Symposium for promotion of Indegenious CFD in Engineering Sciences (SPICES'09)

Venue: Computational Research Laboratories (CRL), Pune
Date: July 11, 2009
Participants: ADA, DRDL, CRL, SandI.

Focus

A national initiative to evaluate indigenous CFD tools for various features such as accuracy, efficiency, robustness and scalability.

Features evaluated

Accuracy, efficiency, robustness and scalability
  1. Accuracy: To assess the accuracy of CFD codes by comparing the integrated lift and drag coefficients predicted by codes with the experimental results as well as by assessing capability of the code to predict CLmax and αmax.
  2. Efficiency: The purpose of the efficiency studies is to bring out the capability of the CFD codes to generate the required design data quickly on given high performance computing platform. For this purpose the residue histories and convergence of the integrated force coefficients plotted against computer time were sought.
  3. Robustness: The robustness studies involves presenting the sensitivity of the convergence parameters to α sweep and the algorithmic scalability of the code. The algorithmic scalability means that the numerical performance of a CFD code does change due to parallelization.
  4. Scalability: Scalability measures the ability of a CFD code to exploit fully the parallel architecture of a given massively parallel machine. Under this studies the speed-up curve and the parallel efficiency are presented.
     

 

Configuration

NASA TRAP WING, a three element high lift configuration consisting of body pod (fuselage), leading edge slat, main element and full-span flap.
Configuration

Inferences

This study has clearly demonstrated

  1. The accuracy of HiFUN in generating the required design data for this class of problems.
  2. The efficiency of the HiFUN code in generating the entire drag polar in less than a day.
  3. The robustness of HiFUN for the entire alpha sweep while at the same time the serial performance of the code is not compromised even on large number of processors.
  4. The superior scalability of the HiFUN code in exploiting the full potential of massively parallel computer like EKA.

Grids

Grids were provided by the workshop technical committee. Two unstructured grids, a coarse with 12 million volumes referred to as UG1 and the other fine with 38 million volumes referred to as UG2 were provided by the organizers.

UG1 UG2
Unstructured hybrid grid with prismatic and tetrahedral elements

Free stream conditions

Mach number 0.2
Reynolds number 4.3 million
Angle of attack −4 ° to 38 °

Solver settings

Fully turbulent flow simulations were performed.

Results

Accuracy study

Accuracy study for HiFUN is carried out for the entire α sweep.

CL vs α CD vs α
Comparison of Force Coefficients

 

  αmax CLmax
Experiment 32.993 3.0306
UG1 32 2.9735
UG2 33 3.0565
  1. Excellent comparison of lift and drag curves with experimental values on both the grids.
  2. On both the grids, HiFUN captures the slope discontinuity in the lift curve at α = 3 °.
  3. While HiFUN marginally underestimates CLmax and αmax on coarse grid UG1, it is very close to experimental values on fine grid UG2.
Efficiency study

Efficiency study for HiFUN is carried out on the grid UG2 at post-stall angle of attack equal to 34 ° using 2048 processors on EKA.

Residue CL & CD Pressure fill plot
Efficiency Studies, α = 34 °
  1. In HiFUN, convergence is declared when the relative residue of density/energy falls by eight decades and the change in lift and drag coefficients over 100 iterations is less than 1 count (1 lift count = 0.001 and 1 drag count = 0.0001)
  2. Solution convergence for this run is achieved in about 135 minutes/2.25 hours.
  3. Therefore, this study amply demonstrates the efficiency of HiFUN in generating the entire drag polar with about 10 points in less than a day on a machine like EKA.

Robustness study

In this study, robustness of a code is evaluated by studying the sensitivity of convergence of HiFUN to α sweep and algorithmic scalability on the fine grid UG2.

Sensitivity of convergence to α sweep

α = 0 ° α = 2 ° α = 4 °
 
α = 28 ° α = 32 ° α = 34 °
Sensitivity for varying angle of attack
  • Owing to a robust implicit procedure incorporated in HiFUN, it can be seen that 8 to 10 decades of fall in relative residue is achieved for all the angles of attack.

Algorithmic scalability

In order to demonstrate algorithmic scalability of HiFUN, it is run on UG2 with two sets of processors, namely, 1024 and 2048.

Density residue Energy residue CL & CD
Algorithmic scalability, α = 34 °
  1. Although there is marginal change in the convergence pattern between 1024 and 2048 processors after 7 decades of residue fall, the integrated force coefficients do not show any perceptible difference in the convergence. It should be remarked that these marginal differences in the convergence history are observed only in the post-stall region and along the linear leg of the lift curve, the code shows near identical convergence on varied number of processors.
  2. This study demonstrates that HiFUN exhibits high levels of algorithmic scalability.

Scalability

Parameters describing scalability or parallel performance of code are speed up and parallel efficiency. The definitions of these parameters are as follows:

Actual speed up: It is defined as the ratio of time per iteration on reference set of processors to the time per iteration on given set of processors.
Ideal speed up: It is defined as the ratio of number of processors in given set to number of processors in reference set.
Parallel efficiency: It is defined as the ratio of actual speed up to ideal speed up.

Speedup parallel efficiency
Scalability on UG1
  1. The reference set of processors for grid UG1 is 8
  2. For grid UG1, the code HiFUN exhibits super-linear speed up till 512 processors and near ideal speed up for 1024 processors.
Speedup Parallel efficiency
Scalability on UG2
  1. The reference set of processors for grid UG2 is 64.
  2. For grid UG2, HiFUN exhibits super-linear speed up till 1024 processors and near ideal speed up for 2048 processors.

The scalability study has amply demonstrated that HiFUN exhibits high level of parallel performance on large sets of processors and can fully exploit the capability of massively parallel machines like EKA.

References

Ravindra K, Nikhil Vijay Shende, N Balakrishnan, ``Performance of the code HIFUN in SPICES–09", Proceedings of AeSI 2009 CFD Symposium, 2009, Bangalore, India

AeSI-2009 presentation: PERFORMANCE OF CODE HIFUN IN SPICES 09

SPICES09 presentation: Performance evaluation of the flow solver HIFUN using NASA TRAP wing configuration
.