software
Integrated monitoring & measurement facility
Fiber splitter passive monitors
Technology transfer
Strong leadership in FAST user community (Newman)
netlab.caltech.edu
Some potential projects
TCP: FAST, HSTCP(Floyd, ICIR), TCP Westwood(Gerla, UCLA), Scalable
TCP(Kelly, Cambridge/CERN), XCP (Dina, MIT)
AQM: REM(Low), PI(Misra/Towsley), AVQ(Srikant, UIUC)
Protocol decomposition (Doyle, Low, Caltech)
Network self-management (Yemini, Columbia)
Content distribution (Bruck, Low, Caltech, Xu, Washington U)
Optical switching (Low, Psaltis, Caltech)
Network separation theory (Doyle, Low, Caltech Paganini, UCLA)
Real-time control over high performance networks (Dolye, Low, Murray,
Caltech)
Simple Optics Smart Router (SOSR) (Yates, AT&T Research)
Optical protection, recovery (Yates, AT&T Research; Nirmalathas,
Melbourne U)
Dynamic lightpath configuration & provisioning (Tucker, Melbourne U)
Active probing (Veitch, CUBIN)
Passive monitoring (Veitch, CUBIN)
Building & testing firewalls (Hoffman, U of Victoria)
High performance active network node (Turner, Washington)
netlab.caltech.edu