Search for: All records

The rapid advancement of large language models (LLMs) has caused their parameter sizes to grow beyond the memory capacity of a single GPU. Although distributed inference across multiple GPUs is a solution in enterprise settings, it remains inaccessible for most non-commercial users. Thus, there is a growing demand to run LLMs on a single GPU when the model does not fit entirely in GPU memory. A common approach is to offload parts of the model from the GPU to the CPU during inference. However, repeatedly transfer- ring parameters between these devices incurs significant overhead. To address this challenge, we propose a new buffer management policy, LIRS-M, which maximizes buffer hits and minimizes data transfer. Experimental results show that our approach achieves a 2.0× speedup compared to SoTA offloading techniques while delivering robust buffer-hit performance. 

Query optimization is critical in relational database management systems (DBMSs) for ensuring efficient query processing. The query optimizer relies on precise selectivity and cost estimates to generate optimal query plans for execution. However, this static query optimization approach falls short for DBMSs handling machine learning (ML) queries. ML-centric DBMSs face distinct challenges in query optimization. First, performance bottlenecks shift to user-defined functions (UDFs), often encapsulating deep learning models, making it difficult to estimate UDF statistics without profiling the query. Second, optimal query plans for ML queries are data-dependent, requiring dynamic plan adjustments during execution. To address these challenges, we introduce Aero, an ML-centric DBMS that utilizes adaptive query processing (AQP) for efficiently processing ML queries. Aero optimizes the evaluation of UDF-based query predicates by dynamically adjusting predicate evaluation order and enhancing UDF execution scalability. By integrating AQP, Aero continuously monitors UDF statistics, routes data to predicates in an optimal order, and dynamically allocates resources for evaluating predicates. Aero achieves up to 6.4x speedup compared to a state-of-the-art ML-centric DBMS across four diverse use cases, with no impact on accuracy. 

In this demonstration, we will present EVA, an end-to-end AI-Relational database management system. We will demonstrate the capabilities and utility of EVA using three usage scenarios: (1) EVA serves as a backend for an exploratory video analytics interface developed using Streamlit and React, (2) EVA seamlessly integrates with the Python and Data Science ecosystems by allowing users to access EVA in a Python notebook alongside other popular libraries such as Pandas and Matplotlib, and (3) EVA facilitates bulk labeling with Label Studio, a widely-used labeling framework. By optimizing complex vision queries, we illustrate how EVA allows a wide range of application developers to harness the recent advances in computer vision. 

In recent years, deep learning models have revolutionized computer vision, enabling diverse applications. However, these models are computationally expensive, and leveraging them for video analyt- ics involves low-level imperative programming. To address these efficiency and usability challenges, the database community has de- veloped video database management systems (VDBMSs). However, existing VDBMSs lack extensibility and composability and do not support holistic system optimizations, limiting their practical appli- cation. In response to these issues, we present our vision for EVA, a VDBMS that allows for extensible support of user-defined functions and employs a Cascades-style query optimizer. Additionally, we leverage RAY’s distributed execution to enhance scalability and performance and explore hardware-specific optimizations to facilitate runtime optimizations. We discuss the architecture and design of EVA, our achievements thus far, and our research roadmap.