Transcript Network Emulation for Researching Long-Fat

Network Emulation for Researching Long-Fat
Networks, Delay Tolerant
Networks and Grid Environments
Presented by:
Eric Coe
Computer Systems Research Department, M1–102
The Aerospace Corporation
www.aero.org
[email protected]
(310) 336-1911
February 9, 2005
Outline
• Aerolab: Aerospace Testbed based on EMULAB
– What it is, how it works, why we need it
• Define the following and motivate why network
emulation is appropriate
– Long Fat Network (aka High-Speed Bandits)
– Delay or Disruption Tolerant Networks
– Grid Environments
• Demo: So how easy is it?
• Future Directions
2
What is Aerolab?
Time- and Space-Shared Network Emulation System
• Construction of Complex Topologies
• Emulation of Link Conditions—Delay, Loss and Bandwidth
(dynamic)
• Boot Multiple Operating Systems—FreeBSD, Linux, and U of
Utah’s OS-Kit (real-time Linux, Windows on the Way)
• Runs on Commodity Hardware—PCs and Cisco Switch
• Scales to Hundreds of Physical Nodes and supports Multiple
Logical Nodes per Physical Nodes
Easy-To-Use Front-End
• Automates Configuration of OS, Topology and Links
• Supports both ns2 style scripting and Web-based GUI
Originally developed by University of Utah under a
National Science Foundation Grant
3
Aerolab Architecture
4
Utah’s Emulab
physical
nodes
programmable
switch
control
nodes
5
Why Network Emulation (Utah)
 “We evaluated our system on five nodes.”
-job talk from university with 300-node cluster
 “We evaluated our Web proxy design with 10 clients on
100Mbit ethernet.”
 “Simulation results indicate ...”
 “Memory and CPU demands on the individual nodes
were not measured, but we believe will be modest.”
 “The authors ignore interrupt handling overhead in their
evaluation, which likely dominates all other costs.”
 “Resource control remains an open problem.”
6
Why Network Emulation (Aerospace)
• “Different projects need to evaluate/operate different
networks”
• “Network set-up and running cables is costly and open
to configuration errors (HW/SW)”
• “Aerolab should hide the networking “magic” so that
researchers can experiment at the application or
protocol level”
• “In theory simulation and the real-world are the same
but in practice they are different”
• “Repeatability, Repeatability, Repeatability and
archiving. Old demos can be run w/ little overhead.”
7
Why Network Emulation (more..)
• “You have to know the right people to get access to the
cluster or real network.”
• “The cluster or real network is hard to use.”
• “<Experimental network X> runs FreeBSD 2.2.x.”
• “October’s schedule for <experimental network Y> is…”
• “<Experimental network Z> is tunneled through the
Internet”
8
Aerolab’s Value to Aerospace
Enhances Networking Capabilities
• Large Scale Experiments
• Complex Topologies
• Hardware and Software “In The Loop”
Automates Laboratory Operations
• Reduces Labor and Complexity
• Improves Repeatability and Allows “Archiving” Experiments
• Facilitates Resource Sharing
Emulab is not just the latest cool thing, it’s the
right way to run a networking lab.
9
Outline
• Aerolab: Aerospace Testbed based on EMULAB
– What it is, how it works, why we need it
• Define the following and motivate why network
emulation is appropriate
– Long Fat Network (aka High-Speed Bandits)
– Delay or Disruption Tolerant Networks
– Grid Environments
• Demo: So how easy is it?
• Future Directions
• Potential Collaboration
10
Long Fat Networks (LFN)
What they are:
• Characterized as networks with large bandwidth-delay products
• Typically gigabit speeds and have delays larger than 70 msec
• Examples: networks of compute-clusters; satellite networks;
transcontinental or transoceanic high-speed links.
• Network where standard TCP has poor performance
11
Long Fat Networks (LFN)
What an Aerolab like system provides:
• Install custom kernel images that contain various transport
protocols (HS-TCP, BIC, CUBIC, WEB100, XCP)
• Total control of the networking environment
• Run what-if scenario’s
• Guarantee no other traffic on the network or only “known” traffic
• Packet loss can be attribute to “WHO”.
