Graph Neural Networks (GNNs) have emerged as a powerful tool for analyzing relationship-based data, such as those found in social networks, logistics, weather forecasting, and other domains. Inference and training with GNN models execute slowly, bottlenecked by limited data bandwidths between memory and GPU hosts, as a result of the many irregular memory accesses inherent to GNN-based computation. To overcome these limitations, we present GLEAM, a Processing-in-Memory (PIM) hardware accelerator designed specifically for GNN-based training and inference. GLEAM units are placed per-bank and leverage the much larger, internal bandwidth of HBMs to handle GNNs' irregular memory accesses, significantly boosting performance and reducing the energy consumption entailed by the dominant activity of GNN-based computation: neighbor aggregation. Our evaluation of GLEAM demonstrates up to a 10x speedup for GNN inference over GPU baselines, alongside a significant reduction in energy usage.

Given billions of mobile users, the environmental impact of edge computing is significant. To address this, future applications need to execute computations on a green component which is fueled by renewable energy sources. However, because of the intermittent nature of the renewable energy sources, the carbon intensity of computing components can significantly vary with location and time of use. This poses a new challenge for edge applications – deciding when and where to run computations across consumer devices at the edge and servers in the cloud. Such scheduling decisions become more complicated with the amortization of the rising embodied emissions and stochastic runtime variance. This work proposes GreenScale, an intelligent execution scaling engine that accurately selects the carbon-optimal execution target for edge applications in different runtime environments. Our evaluation with three representative categories of applications (i.e., AI, Game, and AR/VR) demonstrate that the carbon emissions of the applications can be reduced by 35.2%, on average, with GreenScale.

As machine learning models become more widespread, they are being increasingly applied in applications that heavily impact people’s lives (e.g., medical diagnoses, judicial system sentences, etc.). Several communities are thus calling for ML models to be not only accurate, but also explainable. To achieve this, recommendations must be augmented with explanations summarizing how each recommendation outcome is derived. Explainable Random Forest (XRF) models are popular choices in this space, as they are both very accurate and can be augmented with explainability functionality, allowing end-users to learn how and why a specific outcome was reached. However, the limitations of XRF models hamper their adoption, the foremost being the high computational demands associated with training such models to support high-accuracy classifications, while also annotating them with explainability meta-data.

With the rise of graph-based algorithms in many applications, dynamic graphs have become critical for applications that work with real-time or time-series relationship data. Due to their large storage footprint, these applications require high parallelism and locality. Many-core architectures offer high parallelism for both static and dynamic graphs. However, systems operating on dynamic graphs must continuously repartition graph data across storage units to achieve good locality, load balancing, and performance. In this work, we examine the effectiveness and efficiency of different vertex-pressure repartitioning schemes, which move vertices so to co-locate them near their most relevant neighbors. We describe key repartitioning design choices and provide a thorough evaluation of the impact of a range of design features with different datasets. Our evaluation indicates that optimized dynamic repartitioning techniques can often provide over 2x performance speedup over state-of-the-art static solutions.

Graph-based algorithms have gained significant interest in several application domains. Solutions addressing the computational efficiency of such algorithms have mostly relied on many-core architectures. Cleverly laying out input graphs in storage, by placing adjacent vertices in a same storage unit (memory bank or cache unit), enables fast access during graph traversal. Dynamic graphs, however, must be continuously repartitioned to leverage this benefit. Yet software repartitioning solutions rely on costly, cross-vault communication to query and optimize the graph layout between algorithm iterations.