RealTime OnLine Network Simulation

1
Real-Time On-Line Network Simulation

• PIsBolek Szymanski, Chris Carothers, Shiv
  Kalyanaraman Ken Vastola
• RAs Paul Belemjian, Hema Kaur, Yu Liu, Kiran
  Mandani, Manoj Mehta, Nick Lessard, Anand Sastry,
  B. Sikdar Tao Ye,
• Rensselaer Polytechnic Institute, Troy, NY
• http//www.cs.rpi.edu/szymansk/sonms.html
• email szymansk_at_cs.rpi.edu
• DARPA PI Meeting
• April 2, 2001

2
Novel goals of the research

• On-Line Network Modeling and Simulation scalable
• to multiple domains and hundreds of thousands
  of flows
• Second order traffic and routing control

Topic of this poster
Experiment Design
Network Abstraction And Decomposition
Parallel Discrete Event Simulation
Performance
Processor 1
Processor 2
Domain 2
Domain 1
Parameter 2
Parameter 1
P1min
P1max
Processor 3
Processor 4
Current operating point
Domain 3
Trial operating point triggering simulation
router
Link, simulated at packet level
Models of Inter-domain flows
Links crossing processor Boundary may cause
rollback
All three trial points can be Evaluated
concurrently
3
Real-Time On-Line Network Simulation

• Space decomposition partition large network into
  disjoined individual domains, each simulated
  independently and concurrently with others.
• Time decomposition partition simulation time
  into separate intervals, each interval iterated
  over until all domain simulators converge.
• Synchronization exchange packet delay and loss
  information on flows originated externally to
  each domain at the end of each interval
  simulation (iteration). Message passing via
  sockets is used in farmer-worker parallel
  architecture.
• Basic domain simulation uses currently ns to
  support portability of the results.

4
Global view - abstract configuration
5
Extensions to ns

• Domain definition in network simulation
  definition (Tcl script)
• Fake source and fake link definition to
  represent inflow and outflows to the domain and
  account for packet delay and drop outside the
  domain
• Checkpointing of the simulation state to enable
  iterations over time intervals
• Freeze event to enable synchronizing simulations
  and exchange of data at the end of each iteration

6
Concurrent simulations of domains freeze,
exchange of messages, checkpointing
7
Experiments
64-node (above) and 27-node (right) configurations
8
Advantages of the Approach

• Efficiency execution time t(n) of network of
  size n is growing faster than n
• nlogn term from processing future event queue
• nn term from processing routing
• An iteration with n processors, each running a
  domain of 1/n of nodes run faster than 1/n of
  entire network simulation time.
• Fault tolerance if a domain processor fails,
  the data from last iteration can be used
• Integration of models into simulation a cloud
  of unknown structure could be represented by path
  delays and packet drop probabilities
• Full processing distribution processors
  simulating each domain can be located in the
  domain

9
Whats Next?

• Improvements in implementation
• Synchronization in a tree-like structure
• Checkpointing to local disks
• Aggregating external sources into single external
  link
• Experiments with TCP traffic
• Integration with fast domain simulators (ROSS)
• Linking with an On-line Data Collection
• Integration with Experiment Design component for
  network management

10
ROSS Rensselaers Optimistic Simulation System

• ROSS demonstrates that highly efficient
• execution is possible when using little
• optimistic memory.
• Extreme Performance
• 1,250,000 events/sec, 4 PE case
• uses COTS PC hardware
• Low Memory
• less than 1 optimistic memory for large-scale /
  low event grain models.
• Target Application
• very low event granularity models
• wireless / packet-level network models

• ROSS performance is achieved by...
• optimistic synchronization
• Pointer-based, modular implementation framework
• Reverse computation
• Fujimotos GVT algorithm
• Kernel Processes (KPs)

As a demonstration of ROSS performance, we have
conducted an initial comparison with ns.
11
ROSS Embedded Capabilities

• A version of ROSS current runs as an embedded
  system inside the Linux OS (directly linked).
• User programs invoke ROSS thru the ross system
  call.
• Results and config parameters are pass thru
  system call.
• New capabilities
• allows for parallel simulations to be embedded
  into network elements.
• allows for fine grain control of simulator CPU
  resources.
• allows direct access to OS level network
  performance statistics
• improves simulator performance by 10 to 15 over
  user-space parallel performance.
• potential for improving stability of optimistic
  synchronization.

12
Future Work on ROSS/TCP Model...

• Strong validation between ns and ROSS/TCP models
  across a wide rang of configurations
• Implement RED into TCP model.
• Implement PGM model on-top of TCP/IP model
• Flexible specification of network topology.
• Experiment with ROSS as an embedded simulation
  environment.