• Dynamic Link Characteristics (satellite)
• Varying bandwidth
• Varying error rates, probability of packet loss
• Repeatability: When results from this week do not match last
weeks. .
12
Delay Tolerant Networks (DTN)
What they are:
• Characterized as networks with no contemporaneous end to end
path
• Occur due to intermittent connectivity which is caused by
• large propagation delays
• nodes going up and down frequently
• with highly mobile nodes in sparse network
• Networks where standard TCP has no performance due to the lack
of an end to end path
13
Delay Tolerant Networks (DTN)
14
Delay Tolerant Networks (DTN)
How we classify:
• Application basis
• Interplanetary Network (networking the solar system)
• Sensor networks
– Static nodes (due to nodes choosing their on and off periods
– Mobile nodes (mobility process dictates links presence or
absence)
– Hybrid case
• Link types
• Deterministic: Contacts can be scheduled
• Probabilistic: Contacts happen based on some distribution
• Dynamic: Contacts happen in a purely opportunistic way
15
Delay Tolerant Networks (DTN)
What an Aerolab like system provides:
• Dynamic link conditions for delay, bandwidth, and even link
availability
• Ability to run Bundling Protocol on top of a topology and over
different topologies
• Ability to evaluate routing protocols across the variety of networks
discussed earlier
• A common facility for other researchers to conduct experiments
Avoids everyone having their own “version/vision” of what constitutes a
DTN
16
Grid Environments
What they are:
• Networks of loosely coupled computational resources
• Heterogeneous by nature in both capacity and availability
• Basic grid services and applications must inhabit a dynamic,
heterogeneous, distributed environment
• Issue: How to quantitatively evaluate grid services and
applications in such an environment?
17
Quantitative Evaluation of Grids
What an Aerolab-like system could support for grid research:
• Playback mechanism for network behavior, cross traffic
• Injecting errors for testing fault tolerance, robustness
• Testing autonomic computing methods
(AC control cycle: monitor, analyze, plan, execute)
• Remove a link or halt processes on a host
• Evaluate response of autonomic control system
•
•
•
•
Testing information or resource discovery on larger networks
Quantitatively evaluate scalability in larger topologies
Quantitative evaluation wrt topology structure and connectedness
Tools needed to provide these capabilities in a general way
18
Example: Evaluation of Grid Application Performance
 DARPA Active Networks Project
 Exploit application topology to
compute lower-bound timestamp in
distributed simulations
 Globus used to manage Xbone
and start HLA-compliant simulation
 Very hard to run large test cases
USC
ISI
Aerospace
discover
APSimScript
Globus
overlay
request
start
overlay
config
MDS-2
register
XBone
Manager
APSim
APSim
APSim
HLA/RTI
HLA/RTI
HLA/RTI
ATMD
ATMD
overlay
HLA Time Mgmt done
“in the network” by
overlay of active time
mgmt daemons
19
Example: Evaluation of Grid Application Performance
 Solution: Emulate larger topologies on Utah EmuLab
 Set of quasi-random, tree topologies generated
 4, 8, 16, 32, 64 end-hosts
 9, 15, 22, 29, 34 interior service hosts
 Tree topologies constrained to have average 2.5 degree of connectedness
 550 Lower Bound Timestamp (DRN) calculations done for each topology
Example: topology w/ 32 end-hosts
20
Optimistic Mesh Generation on the Grid
 Optimistic Delauney Mesh Generation
 Optimism controlled by number of cavity
expansion msgs – affects BW demand, latency
 Already run on clusters and TeraGrid
 Work in Progress: run on AeroLab to do repeatable
parametric tests controlling available bandwidth
Nave, Chrisochoides, Lee, Deerfield, International Meshing Roundtable, Sept. 2004
21
Aerolab
Demo: So how easy is it?
#1. Create a topology
#2. Do something useful
22
Future Directions
• Different OSes, namely Solaris, real-time Linux and
Windows
• Incorporate Hardware emulators (IXP,SPIRNET)
• Incorporate passive network sniffers (layer 1)
• Wireless nodes, possible mobile campus nodes
• Merge with the Aerospace cluster (more nodes)
• Integrate 10 Gigabit capability
• Federate with other testbeds
• More nodes...more nodes...more nodes
23
Aerolab
Thank You!
Questions?
24