|
|
Parallel/Distributed Execution of Network Models
Until recently, simulation of network protocol models was limited by sequential execution to a few thousands of nodes. Our group's efforts have substantially increased this capability to a few million nodes, thereby opening up new experimentation approaches that were simply inconceivable before.
My efforts in this direction were focused on systems issues in scaling the parallel execution to large parallel/distributed platforms (e.g., supercomputers), and in efficiently performing time-synchronized communication required for correct results when discrete-event models are run in parallel.
See our research in the news for an account of the impact and recognition of our work.
Fidelity vs. Scalability
Below is my classification of existing methods of modeling network protocols and applications [reported in our ACSAC'04 paper]. It is clear from this picture that packet-level simulation represents the best trade-off between fidelity and scalability. It is precisely here that our large-scale network simulation work fits -- retaining the fidelity of packet-level yet increasing its scalability via efficient parallel execution.
Scalable Parallel Performance
The chart above shows results from some of my largest runs with an enhanced version of pdns on the Lemieux at Pittsburgh Supercomputing Center. As a result of a scalable time-synchronization algorithm that I developed, the synchronization overhead is kept low despite increasing the number of processors to over 1500. Please see our MASCOTS'03 paper for further details.
Applications of Large-scale Network Simulation
Application to Network Security
An important application of large-scale network simulation capability is in detailed simulation of packet-level operation of malicious applications (e.g., DDoS attacks, or worm propagation) and/or network defense mechanisms (e.g., monitors, filters). Our ACSAC'04 paper explores a few such approaches. In particular, the use of large-scale network simulation for testing Honeypot installations is shown to be feasible. This integrates large network simulations with operational (real) Honeypots via a technique called constructive emulation.
Application to On-line Network Optimization
Another important application is in an "on-line optimization" method for rapidly deployed complex networks, as shown in the picture below. We have demonstrated its applicability in some military network scenarios (rapidly exploring the effect of various future configurations, and choosing the best among the alternatives). Our PADS'02 paper documents additional details.
Visualization
To visualize large and complex networks rapidly, I developed a network visualizer & animator, called NetAnim. It is written in Java, and has been used to visualize very large networks (tested up to a million nodes). Here are some snapshots. Blue circles are subnet boundaries, yellow dots are network nodes, red lines are network links.
Related Publications
- High-Fidelity Modeling of Computer Network Worms
Kalyan Perumalla and Srikanth Sundaragopalan
Annual Computer Security Applications Conference (ACSAC), Tucson, AZ, December 2004.
[Paper][Slides]
- Conservative Synchronization of Large-scale Network Simulations
Alfred Park, Richard Fujimoto and Kalyan Perumalla
ACM/IEEE/SCS Workshop on Parallel and Distributed Simulation (PADS), June 2004.
[Paper][Slides]
- Large-Scale Network Simulation -- How Big? How Fast?
Richard Fujimoto, Kalyan Perumalla, Alfred Park, Hao Wu, Mostafa Ammar and George Riley
IEEE/ACM International Symposium on Modeling, Analysis and Simulation of Computer Telecommunication Systems (MASCOTS), October 2003.
[Paper]
- Experiences
Applying Parallel and Interoperable Network Simulation Techniques in On-line
Simulations of Military Networks
Kalyan Perumalla, Richard Fujimoto, Thom McLean and George Riley
ACM/IEEE/SCS Workshop on Parallel and Distributed Simulation (PADS), Washington, DC, May 2002.
[Paper][Slides]