incite – edge-based traffic processing for high-performance networks
DESCRIPTION
INCITE – Edge-based Traffic Processing for High-Performance Networks. R. Baraniuk, E. Knightly, R. Nowak, R. Riedi Rice University L. Cottrell, J. Navratil, W. Mathews SLAC W. Feng , M. Gardner LANL web site: incite.rice.edu. INCITE Project. - PowerPoint PPT PresentationTRANSCRIPT
INCITE –Edge-based Traffic Processing
for High-Performance Networks
R. Baraniuk, E. Knightly, R. Nowak, R. Riedi Rice University
L. Cottrell, J. Navratil, W. MathewsSLAC
W. Feng, M. GardnerLANL
web site: incite.rice.edu
INCITE Project – Rice, SLAC, LANL 2 incite.rice.edu
INCITE Project• InterNet Control and Inference from The Edge
on-line tools to characterize and map host and network performance as a function of time, space, application, protocol, and service
INCITE Project – Rice, SLAC, LANL 3 incite.rice.edu
INCITE Thrusts and Tools
Thrust 1: Multiscale traffic analysis and modeling techniques
o wavelet, multifractal, connection-level models
Thrust 2: Inference and control algorithms for network paths, links, and routers
o end-to-end path probing and modelingo network tomography and topology discoveryo advanced high-speed protocols
Thrust 3: Data collection tools
o active measurement infrastructureo passive application-layer measurement
INCITE Project – Rice, SLAC, LANL 4 incite.rice.edu
pathChirp• Goal
– estimate instantaneous available bandwidth (ABW) on an end-to-end network link
• Basic probing paradigm– stream packets at some rate
no queuing delay rate<ABW queuing delay builds up
rate>ABW• Until now: tradeoff
– high accuracy has required high volume probing (inefficient)
• Unique to pathChirp – variable rate probe packet train
(exponentially spaced chirp)– 10x more efficient than
competing techniques
INCITE Project – Rice, SLAC, LANL 5 incite.rice.edu
Network TomographyFrom end-to-endmeasurements…
… infer internal topology and delay/loss characteristics
INCITE Project – Rice, SLAC, LANL 6 incite.rice.edu
TCP - Low Priority
• TCP alone 745.5 Kb/s
• TCP plus 739.5 Kb/sTCP-LP 109.5 Kb/s
• TCP-LP is invisible to TCP
• Goal– utilize excess bandwidth in a
non-intrusive fashion • Methodology
– sender-side modification of TCP: delay-based approach
• Applications– bulk data transfers– available bandwidth monitoring– P2P file sharing
• High-speed TCP-LP– TCP-LP + HSTCP– implementation
Linux-2.4.22-web100
– experiments Stanford - Ann Arbor Stanford - Gainesville
R 1 R 2
TC P-L P
TC P
C = 1 .5 M b/s
cro s s - t ra f f ic
INCITE Project – Rice, SLAC, LANL 7 incite.rice.edu
Advanced TCP stacks• Standard TCP (Reno) has problems on today’s long-
distance high-speed networks (e.g. trans ocean/continent > hundreds of Mbits/s)
• Advanced TCP stacks (e.g. FAST, High-speed, TCP-LP …) and new rate based UDP transports address this issue
• We have evaluated many (~10) new implementations for throughput, stability, fairness, ease of use etc.
• BaBar (HENP) tier A sites (e.g. SLAC, IN2P3 (Lyon Fr) and FZK (Karlsruhe)) now starting to use chosen TCP stack for production transfer of Monte Carlo data to SLAC– Easier to use than multi-stream TCP, only optimize
one parameter (window size)
INCITE Project – Rice, SLAC, LANL 8 incite.rice.edu
Changes in network topology (BGP) can result in dramatic changes in performance
Snapshot of traceroute summary table
Samples of traceroute trees generated from the table
ABwE measurement one/minute for 24 hours Thu 9 Oct 9:00am to Fri 10 Oct 9:01am
Drop in performance(From original path: SLAC-CENIC-Caltech to SLAC-Esnet-LosNettos (100Mbps) -Caltech )
Back to original path
Changes detected by IEPM-Iperf and AbWE
Esnet-LosNettos segment in the path(100 Mbits/s)
Hour
Rem
ote
host
Dynamic BW capacity (DBC)
Cross-traffic (XT)
Available BW = (DBC-XT)
Mbit
s/s
Note:1. Caltech misrouted via Los-Nettos 100Mbps commercial net 14:00-17:002. ESnet/GEANT working on routes from 2:00 to 14:00
Los-Nettos (100Mbps)
INCITE Project – Rice, SLAC, LANL 9 incite.rice.edu
Crossing the Application/Network Divide
Application
TCP
IP
Data Link
Network
Send dataover network
Segmentation
Fragmentation
Flow & Congestion Control
Checksums
::
• Implications to the application?• Insights for high- performance network
protocols?
Network monitors focus here.
INCITE Project – Rice, SLAC, LANL 10 incite.rice.edu
TICKET and MAGNET+MUSETICKET: Traffic Information-Collecting Kernel with Exact Timing
MAGNeT: Monitor for Application-Generated Network TrafficMUSE: MAGNET User-Space Environment
Application
TCP
IP
Data Link
Network
MAGNET
Send dataover network
Segmentation
Fragmentation
Flow & Congestion Control
Checksums
MUSE
TICKET:tcpdump++
::
For more information, go to www.lanl.gov/radiant/pubs.html
INCITE Project – Rice, SLAC, LANL 11 incite.rice.edu
MAGNeT MAGNET Monitoring Apparatus for General kerNel-Event Tracing (at nanoscale granularity)
• Why not extend monitoring to kernel events in general? Software Oscilloscope for Cluster and Grids – Debugging
e.g., IdentifiedLinux OS bug in the scheduler for SMPs. Can be used to deploy, debug, and monitor the DOE
UltraNet (UltraScienceNet), e.g., dynamic provisioning.– Performance Optimization
Improved performance of 10GigE adapters by 300%. Can improve end-to-end performance of DOE UltraNet.
– Monitoring Grid Applications Integrated MAGNET with SciDAC’s PERC TAU and
SciDAC’s PERC SvPablo/Autopilot.*– Adaptive Resource-Aware Applications
• SciDAC Deployment: PERC, Supernova Science Ctr, Transit Network Fabric + Terascale Supernova Initiative + Fusion Energy (emerging), and Earth Systems Grid II (emerging).
* For more information, see M. Gardner, W. Deng, T. Markham, C. Mendes, W. Feng, and D. Reed, “A High-Fidelity Software Oscilloscope for Globus,” GlobusWorld 2004, Jan. 2004.