Page 111

3.2

Scalability of Zereal

Number of Simulated Players

5000

10000

15000

20000

25000

T
i
m

e

(
s
)

500

1000

1500

2000

2500

3000

5000

10000

15000

20000

25000

Fig. 4. Zereal scalability (with regression line and prediction interval)

In figure 1 the wallclock runtime performance for 5 minutes simulated time

in the game (i.e. 300 simulated seconds) as a function of the number of simu-
lated players. The simulation is performed on 5 Athlon 1.6 GHz CPUs on an
Linux-cluster. The largest simulation performed so far with Zereal is with 160
thousand agents (simulated players and NPCs) on 20 CPUs. The setup for the
experiment was with simulated players and NPCs with Markov chain type rea-
soning (the largest number of hierarchical planning agents tested so far is 50000).
By doubling the number of players and keeping the number of simulated cycles
fixed the number of sense-reason-act cycles to be performed doubles, this can
explain the close-to-linear scalability.

Conclusion and Recommendations

It has been determined that there is significant interaction between factors in the
experiments (MarkovKillers*Radius, PlanAgents*Radius and Monsters*Radius)

The main factors and including all interactions between them explain ap-

proximately 97% of the response variable, this is a surprisingly high value since
the simulator has highly stochastic behavior of killers and monsters. However,
since each run is over a period of 100 cycles, and for each cycle the killers (and
monsters) perform one action (or inaction), one gets an "averaging" effect on
the measured wallclock time.

Paper D

<< - < - > - >>