How to make RF type simulation faster with MULTI-COREs & MULTI-THREADS

Please upload (use the Options tab when replying) the spectre.out from your simulation; that way I can see what options you've picked, the size of your circuit, and any messages spectre will have given.

Thanks,

Andrew.

Dear Andrew,

I have attched the spectre.out file. This simulation is currently running in APS mode and this is the corresponding spectre.out file.

However, I want the difference between Spectre, Turbo and APS, in performance and accuracy point of views.

Kind Regards,

OK, so it's not limited to 8 threads - you're running with 10 and it says it is running with 10. You have also set process affinity (although I advised you not to - you've actually set the processor affinity to wider than 10 CPUs - which is fine - but you need to be careful that there is not a second job running on the same machine with the same processor affinity; so bestnot to set it and let the OS take care of it).

Turbo is no longer in MMSIM (in MMSIM11.1 or MMSIM12.1 - as I said earlier, it was end of lifed.

There is discussion on APS in the documentation, and I suggest you read that. The idea with APS is that it doesn't compromise accuracy, but does provide accelleration by a combination of optimization techniques on what is being solved (e.g. filtering out parts of the circuit which will have no influence, such as dangling resistors, amongst other things), plus newer algorithms - e.g. a new matrix solver - plus smarter timestep placement - plus the ability to evaluate the device model code in parallel and also to do the matrix solve in parallel.

Regards,

Andrew.

Dear Andrew,

Thanks a lot.

When I sent the spectre.out file. I was running the simulation in APS. APS takes as many cores as one can give. It can take all 16 cores of my machine. 

However, if I run in Spectre option, It was taking maximum 8 cores. That is what I was aking: that Why spectre option is taking maximum 8 cores ?

I will look into the documents.

Thanks a lot.

Kind Regards,

Dear Andrew,

Sorry, to jot down another query( I feel it is relevant to add in this query:-

There is also option remote, local and distributed.

Is it possible to run remotely ( for example a server) through ssh ( I am using IC5141).

It looks like IC5141 can run remotely though rsh.

KInd Regards,

That's right - it doesn't really support ssh in IC5141. About the only way of doing it is to change rsh on the machine to be a link to ssh, which most people are a bit reluctant to change.

Andrew.

Andrew, I could not get you that change rsh on the machine to be a link to ssh

kind regards,

What some people do is:

mv /usr/bin/rsh /usr/bin/rsh_real
ln -s /usr/bin/ssh /usr/bin/rsh

(I've not actually tried doing this myself). You then have to of course ensure that you can ssh without needing a password, or confirming authentication etc. 

If you have IT who are prepared to disallow rsh, then maybe they would be OK to replace rsh with ssh.

Regards,

Andrew.

 Dear Andrew,

Thanks a lot.

Few  more queries. Could you please reply to them. It will be really helpful.

I think I should know the cause of below warning. Earlier I was skipping them. :-

 1:-

 I am getting following warnings ( though it is taken care by the tool): What is batch mode ? 

*WARNING* The Spectre run mode needs to be 'batch' when Spectre high-performance simulation mode is enabled. Automatically setting the run mode to 'batch'

2:-

+rcopt is meant for what ? where it can be set in ADE ? If I set it what basically the tool will do ? Will it affect the accuracy. By the way, it is an extracted Netlist.

Notice from spectre during hierarchy flattening.
    The circuit has many RC networks, using +rcopt in the command line may improve the performance and memory.

3:- Why the below warning comes ? Where I can set "preserve_inst" in ADE

 Warning from spectre during hierarchy flattening.
    WARNING (SPECTRE-8527): There are some nodes and instances removed from the netlist topology check. Check the circuit and correct the topology. Or use preserve_inst to keep them if necessary.

4:-  What is the cause of following warnings.Will it affect accuracy ?

Warning from spectre during hierarchy flattening.
    WARNING (SPECTRE-8528): Terminals are connected together. I48.rTIA_Q_AUX_IN_MINUS_89 is removed.
    WARNING (SPECTRE-8528): Terminals are connected together. I48.rTIA_I_MAIN_IN_MINUS_97 is removed.
    WARNING (SPECTRE-8528): Terminals are connected together. I48.rTIA_I_MAIN_IN_PLUS_131 is removed.
    WARNING (SPECTRE-8528): Terminals are connected together. I48.rRF_IN_AUX_137 is removed.
    WARNING (SPECTRE-8528): Terminals are connected together. I48.rRF_IN_AUX_134 is removed.

5:- What is Dangling Instance ?

WARNING (SPECTRE-8529): Dangling instance. I48.rLATCH_I_PLUS_1457 is removed.
    WARNING (SPECTRE-8529): Dangling instance. I31.rLATCH_I_PLUS_1438 is removed.

6:- Same Dangling stuff but differnt warnning Number ?

 Total removed nodes 2 ( dangling nodes 2, floating nodes 0 )

Warning from spectre during hierarchy flattening.
    WARNING (SPECTRE-8530): Dangling node. net048 is removed.
    WARNING (SPECTRE-8530): Dangling node. net055 is removed.

 Kind Regards,

1. By default spectre runs in ADE in "interactive" mode. This means that ADE starts the simulator, and then if subsequent runs do not change the design or analyses (just change parameters or nothing at all), then that change is communicated interactively to spectre and spectre will re-run without needing to re-start the executable, re-checkout the license, re-parse the netlist - it just changes the parameter value in memory. APS however does not support this model, so ADE automatically switches into "batch" mode, where the simulator starts fresh on every run.
2. This is an enhanced algorithm for large RC networks which is currently not on by default until we've had more feedback from customers - it is not reducing the parasitics, but solving them more efficiently and using less memory.
3. It's at the bottom of the High Performance options form. The warning is coming because some things were removed in order to optimize out unnecessary components (e.g. dangling instances etc below).
4. If a resistor (ideal, with no capacitance) has its terminals shorted, it gets pruned out because it won't affect the simulation (other than slow it down)
5. Simularly resistors which are only connected at one end (and floating at the other) get pruned out - they wouldn't affect the results because there's no current flow.
6. It's a different warning number because it's a different message. This is talking about the nodes being removed (also good for simulation performance)

Kind Regards,

Andrew.