proposal: distributed calculation to solve models

[Emo]

It's years now that I participate in calculation projects distributed by seti a wcg. It is now possible to achieve previously unthinkable goals.

I noticed that for work, long time I dedicate it to making the model, 1 week and the computer is there that does nothing and then I would immediately like the results. instead usually take about 2-3 days.
I therefore note that during the creation of the model waste many cycles of my computer.

I think to solve all this could come into play the distributed calculation. I elaborate in the dead moments the models of others (a small part of course) and then when I need all the power out that of others. you could make a credit system so you have to first process a tot for others and then you can ask for help to others.

Perhaps in a few hours I can get the results of my models! :smile:

What do you think?

 

[stefano.garbin]

not feasible.
a fem model is all about seti@home and beautiful company.
study the theory of finite elements and you will understand why.

 

[Fulvio Romano]

not feasible.
a fem model is all about seti@home and beautiful company.
study the theory of finite elements and you will understand why.

in the sense that the reverse of a scattered matrix is not segmentable?

 

[stefano.garbin]

Yes, it is, the reversal of scattered matrices can be parallelized.
but the question is perhaps not just this.
you have a domain divided into n partitions: these partitions during the resolute process must "talk". with what time latency can this communication become if I don't even know when one or more of these partitions will begin to be calculated? apart from of course the amount of data that must be exchanged at each cycle.. .

 

[Emo]

I had actually not thought about throughput for each processing. cmq would avoid writing "study the theory of finite elements a little", you can't know how much I know, just for an idea I wrote.

cmq I would not say "not feasible" but hardly feasible for the required band.

the idea was born from the fact a company parallel to the one for which work made calculations on 128 dual core machines with dyna and abaqus. obviously paid. of course they were all machines connected locally.

cmq an idea is that of hinges of a solution. for example on dyna contacts test + cinematographically feasible solutions. after you have tried so many choose the balanced and congruent one. it would be right at this stage to use the distributed calculation. throughput in this case would be the only si/no of the solution.

 

[stefano.garbin]

It is not just a matter of band, it is that when your model "turn" must "turn" in cotemporaneous all the partitions in which it was divided.
in any case the explicit cfd and structural models "scale" well. the classic implied fem instead has (at the moment) greater difficulty but I believe that the solution algorithms are also doing them considerable advances towards parallelization.
the idea that you've exposed works great if you're doing optimizations (mono or multi lens), where different copies of a parameterized model turn individually.
"last but not least" I didn't want to be offensive with my statement but only "punzecchiare" a little, and I'm glad that my "puntura" had a good effect! :wink: