Keywords: computational fluid dynamics, parallel computing, scalability, rotating mesh.

Computational fluid dynamics (CFD) has advanced over the last thirty years to the point whereby simulations with million node meshes modelling complex geometries with a comprehensive suite of physics is now commonplace. Moreover, code parallelisation strategies based upon domain, (more precisely, mesh) partitioning are now well established to the point where commercial CFD software tools all include parallel capabilities, which run scalably on parallel computing systems with adequate communications speeds. These approaches were enabled through the emergence of:

a)

message passing tools such as PVM and MPI, which is now the predominant method for exchanging data between processors, and

b)

mesh partitioning tools to enable the distribution of compute effort across the parallel cluster

The mesh partitioning approach to CFD code parallelisation implicitly assumes that the compute load at each element of control volume is identical, and where this is the case codes scale well. However, there are a range of physical flow phenomena where the compute load cannot be guaranteed to be uniform across the mesh, and it is under these conditions that scalability might become significantly limited. Flows through domains with rotating machinery have presented a range of challenges to the CFD community over the last decade or so. There are two key complexities introduced by rotating machinery - the introduction of momentum into the flow field at specific locations, and the loading of the flow on the rotating structure as well as the details of the flow interacting with it. In the paper we consider the three conventional ways to capture the impact of rotating machinery on flows:

a)

The source-sink approach, which essentially evaluates the gross impact of the rotating structure at a specific location and adds this into the momentum equations as a source.

b)

The no-inertial rotating reference frame approach captures the flow relative to the rotating equipment.

c)

The multiple reference frame approach, where part of the flow domain is rotating with the equipment and typically the rest of the mesh is static.

In this paper attention is then focussed upon a straightforward strategy to enable the multiple reference frame approach to run in parallel. In the scheme the element adjacencies change on the interface between parts of the mesh influenced by different reference frames. In order for this technique to run in parallel there is a need to extend the communication schedules so that all data associated with near interface geometry is available on all processors. Although this is safe in computational terms, it is also potentially expensive unless the number of elements bounding the rotating surface is small compared to the overall mesh size.

All the above strategies have been implemented within the CFD module of the multi-physics code PHYSICA, which employs a fully unstructured mesh of any mix of elements up to and including hexahedra together with conventional finite volume solution procedures for multi-component multi-phase free surface reacting flows. The code has been parallelised using a conventional mesh partitioning approach employing the JOSTLE partition tool and the CAPLIB message passing library. Using the above CFD technology the rotating equipment modelling strategies referred to above are all evaluated with regard to their parallel scalability. They are all shown to be surprisingly scalable in parallel, so long as the ratio of the number of elements bounding the rotating reference frame is small compared to the total mesh size.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £72 +P&P)