Range-top-k queries retrieve the top-k elements within an arbitrary subrange of a large array and are a key primitive in real-time analytics. Unlike one-shot top-k selection, practical deployments issue large volumes of queries over varying and often overlapping ranges, frequently interleaved with streaming updates. In this setting, applying conventional GPU top-k kernels per query is inefficient: each query triggers range rescans or O(n)-scale passes that overwhelm HBM bandwidth, thrash on-chip caches, and provide little reuse across overlapping windows.

Large Language Model (LLM) inference has emerged as a fundamental paradigm, however, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose STAR, a decode rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with a dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 75.1% and achieving 2.63 × higher goodput.

Training large language models faces frequent interruptions due to various faults, demanding robust fault-tolerance. Existing backup-free methods, such as redundant computation, dynamic parallelism, and data rerouting, each incur performance penalties, whether from ongoing overhead, lengthy reconfigurations, or post-recovery inefficiencies. We propose Chameleon, an adaptive fault-tolerant system that intelligently selects optimal recovery strategies when a failure occurs. Chameleon achieves this through a unified performance model, expedient execution plan search, accurate performance estimation, and efficient communication optimizations. Experiments on a 32-card cluster show that Chameleon maintains a performance gap of within 11.00% between post-recovery and failure-free training, while preserving model convergence and efficient memory usage. Compared to state-of-the-art methods, Chameleon achieves up to 1.229x and 1.355x higher average throughput than Oobleck and Recycle, respectively.

With the rise of deep learning, many companies have developed domain-specific architectures (DSAs) optimized for AI workloads, with Ascend being a representative. To fully realize the operator performance on Ascend, effective analysis and optimization is urgently needed. Compared to GPU, Ascend requires users to manage operations manually, leading to complex performance issues that require precise analysis. However, existing roofline models face challenges of visualization complexity and inaccurate performance assessment. To address these needs, we introduce a component-based roofline model that abstracts components to capture operator performance, thereby effectively identifying bottleneck components. Furthermore, through practical operator optimization case studies, we illustrate a comprehensive process of optimization based on roofline analysis, summarizing common performance issues and optimization strategies. Finally, extensive end-to-end optimization experiments demonstrate significant model speed improvements, ranging from 1.07× to 2.15×, along with valuable insights from practice.

Random walks are a primary means for extracting information from large-scale graphs. While most real-world graphs are inherently dynamic, state-of-the-art random walk engines failed to efficiently support such a critical use case. This paper takes the initiative to build a general random walk engine for dynamically changing graphs with two key principles: (i) This system should support both low-latency streaming updates and high-throughput batched updates. (ii) This system should achieve fast sampling speed while maintaining acceptable space consumption to support dynamic graph updates. Upholding both standards, we introduce Bingo, a GPU-based random walk engine for dynamically changing graphs. First, we propose a novel radix-based bias factorization algorithm to support constant time sampling complexity while supporting fast streaming updates. Second, we present a group-adaption design to reduce space consumption dramatically. Third, we incorporate GPU-aware designs to support high-throughput batched graph updates on massively parallel platforms. Together, Bingo outperforms existing efforts across various applications, settings, and datasets, achieving up to a 271.11x speedup compared to the state-of-the-art efforts.

The Louvain algorithm is one of the most popular algorithms for community detection. Observing that existing implementations suffer from inaccurate pruning and inefficient intermediate state management, we introduce GALA, GPU-Accelerated Louvain Algorithm, which incorporates two key innovations. The first innovation is a novel modularity gain-based pruning strategy, supported by rigorous theoretical guarantees of optimality and able to reduce up to 76% of vertices as well as their corresponding computations. To take advantage of the memory hierarchy and parallelism of GPUs, the second innovation is workload-aware kernels, featuring a shuffle-based kernel founded on the warp-level primitives for exchange states and a hash-based kernel that prioritizes shared memory in hashtable design. GALA further scales to multiple GPUs by minimizing the synchronization overhead between GPUs through a dense-sparse synchronization strategy. We evaluate the performance of GALA through theoretical analysis and practical experiments on various real-world graphs. The experimental results confirm that GALA significantly improves the performance of the parallel Louvain algorithm on GPUs, surpassing state-of-the-art solutions by 6× on average.

