How many simulation replications do I need?
Choose replication count from the precision you need instead of using a fixed rule of thumb.
If arrivals or service times are random, one run can be unusually quiet or unusually busy. Repeating the same setup with different random streams shows how much the result moves from run to run.
Those repeated runs are replications. For example, if the output is mean waiting time, each full run gives one mean waiting-time value for the comparison.
Start by deciding how much error you can tolerate. Run an initial batch, calculate the confidence-interval half-width, and add replications if the interval is still too wide.
A noisy measure may need many more replications than a stable one. A small decision margin also demands more precision than an obvious difference.
target: confidence-interval half-width ≤ the error you can tolerate
Each replication should normally use a different random stream. Reusing the same seed would simply repeat the same realization and provide no new information about run-to-run variation.
When comparing two alternatives, you can give them matched random streams. That technique is called common random numbers.
Increase the number of replications below. The individual results do not stop varying, but the estimate of their mean usually becomes more precise.