Learning Forum

[Filip Parol]

I'm performing 3-d static analysis with imported temperature field and fracture tool (arbitrary crack), 3mln nodes and some contact in the analisys, using PCG solver with gpu accelatetion.

To solve one step it takes around 5 min and "writing results file" between steps takes 2-3h. i'm clueless why it happens, my cpus are running 100% and there is no read or write activity on my ssd drive.

Do you have any idea what's wrong?

[Chandra Sekaran]

With SSD drive it should be fast. Have you saved the project to the SSD drive before executing SOLVE? Otherwise the solution is done on a scratch directory elsewhere.

You can do a trial run by adding "outre,all,none" as a command snippet. This basically should prevent anything from being written to the rst file. See if this is still slow.

You can also add a "*GET,CPUTIME,active,,time,cpu" and "*GET,walltime,active,,time,wall" as a command snippet that is active for all load steps. Then when you look at the solve.out file (or Solution Information in Mechanical) it may give you an idea.

[Filip Parol]

switching to only 8 cores without gpu accelleration it take significantly longer to solve but the writing time is normal. althout i get this warrning:

*** WARNING ***                         CP =    5295.719   TIME= 14:05:32
 During this session the elapsed time exceeds the CPU time by 71%.      
 Often this indicates either a lack of physical memory (RAM) required   
 to efficiently handle this simulation or it indicates a particularly   
 slow hard drive configuration.  This simulation can be expected to run 
 faster on identical hardware if additional RAM or a faster hard drive  
 configuration is made available.  For more details, please see the     
 ANSYS Performance Guide which is part of the ANSYS Help system.

 

Total CPU time for main thread                    :     3178.6 seconds
Total CPU time summed for all threads             :     5295.7 seconds

Solver uses 25GB out of 64 avelible RAM, i think partialy long elapsed time is due to contact element and fracter parameter creation, can it be righte?

But still it dasn't explain why with gpu accelleration it take 3-4h to "write results file".

PS. i have Samsung evo 970 Pro SSD drive

[mrife]

Hi Filip Parol & Chandra Sekaran

Watch out as not all SSD drives are created equally.  Some of the "consumer" level SSDs perform very well when there is one process writing/reading to the drive.  But not when there are multiple processes trying to read/write concurrently.  And distributed parallel solving means that if run on N cpu cores there are N solver processes usually all trying to read/write at the same time.  You may want to search for any publised tests of the concurrent process read/write tests of the specific SSD on your system.  Or ask your IT departments hardware people if you happen to work at a company/university that has a IT department.  Or test it your self by solving a medium sized model with 1 cpu core, then 2, then 4 up to the limits of your system.  

Mike

[Filip Parol]

thx for your reply. i'm, performing 1-5 analysis every day usualy using 4-8 cores, it's first time i encounter such problem. i checked on other 3,5mln nodes model how it solves with 8cores and 1 gpu and it takes sekundes to write.

i'm using samsung evo 970 pro ssd drive.

as i have writen in previus reply, when solving without gpu accelleration there is no problem with writing solution file.

[Filip Parol]

.