13
Experiment Design Goals

• Problem definition Search a large parameter
  state-space
• Expected search method properties
• Exponential improvement rate
• Emphasis not on full optimization but finding a
  better operation point soon

Rastrigin test function many local optima
14
Heuristic Search Algorithms Structure
Explore examine new, unknown regions of the
state space. E.g., random sample, random
walk. Exploit attempt to quickly converge to the
optimum of a region of interest, e.g.,
hill-climbing, pattern search Balance Strategy
adjust the computing resource allocation between
explore and exploit processes.
15
Hybrid Search Algorithm

• Generic heuristic algorithms sacrifice
  performance for wide applicability
• Eg Genetic algorithms, Simulated Annealing
• Hybrid Search Algorithm
• Combines multiple search techniques
  hill-climbing, tabu search, simulated annealing,
  etc.
• Dynamically adjust the balance between exploit
  and explore
• Automatic learning and aggressive exploit of
  response surface features

16
Test Results for Hybrid Algorithm
Schwefel's (Sine Root) test function multiple
similar local optima
Average performance Hybrid algorithm converges
faster and better.
17
Unified Search Architecture

• Optimization find as good a result as possible
  within the minimum time using limited computing
  resources
• No Free Lunch Theorem a search algorithm is
  efficient only for a certain class of search
  spaces.
• Unified search architecture
• Allows dynamic combination of multiple search
  techniques
• 3-dimensional dynamic balance strategy
• Explore, exploit, resource-management

18
Unified Search Architecture
Exploiter 1 20
Exploiter 2 10
...
Explorer 1 40
Explorer 2 0
Sample Space
(Memory)
...
19
On-line Simulator Data Flow
Console/Monitor
Control
Command
On-line Simulator
Network
Configuration
w
Testbed
w
Simulation Script
Statistics
Experiment
Result
Farmer
...
...
Worker
20
Future Work

• Modularize other explore and exploit techniques
  and integrate them into unified architecture
• Provide programmable interface for dynamic
  balance strategy so as to easily synthesize new
  search strategies with a basic set of building
  blocks (explore and exploit methods)
• Deploy architecture on a scalable computation
  infrastructure
• Investigate on more advanced balance strategy to
  optimize the utilization of computing resource
• Integration with other parts to accomplish
  scalable network management

21
Routing Management Using Online Simulation
Previous work

• Goal To demonstrate improvement in routing
  performance using Online Simulation without
  modifying existing routing algorithms
• Demonstrated in simulation that parameter tuning
  can significantly improve performance
• Potential effect of parameter tuning
• Increased throughput (10-20)
• Lower end-to-end delay (20-35)
• Reduced number of routing updates (up to 90)
• Demonstrated Online simulation conceptually
• Conclusion Routing update flooding and route
  flapping have significant impact on performance
• New Goal Improve routing without excessive
  routing updates
• Output of online simulation to change interface
  costs
• Routing with new interface costs would converge
  to a desired set of routes
• Demonstrate Online Simulation in a real Testbed
  Network

22
Routing Management Using Online Simulation
Current Work

• Developed a new scheme for stable
  congestion-sensitive routing
• New scheme is based on Additive Increase
  Multiplicative Decrease cost metric on congested
  links with diminishing increments
• Demonstrated Online Simulation in a test network
  of Linux routers
• Non-trivial test network topology for
  demonstrating routing management
• New stable scheme for congestion sensitive
  routing is used in simulation
• Online simulation is used to obtain
  ospfIfMetricValue for various interfaces in the
  network
• OSPF converges to desired set of routes when new
  values of ospfIfMetricValue are deployed in the
  network

23
Routing Management Using Online Simulation
Future Work

• Test and analyze the dynamics of proposed stable
  congestion-sensitive routing scheme
• Use online traffic monitoring and modeling
• Investigate the impact of RED on congestion
  sensitive routing with TCP traffic
• Use prediction using self-similar traffic models
  to relax the stationary assumption in congestion
  sensitive routing
• Validate online simulation in simulation using
  routing at multiple time scales in MPLS for
  setting up failover Label Switched Paths (LSPs)

24
Traffic Modeling Current Work

• Our previous work identified timeouts and
  exponential backoffs as cause behind TCP traffic
  self-similarity
• We used simulations to verify that short TCP
  flows can also lead to self-similarity (our
  analysis used infinite flows)
• Conducted statistical tests on real network
  traces to verify the assumptions made by our
  analytical model
• Using simulations we show that keeping a buffer
  to number of flows ratio of 2-3 is sufficient to
  prevent self similarity
• Marking retransmitted packets as higher
  priorities can also prevent timeouts

25
Traffic Modeling Future Work

• We are developing efficient ways of simulating
  both individual as well as aggregate TCP flows on
  a link
• We are also investigating the effect of
  aggregation and developing mathematical models to
  characterize aggregate flows parsimoniously
• Currently working on using purely Additive
  Increase Multiplicative Decrease (AIMD) protocols
  as a means of reducing traffic self-similarity
• We are investigating ways to prevent bursty
  losses Active queue management techniques