[Ngspice-devel] ngspice speed comparisons and hints for optimization

[Holger V. <hol...@un...>]

Dear all,

below you will find some speed comparisons of ngspice, using a Pentium4 
with 3.4GHz and 1.5GByte of memory.

The input file is:
/tests/tcl-testbench4 example.cir
with the settings:
+VERSION = 3.3.0 
and:
.TRAN 0.02n 50n 0n 0.5n
No other changes werde made to the file.

WindowsXP:

ngspice_m (-O3)                      162.3 sec
ngspice_vc (/O2 /Oi /Ot /GL)     94.8 sec
ngspice_vc debug (/Od)            152.6 sec

168364 iterations were needed.

_vc = MS Visual Studio 2008
_m  = MINGW gcc 3.4.5

For comparison plain old spice (DuSpice.exe with MS Visual Studio 2008, 
/O2 /Oi /Ot /GL) was used:
DuSpice (vc)                               52.0 sec

The speed advantage, however, is not due to faster algorithms, but to a 
reduced number of only 87254 iterations. This might be a hint how to 
increase simulation speed.

Linux Suse 10.2 gcc 4.1.2 gave:
ngspice (-O3)                            135,9 sec.

Then I tried to figure out the use of code symbols, applying  
CodeAnalyst from AMD (a free profiling tool for Windows) and using 
ngspice_vc debug  with model BSIM 3.3.0:

BSIM3Load: 54,7%
spFactor:  14,8%
spSolve:    6,9%
CKTterr:    5,4%
spClear:    5,2%

All others are below 2.4% each.

Matrix solving is only about 25-30%. Most of the time the simulation 
spends in the BSIM3 model code (55% in BSIM3Load). Thus it is probably 
not too useful to spend much time for establishing a new matrix solver. 
The paper cited recently by Lionel (T.-H. Weng e.a.: OpenMP 
implementation of SPICE3 circuit simulator, IWOMP 2005/2006, LNCS 4315, 
pp. 361-371, 2008),  points to the same direction. The authors care for 
parallelising the MOS3load() function. As soon as I have a multi core 
processor machine available, I will try to figure out how that might 
benefit ngspice.

Regards

Holger