next up previous contents
Next: About this document ... Up: Technical Notes Previous: Data Structures   Contents

Subsections

Performance

This section gives some notes on some of the performance issues in simulators. It is not intended to be complete or well organized.

Virtual functions

There is a question of the impact on speed from the use of virtual functions. The experiment used here is to use the circuit eq4-2305.ckt from the examples directory, and try several modified versions of the program. I used a 100 point dc sweep, a version between 0.20 and 0.21, and made several modifications for testing purposes. I chose this circuit because it has little to mask the effect, and therefore is sort of a worst case.

I added an int foo to the element class. I made the function il_trload_source call a virtual function virtual_test and stored the result. The local version body has a print call, which should not show, to make sure it calls the other. These functions simply return a constant, determined by which version of the function is called. Run time is compared, with and without this.

With 1 virtual function call (included in load)

               user      sys    total
  evaluate    13.45     0.11    13.56
      load    13.40     0.06    13.47
        lu     1.91     0.09     2.00
      back    22.35     0.27    22.61
    review     0.00     0.00     0.00
    output     0.11     0.11     0.22
  overhead     0.23     0.19     0.42
     total    51.45     0.83    52.28

With 10 virtual function calls (included in load)

               user      sys    total
  evaluate    13.47     0.09    13.57
      load    24.69     0.17    24.87
        lu     2.09     0.02     2.11
      back    22.17     0.35    22.51
    review     0.00     0.00     0.00
    output     0.14     0.11     0.25
  overhead     0.25     0.25     0.50
     total    62.82     0.99    63.81

No extra function calls (included in load)

               user      sys    total
  evaluate    13.41     0.09    13.50
      load    11.75     0.05    11.79
        lu     2.04     0.03     2.07
      back    22.51     0.33    22.84
    review     0.00     0.00     0.00
    output     0.08     0.11     0.19
  overhead     0.31     0.25     0.56
     total    50.10     0.86    50.96

My conclusion is that in this context, even a single virtual function call is significant (10-15% of the load time), but not so significant as to prohibit their use. The load loop here calls one virtual function inside a loop. The virtual function calls an ordinary member function. Therefore, about 30% of the load time is function call overhead.

The impact should be less significant for complex models like transistors because the calculation time is much higher and would serve to hide this better.

Spice uses a different architecture, where a single function evaluates and loads all elements of a given type. This avoids these two calls.

Inline functions

For this test, il_trload_source is not inline. Contrast to "No extra function calls" and "1 virtual function" above, in which this function is inline.

               user      sys    total
  evaluate    13.44     0.15    13.60
      load    13.85     0.14    13.99
        lu     1.73     0.02     1.75
      back    22.89     0.35    23.24
    review     0.00     0.00     0.00
  overhead     0.45     0.17     0.63
     total    52.50     0.94    53.44

This shows (crudely) that the overhead of an ordinary private member function call (called from another member function in the same class) is significant here. The cost of a virtual function call is comparable to the cost of an ordinary private member function call.


next up previous contents
Next: About this document ... Up: Technical Notes Previous: Data Structures   Contents
Al Davis 2002-03-26