Transcript INCITE-poster-March2004

INCITE
R. Baraniuk, E. Knightly, R. Nowak, R. Riedi (Rice), L. Cottrell, J. Navratil (SLAC), W. Feng, M. Gardner (LANL)
Edge-based Traffic Processing and Service Inference for High-Performance Networks
INCITE:
InterNet Control and
Inference Tools at the Edge
1
4
Network Tomography (Rice, Wisconsin)

1
1-by-2
Component
2-by-1
Component
q4
q1
q2
Arrival order fixed
at joining point
q4
q1
q3
q2
q5
• Poor understanding of origins
of complex network dynamics
• Lack of adequate modeling
techniques for network dynamics
• Internal network inaccessible
• Low impact, large scale
monitoring
• Application-driven traffic
modulation
• High-speed measurements
3
Mean
+
PingER/ABwE (SLAC)
8
q6
q3
Common Branch
Different Branch
Point: Arrival order Points: arrival order
usually the same
varies depending on
delays, offset
participate in science due to poor
Internet connectivity
•e.g. 10-20% of HENP
collaborators are from
developing nations
• To understand need simple, low cost, performance
measurements to and within developing regions
providing:
• planning, setting expectations, policy setting
• PingER meets these needs
• < 100bits/s, uses ubiquitous ping
• covers > 100 countries (>90% of world’s Internet
connected population)
Rice LAN
Pinger deployment
Blue=monitoring site
Red=remote site
Arrival Order and Loss
Arrival Order Only
Loss Only
1000 probes
5
=
• Key: both application and network properties important
for traffic modeling
ROC Curve
Technical Challenges
2
99%
• Many scientists are unable to
Arrival Order Based Topology ID
 Approach:
 Active and passive network probing
 Statistical model based inference
beta
alpha
bytes
per
time
plots
1
 Improve throughput over the Internet
for DoE high performance projects
Thrust 1: Traffic analysis and modeling
Thrust 2: Path and tomographic inference
Thrust 3: Data collection tools
(PingER, MAGNeT, +)
• Cause of burstiness in traffic?
• Alpha: cause bursts, large transfers, high rate, low RTT,
few connections
• Beta: not-bursty, low rate, high RTT, most connections,
possess long-range-dependence
Two senders/receivers problem
characterizes network tomography
problem in general
From edge-based traffic
measurements
(loss/delay/arrival order),
infer internal topology, link
level loss rates, queuing
delays
Objectives:
7
Canonical Subproblems:
?
Alpha-Beta Traffic Model (Rice)
pathChirp: Efficient Available Bandwidth
and Tight Link Estimation (Rice)
Chirp: packet train with
increasing rate
When probe rate exceeds
available bandwidth,
queuing delay increases
ABwE tool: abing Characteristics
• Interactive (1 – 2 second response)
• Low network impact (20 packets/host/direction)
• Simple & robust: just need simple responder
installing
• Provides measurements in both directions
• Provides capacity & available bandwidth
• Agrees with more intense/complex methods
• Used in MonALISA, IEPM-BW & PlanetLab
Bandwidth
Impact and Connections
 Impact:
 Optimize performance of demanding
applications (remote visualization, highcapacity data transfers)
 New understanding of the complex dynamics
of large-scale, high-speed networks
 New edge-based tools to characterize and
map network performance as a function of
space, time, resource, application, protocol,
and service
 Highly efficient methods for monitoring in
distributed computing systems.
Connections:
 Rice/SLAC/LANL synergy
• Particle Physics Data Grid
Collaboratory Pilot
(Newman, Cottrell, Mount).
• SciDAC Center for Supernova
Research (Warren)
• Scientific Workspaces of the
Future (ANL, UIC, LANL, BU,
Brown, NCSA).
 Globus
•
•
•
•
•
•
•
•
Teragrid
Transpac at Indiana U.
European GridLab Project
San Diego Supercomputing Center
Telcordia
IEPM-BW
Internet2
ns-2 Simulator
Reduce available
bandwidth on Gigabit
testbed using cross-traffic
generator
Locating tight
links on two
paths sharing
4 common
links
Available bandwidth
estimates decrease in
proportion to the introduced
cross-traffic
UIUCRice tight link
SLACRice tight link
TCP Low-Priority (Rice)
6
Goal: Utilize excessive
bandwidth in a non-intrusive
fashion
Applications: bulk data
transfer, P2P file sharing
High-speed TCP-LP
•TCP-LP + HSTCP [Floyd03]
•Linux-2.4.22-web100
implementation
• TCP alone 745.5 Kb/s
• TCP plus
TCP-LP
739.5 Kb/s
109.5 Kb/
• TCP-LP is invisible to TCP
DoE SciDAC high-performance networking research project: INCITE
The graphs show Abing monitoring data
9

via
MonALISA
Tools: MAGNeT & TICKET (LANL)
MAGNeT:
 Monitor for Application-Generated Network
Traffic
 Monitor traffic immediately
after being generated by the
application throughout the
protocol stack to see how
traffic gets modulated. Is
TCP/IP the obstacle to high
performance?
 TICKET:
 Traffic Information-Collecting Kernel with Exact Timing
 Current solutions to network packet capture (e.g.,
tcpdump) are too slow or too expensive
 Monitor and record traffic at gigabit-per-second
(Gb/s) speeds and nanosecond granularity
INCITE.rice.edu
2004