Transcript FAST v3

FAST TCP
Cheng Jin
David Wei
Steven Low
netlab.CALTECH.edu
GNEW, CERN, March 2004
Acknowledgments

Caltech








Bunn, Choe, Doyle, Jin, Newman, Ravot, Singh, J.
Wang, Wei
UCLA

Paganini, Z. Wang

Martin, Martin-Flatin

Cottrell, Mount

Aiken, Doraiswami, Yip

Fernes

Wu
CERN/DataTAG
Internet2
 Almes, Shalunov
SLAC
Cisco
Level(3)
LANL
FAST project
NSF RI
(2003)
NSF
STI
(2002)
HEP networks TeraGrid
Abilene
DummyNet
Abilene
HEP networks PlanetL
WAN
in Lab
Linux
TCP kernel
NSF
ITR
(2001)
Performance
Stability
Other
platforms
Fairness TCP/IP
IETF
GGF
Deployment
UltraLight
testbed
Experiment
Monitoring
Debugging
Implement
Noise
Random
-ness
Theory
Outline
Experiments
 Results
 Future plan
 Status
 Open issues
 Code release mid 04
 Unified framework
 Reno, FAST, HSTCP, STCP, XCP, …
 Implementation issues
Aggregate throughput
88%
FAST
 Standard MTU
 Utilization averaged over > 1hr
90%
90%
Average
utilization
92%
95%
1hr
1 flow
1hr
2 flows
6hr
7 flows
1.1hr
6hr
9 flows
10 flows
DataTAG: CERN – StarLight – Level3/SLAC (Jin, Wei, Ravot, etc SC2002)
FAST
Linux
Dynamic sharing: 3 flows
Dynamic sharing on Dummynet





capacity = 800Mbps
delay=120ms
3 flows
iperf throughput
Linux 2.4.x (HSTCP: UCL)
FAST
Linux
Dynamic sharing: 3 flows
Steady throughput
HSTCP
STCP
queue
FAST
loss
Linux
throughput
30min
Dynamic sharing on Dummynet
 capacity = 800Mbps
HSTCP
 delay=120ms
 14 flows
 iperf throughput
 Linux 2.4.x (HSTCP: UCL)
STCP
queue
Room for mice !
FAST
loss
Linux
throughput
HSTCP
HSTCP
30min
STCP
Aggregate throughput
ideal
performance
Dummynet: cap = 800Mbps; delay = 50-200ms; #flows = 1-14; 29 expts
Aggregate throughput
large
window
8000
small
window
800pkts
Dummynet: cap = 800Mbps; delay = 50-200ms; #flows = 1-14; 29 expts
Fairness
Jain’s index
Dummynet: cap = 800Mbps; delay = 50-200ms; #flows = 1-14; 29 expts
Stability
stable in
diverse
scenarios
Dummynet: cap = 800Mbps; delay = 50-200ms; #flows = 1-14; 29 expts
Outline
 Experiments
 Results
 Future plan
 Status
 Open issues
 Code release
 Unified framework
 Reno, FAST, HSTCP, STCP, XCP, …
 Implementation issues
Benchmarking TCP
 Not just static throughput
 Dynamic sharing, what protocol does to network, …
 Tests to zoom in on specific properties
 Throughput, delay, loss, fairness, stability, …
 Critical for basic design
 Test scenarios may not be realistic
 Tests with realistic scenarios
 Same performance metrics
 Critical for refinement for deployment
 Just started
 Input solicited
 What’s realistic for your applications?
Open issues: well understood
 baseRTT estimation
 route changes, dynamic sharing
 does not upset stability
 Small network buffer
 at least like TCP
 adapt a on slow timescale, but how?
 TCP-friendliness
 friendly at least at small window
 tunable, but how to tune?
 Reverse path congestion
 should react? rare for large transfer?
Status: code release
 Source release mid 2004
 For any non-profit purposes
 Re-implementation of FAST TCP completed
 Extensive testing to complete by April 04
 Pre-release trials
 CFP for high-performance sites!
 Incorporate into Web100
 with Matt Mathis
Status: IPR
Caltech will license royalty-free if FAST TCP
becomes IETF standard
 IPR covers more broadly than TCP
 Leave all options open
Outline
 Experiments
 Results
 Future plan
 Status
 Open issues
 Code release mid 04
 Unified framework
 Reno, FAST, HSTCP, STCP, XCP, …
 Implementation issues
Packet & flow level
Reno TCP
 Packet level
W

