Performance considerations

This section provides some recommendations for sizing the resources of the Simcenter Amesim job in order to optimize the global load of the entire HPC platform.

Sizing the resource request

resources requirement

For best integration of your jobs on your HPC platform, it is important to understand the model of parallelism of all types of Simcenter Amesim study.

For a single run, it is not necessary to allocate multiple nodes and multiple cores. Single runs are not parallel-processed. Nonetheless, single runs can take advantage of running on an HPC platform:

• The processors are more efficient for numerical applications (better arithmetic calculus) than employee desktops/laptops.

• Better I/O performance.

• Faster memory access, larger cache memory.

Up to 40% time savings have been observed between standard professional laptops and HPC nodes.

Parallelism models

Simcenter Amesim parallel studies follow one of the two parallelism models described hereafter.

• Embarrassingly parallel

Batch run, Design of Experiments and Monte Carlo algorithms have the same parallelism model. They correspond to a set of independent tasks, where only the input parameters will change. Most of the time, the number of runs is high. As it is a good candidate for parallelized computation, a linear speed up is expected. Linear speedup means the time gain is proportional to the amount of resources used.

• Iterative method

The Optimization algorithms available in Simcenter Amesim are iterative methods, where the parallelism is limited to one iteration execution. The nth iteration needs inputs from the previous iteration to proceed. However, for the genetic algorithm, the runs within a population corresponding to one generation are independent and are thus candidates for parallel distribution.

Recommendation for different study types

For batch run, time to solution will decrease linearly according to the number of cores allocated to the jobs. For Monte Carlo simulations, similarly to batch studies, the performance will follow a linear speed up (until the number of runs equal number of cores). That means time to solution will decrease linearly according to the number of cores allocated to the jobs. For Design Of Experiments, the size of the design matrix, corresponds to the number of runs.

Both optimization algorithms within Simcenter Amesim are iterative methods. For NLPQL, the optimal number of cores is limited by the number of parameters, which corresponds to number of computations to estimate the gradient at each step. For the genetic algorithm, the limit corresponds to the population size. Please consult the Design Exploration manual for more details

Scalability is generally better for the genetic algorithm, the population size (of several hundred) is bigger than the number of parameters in NLPQL (less than 10).

The following tab summarizes the sizing recommendations for Simcenter Amesim studies. If the requested resources are greater than this limit, the job will not take advantage of the extra resources available, so the reservation is over-dimensioned. These extra and unused resources will consume your CPU hour budget.

Remarks: Scalability in practice should be lower than the theoretically projected performance, due to the computing architecture or a high working load on the cluster.

Study type	Upper bound speed up limit
single run	1
batch run	nb of batch runs
Design of experiments	size of the design matrix
Monte Carlo	number of runs
Optimization: NLPQL	number of degrees of freedom
Optimization: genetic algorithm	population size

Figure 26: Speed up versus number of resources

Performance factor

The performance of the simulation will be maximized if the working directory is placed on the fastest shared storage of your platform. Ask your system administrator for more details.