Tabular data embedded in PDF files, web pages, and other types of documents is prevalent in various domains. These tables, which we call human-centric tables (HCTs for short), are dense in information but often exhibit complex structural and semantic layouts. To query these HCTs, some existing solutions focus on transforming them into relational formats. However, they fail to handle the diverse and complex layouts of HCTs, making them not amenable to easy querying with SQL-based approaches. Another emerging option is to use Large Language Models (LLMs) and Vision Language Models (VLMs). However, there is a lack of standard evaluation benchmarks to measure and compare the performance of models to query HCTs using natural language. To address this gap, we propose the HumanCentric Tables Question-Answering extensive benchmark (HCTQA) consisting of thousands of HCTs with several thousands of natural language questions with their respective answers. More specifically, HCT-QA includes 1,880 real-world HCTs with 9,835 QA pairs in addition to 4,679 synthetic HCTs with 67.7K QA pairs. Also, we show through extensive experiments the performance of 25 and 9 different LLMS and VLMs, respectively, in an answering HCT-QA's questions. In addition, we show how finetuning an LLM on HCT-QA improves F1 scores by up to 25 percentage points compared to the off-the-shelf model. Compared to existing benchmarks, HCT-QA stands out for its broad complexity and diversity of covered HCTs and generated questions, its comprehensive metadata enabling deeper insight and analysis, and its novel synthetic data and QA generator.

Graph Neural Networks (GNNs) with learnable vertex embeddings enable models to infer rich, task-specific representations even when vertex features are sparse, noisy, or missing. In large-scale multi-GPU training, dynamically updated embeddings, often orders of magnitude larger than model parameters, severely degrade training efficiency. Specifically, loading remote embeddings and synchronizing their gradients collectively account for over 90% of per-iteration time. Traditional caching and parallelism approaches, designed for static embeddings or model parameters alone, are ineffective at mitigating this “data wall” of embedding-related transfers. To address this, we begin with a detailed analysis of vertex access patterns over training iterations and find that infrequently sampled vertices, despite incurring the majority of embedding-loading latency, undergo very few updates, making their embeddings ideal candidates for staleness reuse. Driven by this, we propose GLPilot, a novel system that mitigates embedding-related bottlenecks. GLPilot introduces a staleness-bounded embedding buffering mechanism to reduce remote fetches and a local gradient aggregation technique to minimize redundant communications during synchronization. Additionally, GLPilot utilizes an on-GPU cache for keeping mostly updated embeddings to alleviate CPU-GPU data transfer bottlenecks. Our evaluations on a 32-GPU cluster using two popular GNN models, three datasets and two optimizers demonstrate that GLPilot consistently achieves 1.28–1.93× per-epoch training speedups, in comparison with two strong baselines such as DGL and P3, while maintaining comparable model accuracy.

The growth of Large Language Models (LLMs) has necessitated large-scale distributed training. Highly optimized frameworks, however, suffer significant losses in MFU (Model FLOPS Utilization) due to communication. This paper introduces DHeLlam, a novel micro-structure inspired by DNA that significantly enhances the efficiency of LLM training. Central to DHeLlam is Strand Interleaving (SI), which treats the continuous stream of training micro-batches on a GPU as two interleaved strands. DHeLlam co-schedules their forward and backward passes using operator-level overlap profiling and a dynamic programming-based search. It enables the two strands to share model states and activation memory, requiring$<3 \%$additional HBM space under common model configurations. To our best knowledge, DHeLlam is the first to co-execute two microbatches without requiring model replication. With its unique model folding design, DHeLlam seamlessly integrates with all forms of data and model parallelism, including the challenging pipeline parallelism (with a W-shaped pipeline). We evaluated DHeLlam training with the popular Llama and GPT dense models, plus the Phi Mixture of Expert (MoE) model, across 2 GPU clusters. Results show that it achieves 12-40% throughput (up to 58% MFU) and 5-24% throughput (up to 64% MFU) improvement on the 64-card A40 and A800 clusters respectively, significantly outperforming state-of-the-art methods.

Transaction performance in geo-replicated databases heavily relies on the request location: when not issued by the primary region, transactions are forced to involve costly wide-area communication. While existing systems distribute primary roles across regions, such assignment typically occurs at the shard level, making it difficult to align with geographically dispersed access to individual records. This paper introduces PolyBase, a pioneering architecture to address such misalignment, leveraging the widely adopted Paxos-based log replication mechanisms. It enables flexible row-level consensus group affiliation , which runs on an unchanged Paxos protocol , but dynamically re-assigns database rows between Paxos log replication groups, whose leaders become the primary region, enjoying faster writes and up-to-date versions for reads. With carefully designed data structures and protocols, PolyBase significantly reduces wide-area RTTs without compromising transaction or log replication consistency or reliability guarantees. We implemented PolyBase with optimized re-assignment policies and integrated it into two popular databases (RocksDB and MySQL). Our evaluation on AWS, using a production e-commerce workload and microbench-marks confirms that PolyBase offers significantly higher transaction throughput and lower average/tail latency compared to baselines.

