[Hadstruct] Minutes

[Robert Edwards]

> On Apr 22, 2020, at 2:13 PM, Robert Edwards <edwards at jlab.org> wrote:
> 
>> 
>> Genprops
>> 
>> RE had been working with CE to check on memory footprint for genprops. That would be a concern, notably for GPDs rather than PDFs, and it was probably important to implement (or adapt) FW's hybrid approach. RE noted there was now a matrix version of keys.
>> 
> 
>    the memory issue is independent of whether it is a GPD or PDF. One has to hold the solution vectors. The issue of contractions is just how many times you want to reuse those solution vectors. Frank’s hybrid method only holds the solution vectors for the time slices in the 3pt, not the entire lattice. So, for T=14 and Lt=96, that’s a 4x reduction just for memory. As a reminder, you have a nominal (48^3x96)/(32^3x64) X (N=128/N=64) -> 11x increase in memory, so the 4x helps.
> 

Oops - I have to correct my math. If you switch to the hybrid mode, the memory reduction is  (Lt=96) / (T=14) -> 6.8x. Compare that to the memory increase from 32^3x64 -> 48^3x96 and N=128, so a 11x increase, an offsetting decrease of 6.8x sounds very welcomed. Namely, a real calculation on 48^3x14 with N=128 might even be possible…

		Robert