46 4. Designprocess is used to find the parameter combination that minimizes the aggregatederror between the propagation simulation result and the collected measurements.4.2.2.1 InitializationThe measurements of the signal strength for the individual APs are either collectedby using the mobile devices that should be used as localization targets later orby some special hardware like the WISPY spectrum analyzer1. The measurementsare generated by reading the device specific Wi-Fi-APIs and then transmitted overHTTP to the server instance. There, they are stored for each device separately toretain the possibility to adapt for possible device specific distortions.Since the readings of all APs can be measured parallel at each location, the proce-dure is not time consuming. In the UMIC scenario measuring for 30 seconds at 60locations has been shown sufficient for running the parameter optimization.4.2.2.2 OptimizationThe parameter optimization is driven by a genetic algorithm on the Optimizer com-ponent. A set of free parameters constitute an organism in the context of the geneticalgorithm. The fitness of an organism is evaluated by running raytracer simulationsfor all APs over the geometry of the scene with the corresponding parameters ofthat organism. This can be a time consuming task since a representative raytracersimulation needs more than 30 seconds on a current GPU/CPU and convergence ofthe cost function is reached by evaluating at least 3000 organisms. In the analysedUMIC scenario with 22 APs, this leads to a serial processing duration of around22 days. Through exploiting the parallelizability in the cost/fitness function andthe parallel nature of the breeding phase, the duration of such an optimization runcan be reduced nearly linearly, as long as Ngpus<< Npop× Naps, by employingGPU-nodes. These nodes are accessed through the Simulator component that willdistribute them accordingly.The nodes are designed as lightweight processes that are hard wired to run thePHOTON raytracer and execute some job specific code on the results before pushingthem back to the Simulator component. The job specific code is transferred withthe job package to the node. This allows to bring new types of jobs into the systemby only updating the Server deployment. The security risk imposed by transferringthe executed result inspection code over HTTP, should be minor as the nodes pullthe jobs from a single preconfigured host. The design of the system can easilybe extended to allow automatic spawning of nodes but no real efforts have beenundertaken in that direction.The motivation for choosing such an approach was given by the circumstances thatduring the duration of this thesis the RWTH had deployed a new GPU-Clusterconsisting of around 50 NVIDIA QUADRO 6000 nodes. By using these resourcesa full optimizing run containing 30000 simulations with a duration of 30 sec each,1Although measurements with the WISPY were taken with the goal to derive the free raytracerparameters, the evaluation of the results were inconclusive and have thus been removed from thescope of this thesis.