Searcharxiv⌕ Search

arXiv subjects

Patrick P. C. Lee

Publications and source records attributed to Patrick P. C. Lee.

At least 19 recordsLinked to original sources

FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving

Mixture-of-Experts (MoE) models have become mainstream for scaling language models to hundreds of billions of expert parameters. Despite sparse expert activation, existing inference engines keep all experts GPU-resident, crowding out the key-value cache in large-batch, long-output offline workloads. We present FluxMoE, which decouples experts from physical GPU residency and adapts their footprint to available memory through a new \emph{expert paging} abstraction. FluxMoE combines PagedTensor for transparent remapping, a bandwidth-balanced hierarchy spanning losslessly compressed GPU memory and host DRAM, and a budget-aware residency planner. Unlike CPU-GPU co-inference and whole-layer offloading, FluxMoE streams weights on demand while keeping expert computation on GPUs. We implement FluxMoE atop vLLM and evaluate it on three MoE models. For GLM-4.5 on 8$\times$H20 GPUs, FluxMoE delivers up to 7.2$\times$ vLLM's throughput and 79.0\% lower average Time-Per-Output-Token (TPOT), without measurable model-quality loss using lossless compression. For Mixtral-8$\times$7B-Instruct on 2$\times$L40S GPUs, where weight-resident vLLM cannot fit, FluxMoE delivers 4.3$\times$ KTransformers's throughput and 29.1\% lower average TPOT.

cs.LG↗

Fletch: File-System Metadata Caching in Programmable Switches

Fast and scalable metadata management across multiple metadata servers is crucial for distributed file systems to handle numerous files and directories. Client-side caching of frequently accessed metadata can mitigate server loads, but it incurs significant overhead and complexity in maintaining cache consistency when the number of clients increases. We propose Fletch, an in-switch file-system metadata caching framework that leverages programmable switches to serve file-system metadata requests from multiple clients directly in the switch data plane. Unlike prior in-switch key-value caching systems, Fletch addresses file-system-specific path dependencies under stringent switch resource constraints. We implement Fletch atop HDFS and evaluate it on a Tofino-switch testbed using real-world filesystem metadata workloads. Fletch achieves up to 181.6% higher throughput than vanilla HDFS and complements client-side caching with throughput gains of up to 139.6% on 128 simulated metadata servers.

cs.AR↗

RAVEN: A Regime-Aware Variable-context Expert Network for Financial Time Series Forecasting

Financial time series forecasting presents structural challenges absent from standard benchmarks. Log-returns are non-stationary, exhibit exceptionally low signal-to-noise (SNR) ratios, and are governed by regime-dependent temporal dependencies. We identify a key limitation of state-of-the-art (SOTA) time series models in financial settings. A fixed context window is mismatched to the time-varying optimal look-back of non-stationary price processes. We propose the Regime-Aware Variable-context Expert Network (RAVEN), a Mixture-of-Experts framework designed to adaptively determine the temporal context for each input sample. Instead of relying on a fixed look-back horizon, RAVEN constructs a hierarchy of nested contiguous windows whose lengths are determined by the data itself. Specifically, RAVEN scores patches by learned importance in reverse chronological order and applies the Cumulative Importance Thresholding (CIT) mechanism to derive nested prefix windows, each routed to a scale-specialized expert. A Global Compressed Representation (GCR) branch runs in parallel over the full context, preserving global temporal coherence that local experts cannot guarantee. Because the nested routing induces structured overlap among expert inputs, we introduce a Correlation-Aware Weighting (CAW) to align variable-length expert outputs and penalize pairwise cosine similarity prior to aggregation. Experiments on cumulative log-return prediction (HS300, S&P500) and fund sales forecasting demonstrate that RAVEN achieves SOTA performances, improves Pearson correlation by 9.2% on HS300 and 20.2% on S&P500, and reduces MSE by 18.2% on fund sales forecasting, while achieving the best results in 14 of 16 metrics on four PEMS traffic benchmarks.

cs.LG↗

LUMEN: Coordinated Failure Recovery for Distributed LLM Serving

Modern large language model (LLM) serving clusters distribute inference requests across multiple worker processes on different GPUs, but failures are prevalent at scale. When a worker fails, the cluster simultaneously loses the failed worker's GPU-resident key-value (KV) caches and serving capacity, leaving surviving workers to absorb the redirected traffic while re-running interrupted requests from scratch. Existing fault-tolerant systems either restart interrupted requests from scratch or restore KV caches from checkpoints stored on a fixed neighboring worker, but both approaches route recovery work without considering current cluster load and leave the recovering worker idle during model reload. We present LUMEN, a fault-tolerant LLM serving system that treats recovery as a load-aware coordination problem across three decision points: checkpoint placement before failures, interrupted-request distribution at failure time, and serving capacity restoration during model reload. We evaluate LUMEN using both prototype experiments and large-scale simulations and demonstrate significant improvements in serving and recovery times.

cs.DC↗

Leveraging I/O Stalls for Efficient Scheduling in ANNS

Disk-based graph indexes for approximate nearest neighbor search (ANNS) must serve latency-sensitive queries and throughput-demanding updates concurrently. We observe that over 40% of search-thread CPU time is spent stalling on disk I/O; such idle cycles are invisible to thread-level scheduling yet available for other work. We present LIOS(Leverage I/O Stall), a framework that executes index updates inside search-side I/O stall windows. LIOS introduces three techniques: (i) splitting each update into resumable subtasks small enough to fit within a single stall window; (ii) bounding the expected overrun of update subtasks to a given threshold; and (iii) dynamically adjusting the fraction of idle time devoted to updates to drive end-to-end search latency degradation toward a user-specified target. We integrate LIOS into two update-optimized ANNS systems, FreshDiskANN and OdinANN. LIOS achieves speedups of up to 2.68$\times$ in insertion and 2.18$\times$ in deletion, with search latency degradation maintained near the user-specified target.

