Agent-based simulations are gaining popularity in the social sciences, also in economics. As such simulation models grow in size and features, there is a growing need that these simulations are executed in parallel. We will study the performance behaviour of a parallelized agent-based model, the so called “benchmark” model by Caiani et al. (2016), implemented in the FLAME framework. The parallel performance of the model is tested on a commodity hardware laptop, a multicore machine and the Vienna Scientific Cluster1. We find that the performance of the simulation is a) poor, especially when MFLOPS per second are used as benchmark but also that b) super-linear speedups are possible. We trace both these finding back to an inefficient design of how the agent’s data is read and stored. An exaggerated amount of data is transferred to the CPU, which makes the application I/O bound and leads to the observed poor performance. Furthermore, when the simulation is executed on more processes, data transfers are better distributed leading to the super-linear speedups.
Masterprüfung mit Defensio, Reiter Oliver
22.10.2018 11:00 - 12:30
Organiser:
SPL 5
Location:
Besprechungsraum 4.34
Währinger Straße 29
1090 Wien
1090 Wien