Summary

The SWANS simulator runs over JiST, combining the traditional systems-based (e.g., ns2) and languages-based (e.g., GloMoSim) approaches to simulation construction. SWANS is able to simulate much larger networks and has a number of other advantages over existing tools. We leverage the JiST design within SWANS to:

• achieve high simulation throughput: Simulation events among the various entities, such as packet transmissions, are performed with no memory copy and no context switch. The system also continuously profiles running simulations and dynamically performs code inlining, constant propagation and other important optimizations, even across entity boundaries. This is important, because many stable simulation parameters are not known until the simulation is running. Greater than $10\times$ speedups have been observed.

• save memory: Memory is critical for simulation scalability. Automatic garbage collection of events and entity state not only improves robustness of long-running simulations by preventing memory leaks, it also saves memory by facilitating more sophisticated memory protocols. For example, network packets are modeled as immutable objects, allowing a single copy to be shared across multiple nodes. This saves the memory (and time) of multiple packet copies on every transmission. A different example of memory savings in SWANS is the use of soft references for storing cached computations, such as routing tables. These routing tables can be automatically collected, as necessary, to free up memory.

• run standard Java applications: SWANS can run existing Java network applications, such as web servers and peer-to-peer applications, over the simulated network without modification. The application is automatically transformed to use simulated sockets and into a continuation-passing style. The original network applications are run within the same process as SWANS, which increases scalability by eliminating the considerable overhead of process-based isolation.

In addition to the simulator design, it is also essential to model wireless signal propagation efficiently, since this computation is performed on every packet transmission. The hierarchical binning data structure is allows node location updates in expected amortized constant time and receiver node set computations in time proportional to the number of receivers. The combination of these attributes leads to a flexible and efficient simulator. We hope that the performance of SWANS will facilitate further research into scalable ad hoc network protocols.