spectre tran simulation and multithreading

Hi Marcel,

A few points about what you've said above:

1. Setting the number of cores in distributed mode doesn't affect the number of threads that APS uses (by default), so that may be why you are not seeing any difference in performance
2. Spectre is not multi-threaded - only APS (and XPS) are - you have to enable APS (or ++aps) to do this and probably specify the number of threads suitably.
3. This means in your OCEAN script you should use something like:
   option( ?categ 'turboOpts
       'numThreads "4"
       'mtOption "Manual"
       'uniMode "APS"
   )
   You could (if using distributed mode and using Sun Grid Engine or LSF) put "sge" or "lsf" or "farm" as the value of 'numThreads, but that will only work if using distributed mode
4. When using drmsCmd, you should be able to use run(?block t) to cause it to block so you don't have to wait via another mechanism.
5. It would be wise to move to a newer version than MMSIM15.1 - very little reason to use a 3-4 year old version of the simulator. Simulation technology continues to advance, and in particular there were some advances in SPECTRE16.1 which meant that there was better utilisation fo the underlying compute resource that should give you a speed improvement
6. Of course, this is assuming that your circuit is big enough to reasonably take advantage of multi-threading. Very small circuits won't (e.g. less than about 1K devices).

Regards,

Andrew.

Hi Andrew,

Thanks a lot for detailed info.

Regarding #1, #2 and #3:

In my case was no ++aps, probable because of that was no difference between the runs.

Regarding the #4 , run(?block t) when I've read the ocean script documentation I see:

A DRMS (Distributed Resource Management System)
command, such as a bsub command for LSF or a qsub
command for SGE (Sun Grid Engine) used to submit a job.
When this argument is used, all other arguments, except
?jobName will be ignored. Moreover, it will not be possible to
call the OCEAN function wait on the jobs submitted using
this argument.
To know more about the command option, refer to the
section Submitting a Job in the chapter Using the Distributed
Processing Option in the Analog Design Environment of the
Virtuoso Analog Distributed Processing OptionUser Guide.

So, my understating is that ?block will be accepted, but it has no effect.

Regarding #5mmsim version 15.1: That's the minimum version that I have to support and I can not decide it - it is a request. I just wanted to be sure that I'm not using features which were not implemented on 15.1

#6. My test case was big enough, it is in order of x10K devices. 

BR,

Marcel

 Hi Marcel,

Not sure from your initial comment on #1,#2,#3 whether you had APS and multithreading set up (and just not ++aps) - if you did and still were only getting less than 100% utilisation with a design with 10K devices, then you should contact customer support so we can investigate.

You're quite right about the ?block t not working with ?drmsCmd. Oversight on my part. You can workaround this by using a blocking ?drmsCmd instead though - the reason why ?block t doesn't work is that it does this by monitoring the job id and that isn't available when you're using ?drmsCmd - only with the natively supported integrations of SGE or LSF (or LBS).

Regards,

Andrew.

Hi Andrew,

My mistake, when I've wrote '++aps' I've meant '+aps' . In other words I had no APS activated.

Best Regards,

Marcel