The recent emergence of commodity persistent memory (PM) hardware has altered the landscape of the storage hierarchy. It brings multi-fold benefits to database systems, with its large capacity, low latency, byte addressability, and persistence. However, PM has not been incorporated into the popular disaggregated architecture of cloud-native databases.

A distributed database utilizing the wide-spread edge computing servers to provide low-latency data access with the serializability guarantee is highly desirable for emerging edge computing applications. In an edge database, nodes are divided into regions, and a transaction can be categorized as intra-region (IRT) or cross-region (CRT) based on whether it accesses data in different regions. In addition to serializability, we insist that a practical edge database should provide low tail latency for both IRTs and CRTs, and such low latency must be scalable to a large number of regions. Unfortunately, none of existing geo-replicated serializable databases or edge databases can meet such requirements.

It is challenging for cloud-native relational databases to meet the ever-increasing needs of scaling compute and memory resources independently and elastically. The recent emergence of memory disaggregation architecture, relying on high-speed RDMA network, offers opportunities to build cost-effective and elastic cloud-native databases. There exist proposals to let unmodified applications run transparently on disaggregated systems. However, running relational database kernel atop such proposals experiences notable performance degradation and time-consuming failure recovery, offsetting the benefits of disaggregation. To address these challenges, in this paper, we propose a novel database architecture called LegoBase, which explores the co-design of database kernel and memory disaggregation. It pushes the memory management back to the database layer for bypassing the Linux I/O stack and re-using or designing (remote) memory access optimizations with an understanding of data access patterns. LegoBase further splits the conventional ARIES fault tolerance protocol to independently handle the local and remote memory failures for fast recovery of compute instances. We implemented LegoBase atop MySQL. We compare LegoBase against MySQL running on a standalone machine and the state-of-the-art disaggregation proposal Infiniswap. Our evaluation shows that even with a large fraction of data placed on the remote memory, LegoBase's system performance in terms of throughput (up to 9.41% drop) and P99 latency (up to 11.58% increase) is comparable to the monolithic MySQL setup, and significantly outperforms (1.99x-2.33x, respectively) the deployment of MySQL over Infiniswap. Meanwhile, LegoBase introduces an up to 3.87x and 5.48x speedup of the recovery and warm-up time, respectively, over the monolithic MySQL and MySQL over Infiniswap, when handling failures or planned re-configurations.

Key-value (KV) stores support many crucial applications and services. They perform fast in-memory processing but are still often limited by I/O performance. The recent emergence of high-speed commodity non-volatile memory express solid-state drives (NVMe SSDs) has propelled new KV system designs that take advantage of their ultra-low latency and high bandwidth. Meanwhile, to switch to entirely new data layouts and scale up entire databases to high-end SSDs requires considerable investment. As a compromise, we propose SpanDB, an LSM-tree-based KV store that adapts the popular RocksDB system to utilize selective deployment of high-speed SSDs . SpanDB allows users to host the bulk of their data on cheaper and larger SSDs (and even hard disc drives with certain workloads), while relocating write-ahead logs (WAL) and the top levels of the LSM-tree to a much smaller and faster NVMe SSD. To better utilize this fast disk, SpanDB provides high-speed, parallel WAL writes via SPDK, and enables asynchronous request processing to mitigate inter-thread synchronization overhead and work efficiently with polling-based I/O. To ease the live data migration between fast and slow disks, we introduce TopFS, a stripped-down file system providing familiar file interface wrappers on top of SPDK I/O. Our evaluation shows that SpanDB simultaneously improves RocksDB's throughput by up to 8.8 \times and reduces its latency by 9.5–58.3%. Compared with KVell, a system designed for high-end SSDs, SpanDB achieves 96–140% of its throughput, with a 2.3–21.6 \times lower latency, at a cheaper storage configuration.

Better understanding of mobile applications' behaviors would lead to better malware detection/classification and better app recommendation for users. In this work, we design a framework AppDNA to automatically generate a compact representation for each app to comprehensively profile its behaviors. The behavior difference between two apps can be measured by the distance between their representations. As a result, the versatile representation can be generated once for each app, and then be used for a wide variety of objectives, including malware detection, app categorizing, plagiarism detection, etc. Based on a systematic and deep understanding of an app's behavior, we propose to perform a function-call-graph-based app profiling. We carefully design a graph-encoding method to convert a typically extremely large call-graph to a 64-dimension fix-size vector to achieve robust app profiling. Our extensive evaluations based on 86,332 benign and malicious apps demonstrate that our system performs app profiling (thus malware detection, classification, and app recommendation) to a high accuracy with extremely low computation cost: it classifies 4024 (benign/malware) apps using around 5.06 second with accuracy about 93.07%; it classifies 570 malware's family (total 21 families) using around 0.83 second with accuracy 82.3%; it classifies 9,730 apps' functionality with accuracy 33.3% for a total of 7 categories and accuracy of 88.1 % for 2 categories.