Adaptiveness, Asynchrony, and Resource Efficiency in Parallel Stochastic Gradient Descent

Abstract: Accelerated digitalization and sensor deployment in society in recent years poses critical challenges for associated data processing and analysis infrastructure to scale, and the field of big data, targeting methods for storing, processing, and revealing patterns in huge data sets, has surged. Artificial Intelligence (AI) models are used diligently in standard Big Data pipelines due to their tremendous success across various data analysis tasks, however exponential growth in Volume, Variety and Velocity of Big Data (known as its three V’s) in recent years require associated complexity in the AI models that analyze it, as well as the Machine Learning (ML) processes required to train them. In order to cope, parallelism in ML is standard nowadays, with the aim to better utilize contemporary computing infrastructure, whether it being shared-memory multi-core CPUs, or vast connected networks of IoT devices engaging in Federated Learning (FL). Stochastic Gradient Descent (SGD) serves as the backbone of many of the most popular ML methods, including in particular Deep Learning. However, SGD has inherently sequential semantics, and is not trivially parallelizable without imposing strict synchronization, with associated bottlenecks. Asynchronous SGD (AsyncSGD), which relaxes the original semantics, has gained significant interest in recent years due to promising results that show speedup in certain contexts. However, the relaxed semantics that asynchrony entails give rise to fundamental questions regarding AsyncSGD, relating particularly to its stability and convergence rate in practical applications. This thesis explores vital knowledge gaps of AsyncSGD, and contributes in particular to: Theoretical frameworks – Formalization of several key notions related to the impact of asynchrony on the convergence, guiding future development of AsyncSGD implementations; Analytical results – Asymptotic convergence bounds under realistic assumptions. Moreover, several technical solutions are proposed, targeting in particular: Stability – Reducing the number of non-converging executions and the associated wasted energy; Speedup – Improving convergence time and reliability with instance-based adaptiveness; Elasticity – Resource-efficiency by avoiding over-parallelism, and thereby improving stability and saving computing resources. The proposed methods are evaluated on several standard DL benchmarking applications and compared to relevant baselines from previous literature. Key results include: (i) persistent speedup compared to baselines, (ii) increased stability and reduced risk for non-converging executions, (iii) reduction in the overall memory footprint (up to 17%), as well as the consumed computing resources (up to 67%). In addition, along with this thesis, an open-source implementation is published, that connects high-level ML operations with asynchronous implementations with fine-grained memory operations, leveraging future research for efficient adaptation of AsyncSGD for practical applications.