Abstract

Large-scale computer platforms that process users' online requests must be capable of handling unexpected spikes in arrival rates. These platforms, which are composed of distributed components, can be configured with parameters to ensure both the quality of the results obtained for each request and low response times. In this work, we propose a dynamic optimization engine based on metric space indexing to address this problem. The engine is integrated into the platform and periodically monitors performance metrics to determine whether new configuration parameter values need to be computed. Our case study focuses on a P2P platform designed for classifying crowdsourced images related to natural disasters. We evaluate our approach under scenarios with high and low workloads, comparing it against alternative methods based on deep reinforcement learning. The results show that our approach reduces processing time by an average of 40%.