Re: [mesa-users] minimum mass fraction values

SourceForge Headquarters 225 Broadway Suite 1600 San Diego, CA 92101 +1 (858) 422-6466

"FWIW, my experience with MESA as well as in other general computing is that increasing memory gives more speedup than increasing cores.   There's probably documentation for this effect, maybe someone more CS trained than I has a reference."
For more cores to be helpful, there have to be multiple pieces of work to do that don't depend on the results of work not already done (CS shorthand for this is "the task must be parallelizable"). An analogy often given is "one woman can produce a baby in 9 months,  and 9 women can produce 9 babies in 9 months,  but 9 women can't produce one baby in a month".
RAM doesn't provide any speedup by itself, but if the information a program is working on doesn't fit in memory,  the operating system will start writing stuff in memory to disk in order to pretend that it has more RAM. Disk access is slow compared to memory access, so this tends to slow things down significantly: If your program needs only a little more RAM than is available, and most other processes in the machine aren't doing anything at the moment,  things will be done almost as quickly as if everything fit in RAM,  but if you need a lot more memory than you have, and/or other programs are active and fighting yours for memory,  performance will fall sharply. Analogously, on an open book quiz,  students that have everything memorized will be done faster than those who have to check their notes,  and someone who can do multi-digit arithmetic in their head will be done with a math problem faster than someone who has to write it out on paper. CPU cache size can impact performance in a similar way to RAM size,  fir similar reasons. 

-------- Original message --------
From: amber lauer <amb...@gm...> 
Date: 3/9/2017  23:07  (GMT-06:00) 
To: Robert Farmer <rjf...@as...> 
Cc: mes...@li... 
Subject: Re: [mesa-users] minimum mass fraction values 

Whenever I run with the adaptive net, the nets in the "nets" folder usually have 70-100 species in my experience. I don't know if that's typical. I am still doing XRB, as well as other models. Generally I run with Max Z, N, and A turned off and use Min_X values ~10e-11. The 10e-21 run was for benchmarking and to see if there was any significant difference, I had the elapsed time column added but I didn't check it the total run time, but I can if you are interested. It was really just a one off. The machine I'm speaking of has 48Gb of memory and it can run that 
adaptive net case with mesa_67.net as the initial net in a day. FWIW, my experience with MESA as well as in other general computing is that increasing memory gives more speedup than increasing cores.   There's probably documentation for this effect, maybe someone more CS trained than I has a reference. 

On Thu, Mar 9, 2017 at 2:30 PM, Robert Farmer <rjf...@as...> wrote:
>So 17Gb is the max mesa can use for the nuc nets matrix solver? 

this is the combination of nuclear network plus structure equations (they are solved simultaneously) matrix,  you may use more memory overall as this doesn't include things like the eos files that are loaded etc

Are you still doing xray bursts? what did you have for the other adaptive net options?

Rob

On Thu, Mar 9, 2017 at 1:21 PM, amber lauer <amb...@gm...> wrote:
So 17Gb is the max mesa can use for the nuc nets matrix solver? I just finished a run with an adaptive net with my min Z set at 10e-21, and I used a starting model with an abundance profile containing roughly 20 species. I just checked the profile in the region of interest for my research, and even with those settings the net only has about 70 species. So it works out OK, at least in my case. 

On Wed, Mar 8, 2017 at 11:53 AM, Robert Farmer <rjf...@as...> wrote:
>Unfortunately in my case, this can be as low as 10e-9. For more final 
models, I have a machine built specially for this purpose with 16 cores,
 so I can dedicate a bit more time for the larger networks. 

There are limits on how big the matrix mesa can solve based on memory limits. Mesa memory scales as ~num_isos**2 * num_zones, with an upper limit based of the size of the 4-byte int used to index the matrix array (2^31-1 which is about ~17GB of memory for double precision numbers). This is about 400 isos with 1000 zones, so if you want more isotopes you will need to reduce your spatial resolution.

>What about when starting from an abundance profile? If I have an isotope
 with abundance of 10e-13, but min x for keep is set at 10e-9, for 
example, will it eliminate it immediately? Or after a few steps?

Immediately, though it might get re-added if there is a reaction that makes it. This is the problem i find with the adaptive nets is once you add an isotope it gets quite hard to remove it, given all the possible reactions that could make it, so the nets never get smaller.

Rob

On Wed, Mar 8, 2017 at 10:18 AM, amber lauer <amb...@gm...> wrote:
Yeah, we've all experienced that vicious slowdown. However when trying to match observables, you have to at least use the values from observation as a lower bound. Unfortunately in my case, this can be as low as 10e-9. For more final models, I have a machine built specially for this purpose with 16 cores, so I can dedicate a bit more time for the larger networks. 

What about when starting from an abundance profile? If I have an isotope with abundance of 10e-13, but min x for keep is set at 10e-9, for example, will it eliminate it immediately? Or after a few steps?

On Wed, Mar 8, 2017 at 10:33 AM, Robert Farmer <rjf...@as...> wrote:
Hi

Those values come from the cited Woosley, Heger, et al, ApJSS, 151:75-102, 2004 paper. I would advise against using them for mesa, in fact i'm going to go change them. If you look in the MESA3 paper at the x-ray burst section, I use much larger values for the different controls, 10**-6 to 10**-3, the smaller the number for xmin the more isotopes you get in the net and you can very quickly get more than mesa can cope with.

Rob

On Wed, Mar 8, 2017 at 9:00 AM, amber lauer <amb...@gm...> wrote:
I'm trying to decide on a good min X, not just for the adaptive net, but also for inclusion in abundance profiles. I see that the 'standard' values in star_job.defaults are 10e-23 and ~10e-18. Is there a numerical/computational argument for a relevant threshold? 

-- 
Amber Lauer. M.S. Physics
PhD Candidate,
Graduate School of Physics,Louisiana State University

------------------------------------------------------------------------------

Announcing the Oxford Dictionaries API! The API offers world-renowned

dictionary content that is easy and intuitive to access. Sign up for an

account today to start using our lexical data to power your apps and

projects. Get started today and enter our developer competition.

http://sdm.link/oxford
_______________________________________________

mesa-users mailing list

mes...@li...

https://lists.sourceforge.net/lists/listinfo/mesa-users

------------------------------------------------------------------------------

Announcing the Oxford Dictionaries API! The API offers world-renowned

dictionary content that is easy and intuitive to access. Sign up for an

account today to start using our lexical data to power your apps and

projects. Get started today and enter our developer competition.

http://sdm.link/oxford
_______________________________________________

mesa-users mailing list

mes...@li...

https://lists.sourceforge.net/lists/listinfo/mesa-users

-- 
Amber Lauer. M.S. Physics
PhD Candidate,
Graduate School of Physics,Louisiana State University

------------------------------------------------------------------------------

Announcing the Oxford Dictionaries API! The API offers world-renowned

dictionary content that is easy and intuitive to access. Sign up for an

account today to start using our lexical data to power your apps and

projects. Get started today and enter our developer competition.

http://sdm.link/oxford
_______________________________________________

mesa-users mailing list

mes...@li...

https://lists.sourceforge.net/lists/listinfo/mesa-users

-- 
Amber Lauer. M.S. Physics
PhD Candidate,
Graduate School of Physics,Louisiana State University

-- 
Amber Lauer. M.S. Physics
PhD Candidate,
Graduate School of Physics,Louisiana State University