Stats

  • Replies 2
  • Subscribers 133
  • Views 87
  • Members are here 0

Job mode control issue, LSCS v.s. ICRP

gangliurfic

I am working in a large company where there is grid to run the jobs.

I have a small simulation which takes less than 1 min simulation time, but I want to do a variable sweep, say 32 points. so I have total 32 simulations.

If I use LCSC mode, it will first launch one job, wait for the server, then get it done, and then submit 32 netlisting jobs followed by 32 simulation jobs. After I get the results, I change the variable sweep and launch the sim again. It will go through the procedure again. Overall the turn around time is much longer, especially when the grid is busy

I prefer to use ICRP mode for these small sims. Same experiment, it will submit 32 jobs once and that's it. When each job is done the results are there. If I launch another sim immediately, the grid server will be available and the new sim will start immediately. The turn around time is much shorter than LCSC mode.

However, my ICRP jobs gets killed by itself.

ERROR (ASSEMBLER-5034): While simulating run Interactive.24, point 27, xxx, received error:
Error:
------------------------------
ERROR - ICRP has stopped the Spectre process because the process consistency check has failed unexpectedly. Possible reasons could be the unavailability of the required license or an unusual delay in launching Spectre.

for the point: (22 27)

Actually the simulation itself started, but it seems something was triggerd at a certain time, maybe around 60s and then killed the sim before it finishes. 

I don't know if this is because of something set on the grid, or it is because of something set by cadence. BTW, the same simulation can finish without any problem in LCSC mode.

I was using ICRP in my old company and things worked well. Now I started in a new company and my ICRP jobs can't run. It won't be resolved through the company IT/CAD as I was told ICRP mode was not recommended.

Therefore I want to bring up the discussion here. If there are some settings I can change on the user side to disable this kind of check, then it'll be good.

Thank you.

  • Andrew Beckett

    You didn't say which IC (and Spectre) version you are using, but I assume it's not the latest. In IC25.1 versions, if a variable is being swept this normally doesn't result in a specific "netlisting job" with LSCS and instead that is done as a step at the beginning of each simulation (it's generally a small text change and doesn't really need the netlister to do any work).

    Secondly, with LSCS you may benefit from changing this cdsenv var:

    envSetVal("maestro.simulation" "exitJobsAfterRunsComplete" 'boolean  nil)

    (the default is t). With LSCS, the default behaviour is for the simulation job to exit once all the runs have completed (a lot of our customers didn't like slots being retained on the grid for the linger time fo the simulation job). Setting this back to nil will keep it running for the linger time, and hence that will mean that the second run will be much faster.

    In general, ICRP is being de-emphasized and the idea is that it will be removed in the future (IC25.1 warns you of this). We're not quite ready to do that yet, but the whole idea of LSCS was to have a more scalable solution; that said, there have been some issues with smaller jobs such as you're seeing - and a number of those have been addressed in IC25.1 (and there are some further netlisting process optimisations coming in IC25.1 ISR6, due at the end of this week).

    The failure that you are seeing with ICRP seems to be related to a validation step with the simulator to ensure it is actually spectre. Are you running with Spectre or a different simulator? I'm not sure why it would fail with ICRP and not with LSCS because the same validation should happen in both cases, but maybe there was a bug in a specific sub-version of whichever IC release you're using.

    Regards,

    Andrew

  • gangliurfic
    in reply to Andrew Beckett

    Thank you Andrew for the quick feedback. A few more comments:

    1. I am still having issues creating new post or editing post. I hope this can be fixed soon.

    2. Accidentally, I noticed that if I enable a Config sweep in Global Viables section, even if I don’t sweep, just use a single view, my ICRP run can finish without the previous error. If I uncheck the Config sweep, then error appears.  Which part is different between these two cases?

    3. With #2, I can run the same sim in both ICRP and LCSC mode to compare

    I am running IC25.1_ISR4,  spectre  version 25.1.0 64bit 03/20/2026 04:15 (mmsimcm48c-2).

    I started virtuoso on a 32core machine, and  I am running my simulation jobs locally, to bypass the pending time on the queue.

    For a very small simulation, ICRP takes about 1:40 min to see the results for 1st run, and about 50s to see the results for following runs. LCSC takes about 2:20 to see the results for 1st run and 1:20 min for following runs. The difference between the 1st run and following runs is because of setting up and loading lots of company stuff for the 1st run. For following runs, this step is much quicker

    For a small simulation but with 30 variable sweep, ICRP takes about 5mins to see all results for 1st run, and about 2mins for following runs. LCSC takes about 6mins to see all results for 1st run and about 2:30 mins for following runs.

    Overall ICRP is slightly faster on local machine. However, I expect the difference becomes bigger when submitting jobs to grid with pending time for all jobs, because LCSC mode has more steps. Suppose each job waits the same amount of time, then LCSC mode wait time is at least doubled. This is my guess and I don’t have data to back up.

    Question on LCSC mode:

    1. is there any other settings I can try to reduce the turn around time?

    2. LCSC has different Netlist and simulation job policies. a 30 point variable sweep will have 30 netlisting job, do they share the same cpu, or each Netlist process has its own cpu? I have a 32 core machine, if I do 1cpu per job, and I set both Netlist and simulation to max. 30, then I get short simulation time, compared to max 30 for simulation and 8 for Netlist.

    Thanks!

Community Guidelines

The Cadence Design Communities support Cadence users and technologists interacting to exchange ideas, news, technical information, and best practices to solve problems and get the most from Cadence technology. The community is open to everyone, and to provide the most value, we require participants to follow our Community Guidelines that facilitate a quality exchange of ideas and information. By accessing, contributing, using or downloading any materials from the site, you agree to be bound by the full Community Guidelines.