سميه الطائي

One statistically significant challenge faced by multi-server systems is efficient load balancing. This is because servers can differ in performance, workloads can be location an‎d time dependent, an‎d workload is further complicated by the need to conserve energy. Load balancing can be done using several techniques, such as Round Robin an‎d Least Loaded, which are simple an‎d predictable, but are unable to adapt to the workload dynamically. This research studies a load balancing system using Deep Reinforcement Learning (DRL) an‎d a Deep Q Network (DQN) to develop a policy which balances adaptive load system while optimizing the utilization of the servers an‎d decreasing the queue per server in the system. A number of virtual environments were built to simulate real-life server conditions to include variable request arrivals an‎d the processes of failover systems. The DQN agent was trained to allocate servers in a sequential manner depending on the load states of the servers when simulation time was stopped to form a decision. This system was tested in simulation environments that were DQN, Round Robin, an‎d Least Loaded. The DQN mechanism outperformed Round Robin, an‎d Least Loaded, by a perceivable margin of average reward per step of 0.85, 0.62, an‎d 0.74, respectively. DQN, in addition to manger improved load balancing, also managed to fail less servers over the sustained test period than the others. The findings emphasize how reinforcement learning can be utilized for intelligent load balancing with sustained performance improvement an‎d real-time responsiveness. Future research aims to extending the method to larger clusters of servers, implementing energy an‎d latency constraints, an‎d using multi-agent systems to improve both efficiency an‎d reliability in practical cloud computing environments.

راند رابين (RR) , كمترين بارگذاري (LL)، , شبكه Q عميق (DQN)،

Round Robin (RR) , Least Loaded (LL), , Deep Q Network (DQN), , Deep Reinforcement Learning (DRL).

Sumayah Al-Taie

DR.NASSER MOZAYANI