A spare part supply system for repairable spares in a repair shop is modeled as a set of heterogeneous parallel servers that have the ability to repair only certain types of repairables. The proposed model minimizes the total cost of holding inventory for spare parts, cost for backorder arising from downtime of the system due to the lack of spare parts and the cost of crosstraining for servers. Simulation-based Genetic Algorithm (GA) is proposed to optimize inventory levels and to determine the best skill assignments to servers, i.e., cross-training schemes. When methodology's performance is compared with total enumeration, tight optimality gaps are obtained.