[FLASH-USERS] Spherical Coordinates Speed up

[Sean Couch]

Hi Nilou,

I’m not 100% sure, but I think there are a few reasons why you are seeing this behavior.  First, I’m not even sure that collectiveIO works with uniform grid mode.  The collective aspect means that blocks are collected on a subset of ranks before being written to disk.  In UG, there is only one block per rank by construction…so I’m just not sure it works together.  

Second, collectiveIO is to overcome issues in parallel writes of contention amongst many ranks attempting to write to a single file.  In my experience, this only matters when the number of ranks is in the 1000s or more.  I don’t think your problem size is anywhere near large enough to see a benefit from collective IO.

Sean


-----------------------------------------------------------
Sean M. Couch
Theoretical Astrophysics Including Relativity
California Institute of Technology
Mail Code 350-17
Pasadena, CA  91125
(626) 395-4282
www.tapir.caltech.edu/~smc <http://www.tapir.caltech.edu/~smc>
> On Mar 8, 2015, at 6:10 PM, niloofar afsari <nlr_afsari at yahoo.com> wrote:
> 
> Hi Flash users,
> 
> I'm running a 2d Sedov explosion (I've changed it to be compatible with spherical coordinates) simulation in a spherical coordinates on a uniform grid:
>  
> ./setup Sedov -auto +ug +parallelIO -2d -nxb=1024 -nyb=1024
> 
> In flash.par, I've set the "useCollectiveHDF5" to true. 
> 
> So far, I've changed nxb=nyb in the range [128, 2048] and iProcs=jProcs in the range [1,3].
> 
> I've never seen any speed-up when running my simulation with -np higher than 1. I'm wondering if this is expected or I'm missing an important speed-up flash/option.
> 
> Thank you for your help!
> 
> Nilou
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://flash.rochester.edu/pipermail/flash-users/attachments/20150310/b2cca346/attachment.htm>