help-gnucap
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Help-gnucap] computing time


From: Felix Salfelder
Subject: Re: [Help-gnucap] computing time
Date: Wed, 2 Jan 2013 12:37:41 +0100
User-agent: Mutt/1.5.20 (2009-06-14)

Hi Romain.

On Wed, Jan 02, 2013 at 09:50:47AM +0100, Romain GAILLETON wrote:
> 1.    Which parameters can we use for reduce this computing time?

reltol, abstol, vntol....? see manual.

> 2.    Is there an option which would allow Gnucap to recalculate only
> the nodes where the voltage have changed?

gnucap asks the devices, whether they want to be re-evaluated. so
basically nothing should be recalculated unless necessary (but checked
nevertheless). devices that don't change their matrix stamp should not
cause recalculation of the LU decomposition.

> 3.    Is there an option which could allow us to send to the output
> only the nodes which are in a range?
> Example: “print tran v(nodes) “  Which nodes are between 0 to 4000

node names are strings, the print command allows patterns. so if your
circuit has the nodes X1 .. X4000 and Y1 .. Y4000, you could simply
choose the first ones using print tran v(X*). maybe there are more
patterns, allowing something like v(0???) v(1???) to select the 1st 2000
without changing your names (haven't tried).

> 5.    Could it be possible to reduce the computing time under 500ms in
> your opinion?

in most cases, faster is possible. how much faster is difficult to tell.
if everyting is running smootly (always converging within 3
iterations... ), it will be difficult to speed up. i'd look out
for bottlenecks first. for example
- if switches are involved: i've had trouble with convergence due to the
  zero transition time in d_switch.cc. this _could_ be a reason for many
  unnecessary re-evaluations. look at what happens at/around the switch
  (during iteration), if it looks weird try implementing a nonzero
  transition time switch (i haven't had the time yet).
- for large circuits, the node order is crucial, as it impacts the
  bandwith of the system matrix. depending on the node order, the LU
  decomposition can take between linear and cubic time. what you can do
  is reorder the circuit. if it changes things, you should investigate
  further.

hth
felix



reply via email to

[Prev in Thread] Current Thread [Next in Thread]