I've been dabbling in the simulation of composite material laminates, but stumbled upon a result I can hardly explain. The same model (a quasi-static explicit decohesion simulation with cohesive surface) yields different maximum strengths when run in parallel or in a single processor.
It is very common that a parallel code gives *slightly* different results compared to a serial. This generally comes e.g. due rounding errors while building sums and products in different orders when bringing together the results of the parallel workers and due a different implementation of the calculation (make a serial algorithm parallel often requires a re-design). The 1st option also accounts to running a parallel code multiple times with the same input, the results will slightly differ but especially for time driven simulations the timelines will get away from each other the more simulation time is covered.
A simple example that illuminates the effect, assuming we can only save two digits behind the comma while building the product, in practice this rounding errors occur at the very last digits, depending on the machines precision and data types used:
Serial case has always the same order:
0.33 * 0.44 * 0.55 = 0.0825
Parallel cases can differ:
0.33 * 0.44 * 0.55 = 0.0825
0.55 * 0.33 * 0.44 = 0.0792
So anyway, one simulation is often only a snapshot out of a big ensemble of possible states with certain probabilities. Therefore, regardless of inaccuracies due parallelism, you might want to run the simulation multiple times and use an average value for your measure.
While I partly agree with Norman's answer, there could be other reasons, especially in case of Abaqus. Some analysis types do not support parallel execution. You need to check the warnings to see if there is any hint about that.
Also, to complete Norman's answer, try this with any programming language you like. Prepare a list of random floating point numbers, say 100. Now, try adding them together from the start of the list. Do that another time, from the bottom and compare the results. You'll see that there is a slight difference between them. This is because of round up errors. The answers are equally "correct" in the sense of approximate floating point math. Remember that these type of error propagates throughout your entire solution. If the difference you get in different executions is not that much, they might be because of this phenomenon.
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running