The use of GPUs has proliferated for machine learning workflows and is now considered mainstream for many deep learning models. Meanwhile, when training state-of-the-art personal recommendation models, which consume the highest number of compute cycles at our large-scale datacenters, the use of GPUs came with various challenges due to having both compute-intensive and memory-intensive components. GPU performance and efficiency of these recommendation models are largely affected by model architecture configurations such as dense and sparse features, MLP dimensions. Furthermore, these models often contain large embedding tables that do not fit into limited GPU memory. The goal of this paper is to explain the intricacies of using GPUs for training recommendation models, factors affecting hardware efficiency at scale, and learnings from a new scale-up GPU server design, Zion.

This paper examines combined spatiotemporal denoising and clustering of functional magnetic resonance imaging (fMRI) time series. Most fMRI denoising methods are implemented either in spatial or temporal domain without taking into account both space and time information. In this work, a spatiotemporal denoising method is developed where spatial denoising is implemented by Bayesian shrinkage that uses temporal prior information obtained by statistical testing on all voxel time courses. After the denoising, a set of spatiotemporal features are extracted and characterized by a Gaussian mixture model, which is applied to detect activated areas. The proposed methods have been tested on both synthetic and experimental data, and the results demonstrate their effectiveness.

Integration of middleware scheduling and resource management services enables open distributed real-time embedded (DRE) applications to meet end-to-end quality of service (QoS) requirements in highly variable operating environments. This paper describes our research on integrating CORBA scheduling and resource management services, and presents experiments we conducted to validate and quantify the benefits of this integration. Our experimental results show that integrating distributed scheduling and resource management in middleware for open DRE systems can offer significant improvements in predictability. Specifically, integrating our stand-alone resource management service with a previously unmanaged experimental baseline application reduced the ratio of missed deadlines from 26% to 10%, and the same application performed even better under the control of integrated scheduling and resource management services, with a missed deadline ratio of only 1%.

We describe the design and implementation of a static real-time data distribution mechanism that uses a real-time event service and a real-time scheduling service to ensure the ontime, and temporally valid delivery of data. The mechanism implements an algorithm, called the just-in-time data distribution algorithm to determine scheduling parameters that will ensure that data that arrives at a requesting target is valid. We use a navy weapons alignment application to demonstrate the necessity and usefulness of the algorithm and implementation. We also present results of tests that demonstrate that our implementation fulfills the guarantees predicted by the theoretical results.