
ISSN Online: 2688-7231
ISBN Online: 978-1-56700-524-0
Proceedings of the 26thNational and 4th International ISHMT-ASTFE Heat and Mass Transfer Conference December 17-20, 2021, IIT Madras, Chennai-600036, Tamil Nadu, India
Assessment of parallelization techniques for an Eulerian-Lagrangian multiphase flow algorithm
Abstract
Turbulent flows carrying a large population
of small, inertial particles are commonly observed in nature
and engineering applications. The complex interactions
between tiny particles and the carrier fluid turbulence
observed in such flows are resolved accurately, both in
space and time, using Eulerian-Lagrangian direct numerical
simulation (DNS). Nevertheless, these computations are
resource-intensive, and therefore, parallel computations are
necessary to cut down the computation time. This paper
presents various parallelization techniques implemented in
the in-house code MGLET-LaParT using Message Passing
Interface (MPI) to perform Eulerian-Lagrangian DNS of
particle-laden turbulent flows. A strong scaling study is
performed on Intel Xeon Silver 4116 CPU cluster ‘Pavan’
to assess the performances of implemented parallelization
techniques. Relative comparisons of the computation costs
of executing critical steps among various parallel algorithms
are presented. We perform simulations addressing two
different scenarios, uniform and highly non-uniform particle
distributions over the MPI subdomains, respectively. We
show that the parallelization based on the Eulerian domain
decomposition is the computationally most efficient strategy
for flow configurations that produce uniform particle
distribution over MPI subdomains. On the other hand, the
Lagrangian data decomposition based parallelization is an
appropriate strategy for flow problems that exhibit highly
non-uniform distribution of particles over MPI subdomains.