Spectre parallel jobs reduce near finish

Hi,

you might try this one:

Article (20338418) Title: How to redistribute Monte Carlo jobs in ADE XL/Assembler to better utilize faster machines?
URL: https://support.cadence.com/apex/ArticleAttachmentPortal?id=a1Od0000000uxPBEAY

MC runs are "grouped" together, so you probably have 100 runs per host. Now one of the hosts might be slow or busy which will 
hold of 100 of your runs.  Above solution discuss how to redistribute such remaining runs.

The variable to control that is:

In .cdsenv:

adexl.monte minGroupSizeSplitAcrossIdleJobs int 2

In .cdsinit:

envSetVal("adexl.monte" "minGroupSizeSplitAcrossIdleJobs" 'int 2)

Kind regards,

Marc 

Thanks Marc but the issue is with PVT corner runs not monte carlo. The spectre jobs seem to automatically decrease as the last corners are approached. 

Hi,

we also do grouping for param/corner  runs. Controled by this:

envSetVal("maestro.simulation" "groupRunE" 'boolean t)       ;  E - Explorer
envSetVal("maestro.simulation" "groupRunA" 'boolean t)       ; A- Assembler

and there is also a setting to redistribute incase one host is slower than the others:

envSetVal("maestro.simulation" "minGroupSizeSplitAcrossIdleJobs" 'int 2)

Could you determine if it is one host struggling  like I assumed? Or are there other problems?  Maybe you can check the 
Job logfiles to see what' s going on. Lots of guessing at the moment from my side. 

Kind regards,

Marc

Hi,

we also do grouping for param/corner  runs. Controled by this:

envSetVal("maestro.simulation" "groupRunE" 'boolean t)       ;  E - Explorer
envSetVal("maestro.simulation" "groupRunA" 'boolean t)       ; A- Assembler

and there is also a setting to redistribute incase one host is slower than the others:

envSetVal("maestro.simulation" "minGroupSizeSplitAcrossIdleJobs" 'int 2)

Could you determine if it is one host struggling  like I assumed? Or are there other problems?  Maybe you can check the 
Job logfiles to see what' s going on. Lots of guessing at the moment from my side. 

Kind regards,

Marc

Hi Marc, this seems to be background redistribution to reduce overall simulation time across all runs. If Max Jobs is set to 10, no matter what the clever redistribution is happening in the background in the Explorer Run Summary wouldnt you expect the jobs to remain as 10 for the whole corner set and not reduce as the last corners are approaching?

Hi,

at the start of the simulations we group points together and assign them to a job.  That job is responsible for that group of points and will  work through them.   Without the "minGroupSizeSplitAcrossIdleJobs"  setting it will stay like that initial grouping till the very end. So there are situations where you see one job remaining, struggling with one group while all the other jobs did finish their groups already.

The reason for that "struggling" could be  slow/busy hosts or simulations point with difficult convergence conditions, whatever.

The redistribution system was introduced some time after the first releases of Explorer/Assembler  for exactly the reasons you bring here. Make use of the idling jobs to help finishing the rest of the points.

Kind regards,

Marc