W + 1/W
Loss: W

W – 0.5W
ACK:
 Flow level
 Equilibrium
 Dynamics
pkts
(Mathis formula)
Reno TCP
 Packet level
 Designed and implemented first
 Flow level
 Understood afterwards
 Flow level dynamics determines
 Equilibrium: performance, fairness
 Stability
 Design flow level equilibrium & stability
 Implement flow level goals at packet level
Reno TCP
 Packet level
 Designed and implemented first
 Flow level
 Understood afterwards
 Flow level dynamics determines
 Equilibrium: performance, fairness
 Stability
Packet level design of FAST, HSTCP, STCP, H-TCP, …
guided by flow level properties
Packet level
 Reno
AIMD(1, 0.5)
 HSTCP
AIMD(a(w), b(w))
 STCP
MIMD(a, b)
 FAST
W

W + 1/W
Loss: W

W – 0.5W
ACK:
W

W + a(w)/W
Loss: W

W – b(w)W
ACK:
W

W + 0.01
Loss: W

W – 0.125W
ACK:
baseRTT
RTT : W  W 
a
RTT
Flow level:
Reno, HSTCP, STCP, FAST
 Similar flow level equilibrium
MSS/sec
a = 1.225 (Reno), 0.120 (HSTCP), 0.075 (STCP)
Flow level:
Reno, HSTCP, STCP, FAST
 Common flow level dynamics
window
adjustment
=
control
gain
flow level
goal
 Different gain k and utility Ui
 They determine equilibrium and stability
 Different congestion measure pi
 Loss probability (Reno, HSTCP, STCP)
 Queueing delay (Vegas, FAST)
FAST TCP
 Reno, HSTCP, and FAST have common flow level
dynamics
window
adjustment
=
control
gain
flow level
goal
 Equation-based
 Need to estimate “price” pi(t)
 pi(t) = queueing delay
 Easier to estimate at large window
 k(t) and U’i(t) explicitly designed for
 Performance
 Fairness
 Stability
Window control algorithm
 Full utilization
 regardless of bandwidth-delay product
 Globally stable
 exponential convergence
 Intra-protocol fairness
 weighted proportional fairness
 parameter a
FAST tunes to knee
Goal:
• Less delay
• Less jitter
FAST
TCP
stabilized
oscillation
Window adjustment
FAST TCP
netlab.caltech.edu/FAST
 FAST TCP: motivation, architecture,
algorithms, performance
IEEE Infocom March 2004
 FAST TCP: from theory to experiments
Submitted for publication April 2003
Panel 1:
Lessons in Grid Networking
Metrics
Performance
 Throughput, loss, delay, jitter, stability,
responsiveness
Availability, reliability
Simplicity
 Application
 Management
Evolvability, robustness
Constraints
Scientific community
 Small & fixed set of major sites
 Few & large transfers
 Relatively simple traffic characteristics and
quality requirements
General public
 Large, dynamic sets of users
 Diverse set of traffic characteristics &
quality requirements
 Evolving/unpredictable applications
Mechanisms
Fiber infrastructure
Lightpath configuration
Resource provisioning
Traffic engineering, adm control
Congestion/flow control
Months - years
Mintes - days
Service: sec - hrs
Flow: sec - mins
RTT: ms - sec
Timescale: desired, instead of feasible
Balance: cost/benefit, simplicity, evolvability