Distributed training is a crucial methodology for training complex machine learning models that are too large to fit into the memory of a single processor, enabling the handling of massive datasets by distributing the workload across multiple processors, known as worker nodes. This process is executed through two main approaches: data parallelism, which splits data across workers who each hold a replica of the model and perform training on their data subset, and model parallelism, which divides the model itself to run concurrently across different workers. Synchronous training ensures all workers update weights simultaneously, while asynchronous training allows workers to operate independently, often using a parameter server to manage model parameters. Distributed training can be centralized, involving a parameter server, or decentralized, involving peer-to-peer communication among nodes, and it provides benefits such as fault tolerance, efficiency, scalability, and cost-effectiveness. Frameworks like Horovod, Elephas, Amazon Sagemaker, TensorFlow, and PyTorch support distributed training, aiding in scaling deep learning models across multiple machines and enhancing performance in handling complex tasks involving large amounts of data.