Architectural frameworks to leverage new hardware technologies for emerging data-intensive applications

点击翻译按钮获取中文摘要

Important emerging data-intensive applications in machine learning and robotics utilize computing systems at different levels: datacenters, edge devices, and on resource-constrained autonomous vehicles (e.g., drones). These applications require processing significant amounts of data under various system constraints and their scale and complexity are today limited by constraints in memory, compute, and network capabilities. Advances in hardware technologies such as disaggregated memory, near-data processing, and application-specific acceleration offer promising opportunities to address these bottlenecks. Leveraging these technologies effectively requires co-designing algorithms, systems, and architectures. This research program aims to develop new architectural frameworks and cross-layer solutions to leverage new hardware technologies in three contexts: data centers, edge devices, and low-power autonomous vehicles. The proposed research aims to tackle the following major thrusts. First, we aim to investigate the system challenges to efficiently incorporate memory disaggregation in datacenters. Disaggregation rethinks the traditional notion of datacenters comprising monolithic servers with memory attached over the memory bus. Instead processors are connected to network-attached pools of memory that are independently operated. Disaggregation offers an opportunity to meet the significant memory needs of emerging data-intensive applications such training of large-scale machine learning models at low cost. Second, we aim to investigate the implications of the push towards edge-based computing in large-scale data-intensive applications. Important applications for machine learning in health care, mobile applications, financial institutions require privacy of user data, necessitating partial training of data on edge devices/servers. These applications typically use federated and incremental training to train machine learning models while preserving the privacy of user data. We aim to tackle the architectural and programmability challenges associated with efficient deployment of federated learning in edge devices. Third, we aim to investigate the architectural challenges of supporting robotics tasks on resource-constrained autonomous vehicles such as Unmanned Aerial Vehicles (UAVs) and vehicles for micromobility. These systems are projected to have tremendous growth in demand with use cases in healthcare, rescue, delivery, and mobility. These vehicles are required to support data-intensive tasks such as processing sensor data from LIDAR, cameras, etc., complex localization and mapping, and DNN inference, while still being heavily constrained in power and compute capability. This thrust will involve 1) investigating key compute bottlenecks for each task; 2) developing infrastructure to enable cross-layer research for each application domain; 3) developing hardware-software co-designs to enable efficient processing of these tasks.