This paper presents an algorithm which prevents a simulation user
from choosing a simulation length. This choice is always tricky
and often leads to CPU-time waste, not to mention user-time waste.
Too often, simulation users forget to compute confidence intervals:
they only guess a simulation length and ignore the confidence on the
simulation results. Those who do compute them generally try several
lengths (and thus run several simulations) so as to obtain small enough
confidence intervals.
The algorithm aims at optimizing this length choice by running only
one simulation and by stopping it nearly as soon as possible, i.e.
when some predefined relative confidence intervals on each of the
performance criteria are reached. For this purpose, the confidence
intervals are periodically computed, at run-time, with the batch
mean method. According to these intermediate results and to estimators
properties, a mobile simulation length is (also periodically) predicted.
The algorithm automatically determines batch size and batches number.
This process goes on until all confidence intervals are smaller than
the predefined thresholds.
This algorithm is implemented in MIMESIS, a computer architecture
performance evaluation tool.