The Experimental Computing Laboratory

Active Research Projects

• Parallel and distributed simulation and the WARPED Time Warp simulation kernel. Studying primarily the time warp synchronization strategy.

• Opportunities with Beowulf clusters with virtualizable processor nodes. Specifically we are looking at running application specific operating systems developed for fine-grained distributed computation such as that found in most Time Warp synchronized simulations. As a first step toward this end, we are currently focusing on accelerating the network latency delays that message encounter as they traverse the host/guest layers.

• Parallel simulation on multi-core and many-core Beowulf clusters. Two primary ongoing efforts:

  • Threading the WARPED code. Currently we are exploring and building multiple software architectures for deployment in multi-core and many-core environments.

  • Dynamic Frequency adjustement. Some multi-core/many-core processors have the ability to independenly control (with software) the frequency of a core (or group of cores). We are exploring the runtime control core frequencies in a threaded Time Warp simulator to accelerate the critical path and slow the off-critical path; ideally reducing rollbacks and potentially allowing other dynamic algorithms such as dynamic checkpoing to use settings that further reduce time warp overheads (e.g., reducing the checkpoint frequency and therefore the checkpointing costs).

• Revisit the old MIMD on SIMD work that we did on the old MasPar SIMD computers (see papers TPDS, LNCS #854) using GPGPUs.

• Explorations of non-traditional approaches to building/deploying parallel and distributed programming solutions to problems. An example of this is Time Warp cloning develoepd by M. Hybinette. What I hope is interesting about cloning is the opportunity to take simulation models that are run with multiple initial seeds and fold them into one large parallel simulation. What could be important about this would be to build the simulation so that the unchanging state of the multiple simulation runs is computed only once and shared by the parallel codes. Depending on how much of the state is unchanged between the multiple runs, substantial speedups may be possible.

  Ultimately we would like to depart from the traditional methods of building parallel programs (parallelizing sequential solutions) and explore the pursuit of parallelization solutions from other perspectives.

Previous Research Activities

• Explorations in the automatic/transparent exploitation of parallelism in multi-core/many-core processors. Function precomputation with argument predition.
• VHDL Related: Overview Picture
  • SAVANT: Standard Analyzer of VHDL
  • AIRE: Standard Intermediates for VHDL, Version 4.6 pdf
  • The TyVIS VHDL Simulation Kernel

List of Publications

Recent Theses and Dissertations

The Lab Creed

Programming Practices