[Nek5000-users] Evaluate the validity of simulation results

nek5000-users at lists.mcs.anl.gov nek5000-users at lists.mcs.anl.gov
Thu Feb 16 16:26:14 CST 2012


Thank you very much for your detailed answers, Paul.


-Dong

-----Original Message-----
From: nek5000-users-bounces at lists.mcs.anl.gov [mailto:nek5000-users-bounces at lists.mcs.anl.gov] On Behalf Of nek5000-users at lists.mcs.anl.gov
Sent: Thursday, February 16, 2012 7:22 AM
To: nek5000-users at lists.mcs.anl.gov
Subject: Re: [Nek5000-users] Evaluate the validity of simulation results


Hi Dong,

I've been pondering this question for several days.

Some errors might be catastrophic, some might just be annoying,
and some might be totally benign.  It often depends on the problem.

1) Examples of catastrophic include:

    .Altering the velocity field in a stability calculation

    .Altering a vector in the middle of a solver so that
     conjugate gradient perceives the operator or preconditioner
     as not being symmetric positive definite.

2) Examples of annoying include:

    .Altering the velocity field in a turbulence computation, where
     the flow field is changed but rapidly pulled back to its
     average behavior.

3) Examples of benign include:

    .Altering a vector mid-solver but not driving the result from
     the ball of convergence (i.e., the error recovers and contracts).


The eddy problem is a good test case, given that we have an exact
answer.  It is not as sensitive as a stability computation (e.g.,
fs_hydro).

Paul




On Fri, 10 Feb 2012, nek5000-users at lists.mcs.anl.gov wrote:

> Hi guys.
> I am a computer science guy without any background in fluid dynamics solver.
> I am using Nek5000 for the fault tolerance research. In particular, I inject
> random errors into data structures of Nek5000 and then record how Nek5000
> reacts. Currently, I have a problem about how to judge the performance of
> Nek5K with injected faults is normal (i.e., tolerating fault) or not. I can
> simply record the execution time and then compare time difference. However,
> some injected errors might not lead to time variance while making the
> simulation result invalid. Is there any metric I can use to evaluate if the
> simulation result is valid/reasonable (especially for the eddy and/or vortex
> problem)? For example, a X error rate beyond a threshold after a fixed number
> of timesteps, or something?
>
> Thank you very much.
>
> -Dong
>
>
_______________________________________________
Nek5000-users mailing list
Nek5000-users at lists.mcs.anl.gov
https://lists.mcs.anl.gov/mailman/listinfo/nek5000-users



More information about the Nek5000-users mailing list