Speed up EMOD simulations¶
While small simulations can be run quickly on a local computer, the time and memory needed to complete a simulation can grow significantly when simulations become larger and more complex. For example, a single geographical node simulation might use 3 GB with one core in order to run successfully in one minute or less, while another simulation job may require 32 GB in a dual-core system in order to complete in approximately the same amount of time.
As your simulations grow, you will likely want to run simulations on an HPC cluster or take other steps to improve performance and reduce processing time. This topic describes many of the steps you can take to speed up EMOD simulations.
The EMOD executable (Eradication.exe) is “single threaded”, meaning that for processing, it will use only one core. However, you can use the Message Passing Interface (MPI) to run multiple copies of Eradication.exe in parallel, either locally or on an HPC cluster. To run a simulation in parallel, you must invoke Eradication.exe with the mpiexec command. For more information, see Run a simulation using mpiexec.
You can also set various parameters in the configuration file that will improve performance by scaling down the amount of data used or optimizing data processing. See the parameters listed in the sections below for guidance on making performance adjustments. See Configuration parameters for more information about each of the parameters.
Scaling and sampling¶
Obviously, simulating a shorter timespan will take less processing time. However, the processing time is often driven by the number of infections or immune updates, so running a simulation after all infections have cleared may not increase processing time much. For more information, see Simulation setup parameters.
Instead of the default Individual_Sampling_Type setting of “TRACK_ALL”, you can speed up performance by sampling such that each individual object represents more than one person. For example, a simulation with a population of 1 million and sample rate of 0.1 would simulate 100,000 individuals, with each given a weight of 10. Sampling can be fixed at a particular rate or can adapt the rate based on certain criteria, such as age or immune state. However, you should be especially careful not to undersample simulations to the point where they are overly sensitive to rare stochastic events. For more information, see Sampling parameters.
- Population_Scale_Type and x_Base_Population
Alternatively, you can simply reduce the total population of the simulation using Population_Scale_Type set to “FIXED_SCALING” and x_Base_Population set to less than one. For more information, see Scalars and multipliers parameters.
For large, spatially distributed simulations, running the intra-node dynamics (for example, infection and immune dynamics) in parallel on multiple cores may be very advantageous. Ideally, the timing would be reduced inversely to the number of cores. However, there are costs to serializing individuals for migration over MPI, as well as considerations for balancing the CPU load on each core. These issues may be mitigated using a load balancing input file that is suitable to the geography being simulated. See Load-balancing files for more information.
Vector and malaria¶
“VECTOR_SIM” will run much faster than “MALARIA_SIM” if detailed intra-host dynamics are not “required.
- Vector_Sampling_Type and Mosquito_Weight
The individual vector simulation (“TRACK_ALL_VECTORS”) will run more slowly than the cohort model (“VECTOR_COMPARTMENTS_NUMBER”). If the individual vector model must be used, you can improve performance by sampling some fraction of all vectors (“SAMPLE_IND_VECTORS”) where each mosquito has a weight configured by Mosquito_Weight, which is analogous to individual human sampling.
Using fewer intra-day updates to the malaria intra-host model will speed up simulation time. However, updates longer than three hours, such as eight per day, result in increasingly inaccurate dynamics.