MRIP - Multiple Replications In Parallel
Quantitative stochastic simulation is a useful tool for studying performance of stochastic dynamic systems, but it can consume much time and computing resources. Even with today's high speed processors, it is common for simulation jobs to take hours or days to complete.
Processor speeds are increasing as technology improves, but there are limits to the speed that can be achieved with a single, serial processor. To overcome these limits, parallel or distributed computation is needed. Not only does this speed up the simulation process, in the best case proportionally to the number of processors used, but the reliability of the program can be improved by placing less reliance on a single processor.
One approach to parallel simulation is to divide up the simulation model and simulate a part of it on each processor. However, depending on the nature of the model it can be very difficult to find a way of dividing it up, and if the model does not divide up readily, the gain from parallelising it will be less than proportional to the number of processors. Even in cases where the model can be parallelised easily, more work is required to implement a parallel version of the simulation than a serial one.
Akaroa takes a different approach to parallel simulation, that of multiple replications in parallel or MRIP. Instead of dividing up the simulation program, multiple instances of an ordinary serial simulation program are run simultaneously on different processors.
These instances run independently of one another, and continuously send back to a central controlling process observations of the simulation model parameters which are of interest. The central process calculates from these observations an overall estimate of the mean value of each parameter. When it judges that it has enough observations to form an estimate of the required accuracy, it halts the simulation.
Since the simulations run independently, if there are n copies of the simulation running on n processors they will on average produce observations at n times the rate of a single copy, and therefore produce enough observations to halt the simulation after 1/nth of the time. So the MRIP technique can be expected to speed up the simulation approximately in proportion to the number of processors used.
MRIP also provides a degree of fault tolerance. It doesn't matter which instance of the simulation the estimates come from, so if one processor fails, the program it was running can be restarted and the simulation continued without penalty. Alternatively, the simulation can simply be continued with one less processor and take proportionately longer to complete.
In summary, the advantages of the MRIP technique are that it can be applied to any simulation program without the need to parallelise it or modify it in any way; it provides a speedup proportional to the number of processors; and it improves the reliability of the simulation.
Last modified: 9 Feb 2011