Memory Estimates in epsilon.real.x

Submitted by aohara on Thu, 02/26/2015 - 09:10


Bug reports

I just wanted to report what I believe to be a minor bug with no effect on calculated results, just leads to some inconveniences. It appears that the memory estimate report for epsilon.real.x is very far off. I am using the 1.1 beta2 version of BerkeleyGW. When I do a calculation on the q-> 0 point for an epsilon_cutoff of ~100 and number of bands ~10000, the estimate reports that it will require 799.1 MB per PE while teh available memory is 1824.9 MB per PE. The job, howeve, crashes right before doing the matrix element calculation with "insufficient virtual memory." In other cases, I've had a job crash later at the polarizability step, again having originally reported that enough memory was needed. I can eventually get the job to run with more processors, however it is hard to know a priori how many more to use (for example the current job runs fine but says the memory estiamte is 353.6 MB per PE).

It seems that for non q->0 points, the memory estimate is more reasonable (and probably as close to correct as can be). For example, at my first non q->0 point (with everything the same as above), the memory estimate is 1427.3 MB per PE and it is running without issue.

dvigil's picture

Submitted by dvigil on Mon, 07/13/2015 - 10:34

Hi Andy,
Thanks for the input. I will check why the memory report is incorrect for the q->0 point. This point is treated in a special way, so probably something was missed in the memory report. I will let you know when this is fixed.