[vickmnit]

I have used Novacast, OpenFOAM, and one in house developed Academic CFD code to simulate 0.1 million mesh Casting flow problem. While Novacast (ran on single core) took only less than an hour to simulate flow, the OpenFOAM (single core), and Academic code (MPI enabled ran on 4 cores) took about 5-6 hours to complete the simulation. What could be the reason?

[FMDenaro]

I do not agree ...usually, it was my experience that own made code are less flexible but faster than commercial codes (of course, comparing the same numbers of process on same nodes, etc.)

[vickmnit]

Which means there is something wrong with my implementation or measurement in performance even with OpenFOAM. I will look into this again to confirm the case since other sources also seems to agree with your observation. Thanks for the prompt reply.

[mprinkey]

You also need to be careful to assess what models are being used. I don't know much about Novacast, but I know Barracuda was a similar package that used MP-PIC to simulate casting behavior. MP-PIC, as a modeling scheme, is significantly faster than, say, two-fluid Eulerian simulations just because of the simplifying assumptions being made by MP-PIC. It doesn't make either simulation better or worse...just different. Suitability of each for your simulation task is determined by how detrimental or advantageous those simplifications are.

[sbaffini]

As stated by Michael and Filippo, such kind of test requires the exact same model (or, at least, almost the same) and the exact same hardware. And even in that case, you may notice some differences on different hardwares (a code which is faster on hardware A ends up being slower on hardware B).

However, for how fair this statement can sound, it still hides a lot of details, including, in my opinion, your very first hypothesis.

Typically, academic-research codes are perceived as faster (and mostly are) because they actually solve a different model, where model is really intended in a very broad sense. For example, an academic code might typically use a structured grid and a specific physics implementation which is very fast (and suitable) only for that kind of grids. For example, some pressure poisson equations can be solved exactly and efficiently on some structured uniform grids. The equivalent workhorse for general purpose CFD solvers on unstructured grids is the Algebraic Multigrid, which still can't compete with a fast poisson solver. The underlying equation might be the same, even the discretization for that specific grid, but would you say that the comparison is fairif the very algorithm is different?

Additional topics are robustness and flexibility, algorithmic as well as to user input. Academic-research codes are just not into that business. Most of the commercial business revolves around corner cases, which are not typically part of the picture for non-commercial codes.

Moreover, commercial codes have to (or at least should) respect some programming best practices that academic-research codes tend to neglect (consciously or not).

Finally, there is the compiler/hardware issue. For a given model and a given algorithmic implementation, there is typically room to up to 20x improvements by simply using a better compiler for a given architecture, including all the specific flags for vectorization etc. This might not be common for academic codes, but could be for commercial ones.

Having said so, my experience however is quite the opposite. Commercial codes, when fairly compared, tend to be faster. This, in my experience, comes from the fact that a commercial code developer can allocate resources to work out any single aspect that is considered fruitful to the selling process. And being fast is typically what everyone requires (together with robustness and flexibility).