Relative error between [cloudsim simulation] and [riot]

[ginfung]

error = cloudsim simulation(i.e. true value) - riot (i.e. riot surrogate based sampling prediction)

for each model: x-axis: relative error for objective 1 (makespan) y-axis: relative error for objective 2 (cost) each scatter plot point := relative error for one configuration

[timm]

Waaat? did u get those workflows running on real hardware?

t

[ginfung]

@timm short answer: no. I was still using a simulator.

• recall that RIOT, or earlier paper, the SWAY, is to reduce the number of evaluation.
• In our scheduling problem (again, not workload/performance prediction), evaluation = using simulator, e.g. CloudSim, to assess configurations.
• In our previous submission, we have shown that even with CloudSim ( a very mature toolkit), such simulation is still very time-consuming.

• RIOT can significantly reduce the number of evaluations. But we don't know whether our surrogate-based sampling objective prediction is precision.
• this chart compares [our objective prediction] vs. [cloudsim's evaluation]