Abstract
With advances in technologies such as the Internet of Things (IoT), cryptocurrencies,
and Artificial Intelligence (AI), data being generated has grown exponentially and demands
more compute-intensive resources. This demand has led to an increase in data
centres. But with this growth comes a concerning problem: temperature hotspots that
reduce equipment lifespan and degrade operational efficiency. Whereas traditional rulebased
cooling systems cannot properly respond to such dynamic environments, machine
learning, particularly deep learning, offers a better adaptive and efficient approach to
managing heat distribution. Nonetheless, the greater part of deep learning still operates
on single-thread architectures, thereby limiting their ability to harness fully modern distributed
systems. In order to overcome this limitation, we introduce VACTOR models, a
distributed, actor-based Variational Autoencoder that trains the model through messagedriven
interactions. An Actor processes tasks in isolation and exchanges messages that
pass data such as fragments of datasets, learning rates, and activation functions. VACTOR
achieves distributed processing to optimise resource usage and improve scalability.
Experiments show that this system enhances accuracy, reduces latency, and keeps the
workload more balanced, and it is a promising solution towards temperature hotspots
detection and control for data centres.