[MITgcm-devel] EVP slowdown

[Martin Losch]

Has there been any progress on the underflows in EVP and the associated slow-down?

Martin

On Nov 5, 2009, at 4:57 PM, Martin Losch wrote:

> I think there is some value. As I said in my earlier email about this, we can prevent the terrible slowdown from happening if we change the corresponding code. There must be a reason why the underflows only happen in evp (see your email about "paper1" and Elizabeth's way to correct underflows if a compiler can't do it by itself). If knew how to do this, I would have done it long ago.
> 
> The result would be either code that does not produce underflows, or additional code (as in CICE), that can be enabled via a CPP-flag, if a particular compiler can't handle the underflows properly.
> 
> What do you think?
> 
> Martin
> 
> On Nov 5, 2009, at 4:49 PM, Chris Hill wrote:
> 
>> My fault (as usual). I didn't really understand the explanation!
>> Is there some value in pursuing this? If so I can get Constantinos to
>> look at it - but
>> I don't want to spend time on it unless we think there may be some value.
>> 
>> Chris
>> 
>> On Thu, Nov 5, 2009 at 10:32 AM, Constantinos Evangelinos
>> <ce107 at ocean.mit.edu> wrote:
>>> On Friday 16 October 2009 5:26:05 pm Dimitris Menemenlis wrote:
>>> 
>>>> Constantine, Martin and I are curious whether there is an easy way to
>>>> track down the particular line (or lines) of code, which caused the
>>>> EVP slowdown?  Is there a compiler flag that can track when underflows
>>>> occurs, stop the model, and point to the line of code that is the
>>>> culprit?  D.
>>> 
>>> Oops - sorry I just noticed this e-mail.
>>> 
>>> I've shown Chris how to go about this using the -gsl flag to genmake2, link
>>> with the GSL library and use the proper environment variables for GSL with a
>>> code built with -g.
>>> 
>>> The alternative is to use openspeedshop's exception profiling to see where a
>>> very large number of underflows appears.
>>> 
>>> Let me know which one you want to pursue so I can guide you in more detail. We
>>> could try and do this on Stommel where we have more control over the machine.
>>> 
>>> Constantinos
>>> --
>>> Dr. Constantinos Evangelinos
>>> Department of Earth, Atmospheric and Planetary Sciences
>>> Massachusetts Institute of Technology
>>> 
>>> 
>