next up previous contents
Next: 6.4 User Interaction and Up: 6. Parallel and Distributed Previous: Host Validation and

6.3 Performance Estimation

To illustrate the benefits arising from job-farming let us consider a rigorous calibration of a device simulator. For the estimation of the required simulation time let us assume the following: Transfer curves (ID/VG) with the overall number of N operating points are available, M parameters have to be calibrated, I optimizer iterations are necessary, W workstations are available for computation, and the typical computation time required per operating point is T.

Given that each optimization iteration consists of gradient computation and evaluation, the overall computation time is roughly (M+1) x I x T x N. Parallel evaluation of transfer curves reduces this time to $(M+1)\times I\times
T\times\frac{N}{W}$. For N=30, M=4, W=15, I=100, and $T=1\mbox{min}$, this means that job farming is able to reduce the time compared to operation on a single workstation from approximately 10 days to 16 hours.

Rudi Strasser