What this book covers

Chapter 1, Splitting Input Data, shows how to distribute machine learning training or serving workload on the input data dimension, which is called data parallelism.

Chapter 2, Parameter Server and All-Reduce, describes two widely-adopted model synchronization schemes in the data parallel training process.

Chapter 3, Building a Data Parallel Training and Serving Pipeline, illustrates how to implement data parallel training and the serving workflow.

Chapter 4, Bottlenecks and Solutions, describes how to improve data parallelism performance with some advanced techniques, such as more efficient communication protocols, reducing the memory footprint.

Chapter 5, Splitting the Model, introduces the vanilla model parallel approach in general.

Chapter 6, Pipeline Input and Layer Split, shows how to improve system efficiency with pipeline parallelism.

Chapter 7, Implementing Model Parallel Training and Serving Workflows, discusses how to implement model parallel training and serving in detail.

Chapter 8, Achieving Higher Throughput and Lower Latency, covers advanced schemes to reduce computation and memory consumption in model parallelism.

Chapter 9, A Hybrid of Data and Model Parallelism, combines data and model parallelism together as an advanced in-parallel model training/serving scheme.

Chapter 10, Federated Learning and Edge Devices, talks about federated learning and how edge devices are involved in this process.

Chapter 11, Elastic Model Training and Serving, describes a more efficient scheme that can change the number of accelerators used on the fly.

Chapter 12, Advanced Techniques for Further Speed-Ups, summarizes several useful tools, such as a performance debugging tool, job multiplexing, and heterogeneous model training.