Graph data analysis, particularly local triangle counting, plays a pivotal role in deciphering complex relationships within graph data. This method is invaluable across diverse fields such as social networks, transportation, and cybersecurity. However, this process often involves handling sensitive information, necessitating that the relationship between any two nodes is considered private. Differential privacy (DP) is a formal model to address privacy concerns and can be categorized into two types: the central DP (CDP) model, which achieves better result accuracy, and the local DP (LDP) model, which does not assume a trusted server. To bridge the gap between the two models, we propose Sectric, a server-aided crypto-assisted local triangle counting protocol, in this paper. It can achieve the same result accuracy with the same privacy budget as the CDP model without assuming a trusted server. Sectric also explores a new approach in crypto-assisted graph data analysis algorithms that represents a node's neighbors using a set instead of an adjacency vector, and successfully achieves higher efficiency compared to other crypto-assisted solutions. We also conduct theoretical and empirical evaluations to demonstrate that Sectric achieves the design principles.

Data lakes, increasingly adopted for their ability to store and analyze diverse types of data, commonly use columnar storage formats like Parquet and ORC for handling relational tables. However, these traditional setups fall short when it comes to efficiently managing graph data, particularly those conforming to the Labeled Property Graph (LPG) model. To address this gap, this paper introduces GraphAr, a specialized storage scheme designed to enhance existing data lakes for efficient graph data management. Leveraging the strengths of Parquet, GraphAr captures LPG semantics precisely and facilitates graph-specific operations such as neighbor retrieval and label filtering. Through innovative data organization, encoding, and decoding techniques, GraphAr dramatically improves performance. Our evaluations reveal that GraphAr outperforms conventional Parquet and Acero-based methods, achieving an average speedup of 4452× for neighbor retrieval, 14.8× for label filtering, and 29.5× for end-to-end workloads. These findings highlight GraphAr's potential to extend the utility of data lakes by enabling efficient graph data management.

Butterfly (a cyclic graph motif) counting is a fundamental task with many applications in graph analysis, which aims at computing the number of butterflies in a large graph. With the rapid growth of graph data, it is more and more challenging to do butterfly counting due to the super-linear time complexity and large memory consumption. In this paper, we study I/O-efficient algorithms for doing butterfly counting on hierarchical memory. Existing algorithms of the kind cannot guarantee I/O optimality. Observing that in order to count butterflies, it suffices to "witness" a subgraph instead of the whole structure, a new class of algorithms called semi-witnessing algorithm is proposed. We prove that a semi-witnessing algorithm is not restricted by the lower bound Ømega(|E|2/MB) of a witnessing algorithm, and give a new bound of Ømega(min(|E|2/MB, |E|/|V| √M B)). We further develop the IOBufs algorithm that manages to approach the I/O lower bound, and thus claim its optimality. Finally, we make efforts to parallelize IOBufs to further improve the performance and scalability. We show in the experiment that IOBufs significantly outperforms the state-of-the-art algorithms EMRC and BFC-EM. In addition, IOBufs can scale to conducting butterfly counting on the Clueweb graph with 37 billion edges and quintillions (10^18 ) of butterflies.

Triangle counting is a building block for a wide range of graph applications. Traditional wisdom suggests that i) hashing is not suitable for triangle counting, ii) edge-centric triangle counting beats vertex-centric design, and iii) communication-free and workload balanced graph partitioning is a grand challenge for triangle counting. On the contrary, we advocate that i) hashing can help the key operations for scalable triangle counting on Graphics Processing Units (GPUs), i.e., list intersection and graph partitioning, ii) vertex-centric design reduces both hash table construction cost and memory consumption, which is limited on GPUs. In addition, iii) we exploit graph and workload collaborative, and hashing-based 2D partitioning to scale vertex-centric triangle counting over 1000 GPUs with sustained scalability. In this article, we present Trust which performs triangle counting with the hash operation and vertex-centric mechanism at the core. To the best of our knowledge, Trust is the first work that achieves over one trillion Traversed Edges Per Second (TEPS) rate for triangle counting.