|
In the News: Our team at Georgia Tech have developed some of the largest and
fastest network simulations reported to date. Our achievements have
recently gained good amount of press and publicity, including
Wired Online Magazine [11],
Georgia Tech Press [12],
Pittsburgh
Supercomputing Center News [13].
Here is a historical perspective on our parallel network simulations at
Georgia Tech.
Seed:
If willing to go sufficiently far back, history can be traced to a twin set of NSF
and DARPA projects 1996-99 (collaborative venture with many institutions
including Rutgers, Dartmouth, Univ. of Massachusetts, Bell Labs, Bellcore and of course Georgia
Tech). A few systems took life in those projects, including TeD, SSF and DaSSF.
Those systems presented pioneering performance (of that era), for high-fidelity
network simulations. A drawback of those systems was that they were developed
from scratch, and had a relatively limited user base in network research
community. Existing systems such as ns-2 had much larger base, but were limited
to sequential execution.
Insight:
Circa 1998, building on High Level Architecture experience, Richard Fujimoto
provided the first insight that a “federated approach”[1]
could potentially
provide a blend of rich user base and high-performance execution, for network
simulations. George Riley, under Dr. Fujimoto's & Dr. Ammar’s supervision,
developed a prototype, named pdns, in that direction -- a federated version of
ns-2, which is nothing but several instances of ns-2 integrated together to run
in parallel. This used a HLA-like RTI, called BRTI underneath[2]. Among the
early papers appearing on this federated simulations work include [3,
4]. We
(Ammar, Fujimoto, Perumalla and Riley) started using pdns in later NSF and DARPA
projects at Georgia Tech. Notable application was in the development of the
Dynamic Simulation Backplane[5,6,7].
Ramp Up:
Starting around 2002, we (Park and Perumalla) ventured to take pdns to the next
level, to simulating much faster (efficiently on up to 128 processors). We made
many performance improvements, especially inside the RTI layer to make it
competitive with hand-optimized monolithic parallel simulations. The largest
configurations executed at that time were on 128 processors of the Jedi cluster
at Georgia Tech[8].
Crescendo:
In early 2003, an even bigger challenge was posed to us in the DARPA NMS
project, to scale our simulations to supercomputing platforms. For this, I
highly optimized the communication/synchronization module (called libSynk[9])
used in our federated simulators, and ported it to the Lemieux supercomputer at
the Pittsburgh Supercomputing Center. I also exclusively undertook the
supercomputing challenge with pdns, and ported it to Lemieux.
As a milestone in
this effort, I was able to perform the fastest packet-level simulations to date
– 105 million packet transfers simulated per second[10], scaling up to 1536
processors for a single execution (simulating millions of network nodes).
More:
In late 2003, we experimented with additional improvements on large
configurations, using topology-aware enhancements to distributed time
management[14].
Separately, George Riley ported his new descendant to pdns, called GTNetS,
to Lemieux, and showed it to scale well with respect to network size (although,
GTNetS is a bit slower than pdns, but is less memory intensive, as
reported in [8]).
References:
- G. F. Riley, M. Ammar, R. M. Fujimoto, A. Park, K. S. Perumalla, and D. Xu,
"A Federated Approach to Distributed Network Simulation," ACM Transactions on
Modeling and Computer Simulation, vol. 14, pp. 116-148, 2004.
- R. M. Fujimoto, T. McLean, K. S. Perumalla, and I. Tacic, "Design of
High-performance RTI Software," presented at Workshop on Distributed Simulations
and Real-time Applications, 2000.
- G. F. Riley, R. M. Fujimoto, and M. A. Ammar, "A Generic Framework for
Parallelization of Network Simulations," in Proceedings of Seventh International
Symposium on Modeling, Analysis and Simulation of of Computer and
Telecommunication Systems, 1999.
- G. F. Riley, M. Ammar, and R. M. Fujimoto, "Stateless Routing in Network
Simulations," presented at Workshop on Modeling, Analysis and Simulation of
Computer and Telecommunication Systems, 2000.
- G. F. Riley, M. Ammar, R. M. Fujimoto, K. S. Perumalla, and D. Xu,
"Distributed Network Simulations using the Dynamic Simulation Backplane,"
presented at International Conference on Distributed Computer Systems, 2001.
- G. F. Riley, D. Xu, R. M. Fujimoto, and M. Ammar, "Split Protocol Stack
Network Simulation using the Dynamic Simulation Backplane," presented at
Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication
Systems, 2001.
- K. S. Perumalla, R. M. Fujimoto, T. McLean, and G. F. Riley, "Experiences
Applying Parallel and Interoperable Network Simulation Techniques in On-Line
Simulations of Military Networks," presented at 16th Workshop on Parallel and
Distributed Simulation, Washington, DC, 2002.
- K. S. Perumalla, A. Park, R. M. Fujimoto, and G. F. Riley, "Scalable
RTI-based Parallel Simulation of Networks," presented at Workshop on Parallel
and Distributed Simulation, San Diego, 2003.
- K. S. Perumalla, "libSynk Home Page", Last accessed 2004/05/01,
www.cc.gatech.edu/computing/pads/kalyan/libsynk.htm.
- R. M. Fujimoto, K. S. Perumalla, A. Park, H. Wu, M. Ammar, and G. F. Riley,
"Large-Scale Network Simulation -- How Big? How Fast?," presented at IEEE/ACM
International Symposium on Modeling, Analysis and Simulation of Computer
Telecommunication Systems (MASCOTS), 2003.
- M. Delio, "Net Analysis Gets Turbo Boost", Last accessed 2004/05/18,
www.wired.com/news/infostructure/0,1377,60077,00.html.
- G. T. News, "Georgia Tech Researchers Create the World’s Fastest Detailed
Computer Simulations of the Internet", Last accessed 2004/05/18,
www.gatech.edu/news-room/release.php?id=173.
- PSC News Center (Pittsburgh Supercomputing Center),
"Better Networks,"
www.psc.edu/publicinfo/2003/inprogress/.
- A. Park, R. M. Fujimoto, and K. S. Perumalla, "Conservative Synchronization
of Large-scale Network Simulations," presented at Workshop on Parallel and
Distributed Simulation, 2004.
|