Parallel simulations

HetaSimulator supports parallel simulations on a single machine and in a distributed (cluster) environment. It can be achieved by setting up Julia workers. Distributed package can be used to start workers on a local machine and ClusterManagers package supports a number of job queue systems (SGE, PBS, HTCondor, etc). In the following examples we will use Distributed package to start workers. HetaSimulator implementation of parallel simulations relies on SciML Ensemble Simulations features and inherits EnsembleAlgorithms choice. Parallelization algorithm is defined by parallel_type keyword argument. Currently it supports the following options:

• EnsembleSerial() - No parallelism. The default.
• EnsembleThreads() - This uses multithreading. It's local (single computer, shared memory) parallelism only.
• EnsembleDistributed() - Uses pmap internally. It will use as many processors as you have Julia processes.

Parallelization types

Simulating Scenarios in Parallel

Assuming we have loaded a number of Scenarios: scn1, scn2, scn3 we parallelize simulation or fitting procedures.

using Distributed
addprocs(2)
@everywhere using HetaSimulator

s = sim([scn1, scn2, scn3], parallel_type=EnsembleDistributed())
f = fit([scn1, scn2, scn3], [:k1=>0.1,:k2=>0.2,:k3=>0.3], parallel_type=EnsembleDistributed())

Parallel Monte Carlo Simulations

We can run parallel Monte-Carlo (Ensemble) simulations with parameters taken from distributions or from a pre-generated DataFrame. Parallel setup can work both with parameters vectors and scenarios. Let's assume we have loaded a number of Scenarios: scn1, scn2, scn3 and a DataFrame df with parameters vectors (as rows).

using Distributed
addprocs(2)
@everywhere using HetaSimulator

s1 = mc(scn1, df, 150, parallel_type=EnsembleDistributed())
s2 = mc([scn1, scn2, scn3], [:k2=>Normal(1e-3,1e-4), :k3=>Normal(1e-4,1e-5)], 150, parallel_type=EnsembleDistributed())