Simulator

Our results are based on a simulation engine written for a compiler of the data parallel language UC[Bag] on the IBM-SP2. The compiler translates UC code into Maisie[Bag94] with MPL calls for fast communication and synchronization over the HPS (high performance switch). The resulting program executes as one single-threaded process per processor. The simulation engine is simply a modified version of the same compiler, which:

• Also produces Maisie code. In the simulator, Maisie additionally provides lightweight threads (in order to put many logical simulation processes on one processor) and point-to-point communication between threads (whether they are on the same or different processors).
• Inserts code to calculate local code execution time at the beginning and end of each local code block, and to update simulation time.
• Substitutes MPL communication routines i.e. broadcasts, reduces, barriers and point-to-point communication with the respective communication simulation routines. For example, we discovered the tree algorithm used in the IBM-SP2 for broadcasts, barriers and reduces. The same algorithm is used to simulate each of those routines using point-to-point communications. It is at all these points that the simulation optimizations discussed in the previous section are inserted. Point-to-point message delays are predicted using an analytic model based on linear interpolation and extrapolation of measurements of delays for selected message sizes.