[Nek5000-users] Performance problem

[nek5000-users at lists.mcs.anl.gov]

If you set p94 & p95 to 5 (say), I recommend strongly to set p12 (dt) to

    -3e-5

in your case.   The reason for this is that the projection scheme is much
more stable for fixed dt.

On the whole, however, Stefan's earlier comments about using, say, lx1=8
and fewer elements is a better strategy.   It's also possible that we 
should switch your coarse-grid solve at this scale to AMG.

Paul


On Fri, 6 Aug 2010, nek5000-users at lists.mcs.anl.gov wrote:

>
>
>> - set param(102) and param(103) to 5 (this will turn on the residual 
>> projection)
>
>    Should this be param 94 & 95 ?
>
>    - Paul
>
>
>
>
> On Fri, 6 Aug 2010, nek5000-users at lists.mcs.anl.gov wrote:
>
>> Ok here are some suggestions to improve the performance:
>> 
>> - set timestep, param(12), to -3e-5
>> - set param(102) and param(103) to 5 (this will turn on the residual 
>> projection)
>> - increase lgmres (in SIZE) to 40
>> - you have want to tune the Helmholtz (velocity) and pressure tolerance 
>> (e.g. 1e-8 and 1e-5)
>> 
>> btw: what's the Reynolds number of this flow?
>> 
>> 
>> Stefan
>> 
>> 
>> On Aug 6, 2010, at 11:13 AM, <nek5000-users at lists.mcs.anl.gov> 
>> <nek5000-users at lists.mcs.anl.gov> wrote:
>> 
>>> Dear Mani,
>>> 
>>> I haven't checked your logfile yet but there are my first thoughts:
>>> 
>>> N=4 is low
>>> Your polynomial order (N=4) is low and the tensor-product formulation 
>>> won't buy you much. The performance of all matrix-matrix multiplies (MxM) 
>>> will limited by the memory access times. This is in particular a problem 
>>> on multi-core and multi-socket machines. We have seen that the performance 
>>> drop can be significant.
>>> On top of that you carry around a large number of duplicate DOF and your 
>>> surface to volume ratio is high (more communication). I
>>> 
>>> 
>>> Parallel Performance
>>> Your gridpoints per core (~4700) is quite small!
>>> On Blue Gene (BG) systems we can scale well (e.g. 70-80% parallel 
>>> efficiency) with around 10k gridpoints per core. On other system (e.g. 
>>> Cray XT5) you need much more gridpoints per core (say 80k) because the 
>>> network has a higher latency (NEK is sensitive to latency not bandwidth) 
>>> and the processors are much faster.
>>> 
>>> Cheers,
>>> Stefan
>>> 
>>> On Aug 6, 2010, at 10:51 AM, <nek5000-users at lists.mcs.anl.gov> wrote:
>>> 
>>>> Hi,
>>>>
>>>>   I'm solving for Rayleigh-Benard convection in a 3D box of 37632, 4rth 
>>>> order elements. I fired the job on 512 processors on a machine with 
>>>> quad-core, quad socket configuration (32 nodes with 16 cores each ) with 
>>>> a 20 Gbps infiniband interconnect. In 12 hours it has run 163 time steps. 
>>>> Is this normal or is there maybe some way to improve performance? 
>>>> Attached is the SIZE file.
>>>> 
>>>> Regards,
>>>> Mani chandra
>>>> 
>>>> <SIZE.txt>_______________________________________________
>>>> Nek5000-users mailing list
>>>> Nek5000-users at lists.mcs.anl.gov
>>>> https://lists.mcs.anl.gov/mailman/listinfo/nek5000-users
>>> 
>>> _______________________________________________
>>> Nek5000-users mailing list
>>> Nek5000-users at lists.mcs.anl.gov
>>> https://lists.mcs.anl.gov/mailman/listinfo/nek5000-users
>> 
>> _______________________________________________
>> Nek5000-users mailing list
>> Nek5000-users at lists.mcs.anl.gov
>> https://lists.mcs.anl.gov/mailman/listinfo/nek5000-users
>> 
> _______________________________________________
> Nek5000-users mailing list
> Nek5000-users at lists.mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/nek5000-users
>