Deep Learning (DL) has had an immense success in the recent past, leading to state-of-the-art results in various domains, such as image recognition and natural language processing. One of the reasons for this success is the increasing size of DL models and the proliferation of vast amounts of training data being available. To keep on improving the performance of DL, increasing the scalability of DL systems is necessary. In this survey, we perform a broad and thorough investigation on challenges, techniques and tools for scalable DL on distributed infrastructures. This incorporates infrastructures for DL, methods for parallel DL training, multi-tenant resource scheduling, and the management of training and model data. Further, we analyze and compare 11 current open-source DL frameworks and tools and investigate which of the techniques are commonly implemented in practice. Finally, we highlight future research trends in DL systems that deserve further research.
Stream Processing (SP) has evolved as the leading paradigm to process and gain value from the high volume of streaming data produced, e.g., in the domain of the Internet of Things. An SP system is a middleware that deploys a network of operators between data sources, such as sensors, and the consuming applications. SP systems typically face intense and highly dynamic data streams. Parallelization and elasticity enable SP systems to process these streams with continuous high quality of service. The current research landscape provides a broad spectrum of methods for parallelization and elasticity in SP. Each method makes specific assumptions and focuses on particular aspects. However, the literature lacks a comprehensive overview and categorization of the state of the art in SP parallelization and elasticity, which is necessary to consolidate the state of the research and to plan future research directions on this basis. Therefore, in this survey, we study the literature and develop a classification of current methods for both parallelization and elasticity in SP systems.
Graph neural networks (GNNs) are an emerging research field. This specialized deep neural network architecture is capable of processing graph structured data and bridges the gap between graph processing and deep learning. As graphs are everywhere, GNNs can be applied to various domains including recommendation systems, computer vision, natural language processing, biology, and chemistry. With the rapid growing size of real-world graphs, the need for efficient and scalable GNN training solutions has come. Consequently, many works proposing GNN systems have emerged throughout the past few years. However, there is an acute lack of overview, categorization, and comparison of such systems. We aim to fill this gap by summarizing and categorizing important methods and techniques for large-scale GNN solutions. Additionally, we establish connections between GNN systems, graph processing systems, and deep learning systems.
Federated Learning (FL) has become an established technique to facilitate privacy-preserving collaborative training across a multitude of clients. However, new approaches to FL often discuss their contributions involving small deep-learning models only and focus on training full models on clients. In the wake of Foundation Models (FM), the reality is different for many deep learning applications. Typically, FMs have already been pre-trained across a wide variety of tasks and can be fine-tuned to specific downstream tasks over significantly smaller datasets than required for full model training. However, access to such datasets is often challenging. By its design, FL can help to open data silos. With this survey, we introduce a novel taxonomy focused on computational and communication efficiency, the vital elements to make use of FMs in FL systems. We discuss the benefits and drawbacks of parameter-efficient fine-tuning (PEFT) for FL applications, elaborate on the readiness of FL frameworks to work with FMs and provide future research opportunities on how to evaluate generative models in FL as well as the interplay of privacy and PEFT.
With the emergence of AI regulations, such as the EU AI Act, requirements for simple data lineage, enforcement of low data bias, and energy efficiency have become a priority for everyone offering AI services. Being pre-trained on versatile and a vast amount of data, large language models and foundation models (FMs) offer a good basis for building high-quality deep learning pipelines. Fine-tuning can further improve model performance on a specific downstream task, which requires orders of magnitude less data than pre-training. Often, access to high-quality and low-bias data for model fine-tuning is limited due to technical or regulatory requirements. Federated learning (FL), as a distributed and privacy-preserving technique, offers a well-suited approach to significantly expanding data access for model fine-tuning. Yet, this data is often located on the network edge, where energy, computational, and communication resources are significantly more limited than in data centers. In our paper, we conduct an end-to-end evaluation for fine-tuning the FLAN-T5 FM family on the network edge. We study energy efficiency potentials throughout FL systems - on clients, in communication, and on the server. Our analysis introduces energy efficiency as a real-time metric to assess the computational efficiency of an FL system. We show the stark need for further improvements in communication efficiency when working with FMs and demonstrate the importance of adaptive FL optimizers for FM training.
18 January 2018 in IEEE Transactions on Parallel and Distributed Systems
Recently, graph processing systems such as Pregel, PowerGraph, or GraphX gained popularity due to their superior performance of data analytics on graph-structured data. These systems employ partitioning algorithms to parallelize graph execution while minimizing inter-partition communication. Recent partitioning algorithms, however, unrealistically assume a uniform and constant amount of data exchanged between graph vertices (i.e., uniform vertex traffic) and homogeneous network costs between machines hosting the graph partitions. This leads to suboptimal partitioning decisions and inefficient graph processing. To this end, we developed GrapH, the first graph processing system using vertex-cut graph partitioning that considers both, diverse vertex traffic and heterogeneous network costs. The main idea is to avoid frequent communication over expensive network links using an adaptive edge migration strategy. Our evaluations show an improvement of 10% in graph processing latency and 60% in communication costs compared to state-of-the-art partitioning approaches.