cs.DB↗

TierCheck: Tiered Checkpointing for Fault Tolerance in Large Language Model Training

Large Language Model (LLM) training is frequently interrupted by a heterogeneous spectrum of failures, from common GPU crashes to catastrophic cluster-wide outages. Existing checkpointing systems rely on monolithic, single-tier storage backend, forcing a trade-off between state-saving overhead and recovery speed. We propose TierCheck, a cluster-aware tiered checkpointing system that aligns storage placement with failure heterogeneity. TierCheck adopts a three-tier design that maintains lightweight differential checkpoints in local and peer memory for fast localized recovery, while asynchronously migrating heavyweight base checkpoints to remote persistent storage. It also ensures strict global consistency across tiers without stalling training, and achieves fast cluster-aware checkpoint restoration during recovery. Evaluations on models up to 40 billion parameters show that TierCheck achieves low training overhead, reduces end-to-end checkpointing time to under 10s, and supports high-frequency checkpointing, ultimately striking an optimal balance between low-overhead persistence and fast recovery.

cs.DC↗

Decoupling Vector Data and Index Storage for Space Efficiency

Managing large-scale vector datasets with disk-resident graph approximate nearest neighbor search (ANNS) systems incurs substantial storage overhead due to the co-location of vector data and auxiliary index metadata, which prevents the storage layer from exploiting their distinct compressibility. We present COMPASS, a component-aware compressed storage framework for disk-resident graph vector search. Leveraging data-index decoupling as a foundation, COMPASS losslessly compresses each component according to its distinct compressibility characteristics, thereby significantly reducing storage space. It further adapts the search and update paths to preserve their performance under compressed storage layouts. Evaluation on real-world public and proprietary billion-scale datasets shows that COMPASS reduces storage space by up to 58.7%, while delivering improved or competitive search and update performance compared to state-of-the-art disk-resident graph ANNS systems.

cs.DB↗

Joint Optimization of DNN Model Caching and Request Routing in Mobile Edge Computing

Mobile edge computing (MEC) can pre-cache deep neural networks (DNNs) near end-users, providing low-latency services and improving users' quality of experience (QoE). However, caching all DNN models at edge servers with limited capacity is difficult, and the impact of model loading time on QoE remains underexplored. Hence, we introduce dynamic DNNs in edge scenarios, disassembling a complete DNN model into interrelated submodels for more fine-grained and flexible model caching and request routing solutions. This raises the pressing issue of jointly deciding request routing and submodel caching for dynamic DNNs to balance model inference precision and loading latency for QoE optimization. In this paper, we study the joint dynamic model caching and request routing problem in MEC networks, aiming to maximize user request inference precision under constraints of server resources, latency, and model loading time. To tackle this problem, we propose CoCaR, an offline algorithm based on linear programming and random rounding that leverages dynamic DNNs to optimize caching and routing schemes, achieving near-optimal performance. Furthermore, we develop an online variant of CoCaR, named CoCaR-OL, enabling effective adaptation to dynamic and unpredictable online request patterns. The simulation results demonstrate that the proposed CoCaR improves the average inference precision of user requests by 46% compared to state-of-the-art baselines. In addition, in online scenarios, CoCaR-OL achieves an improvement of no less than 32.3% in user QoE over competitive baselines.

cs.NI↗

GTM: A General Time-series Model for Enhanced Representation Learning of Time-Series Data

Despite recent progress in time-series foundation models, challenges persist in improving representation learning and adapting to diverse downstream tasks. We introduce a General Time-series Model (GTM), which advances representation learning via a novel frequency-domain attention mechanism that captures time-granularity-aware features, an aspect underexplored in prior research. We further propose a novel pre-training strategy that unifies reconstruction and autoregressive objectives through a hybrid masking mechanism. Our pre-training strategy, combined with 2D positional encoding and span shuffling, enhances the robustness and generalization of representations. GTM is established as the first generative-task-agnostic model for time-series analysis, enabling seamless adaptation to various generative tasks without any task-specific modifications. Extensive experiments demonstrate that GTM consistently outperforms SOTA models on various generative tasks and achieves strong classification results with minimal adaptation. Furthermore, GTM exhibits clear scaling behavior, with accuracy improving as model size and pre-training data increase.

cs.LG↗

The Design and Implementation of a High-Performance Log-Structured RAID System for ZNS SSDs

Zoned Namespace (ZNS) defines a new abstraction for host software to flexibly manage storage in flash-based SSDs as append-only zones. It also provides a Zone Append primitive to further boost the write performance of ZNS SSDs by exploiting intra-zone parallelism. However, making Zone Append effective for reliable and scalable storage, in the form of a RAID array of multiple ZNS SSDs, is non-trivial, since Zone Append offloads address management to ZNS SSDs and requires hosts to specifically manage RAID stripes across multiple drives. We propose ZapRAID, a high-performance log-structured RAID system for ZNS SSDs by carefully exploiting Zone Append to achieve high write parallelism and lightweight stripe management. ZapRAID adopts a group-based data layout with a coarse-grained ordering across multiple groups of stripes, such that it can use small-size metadata for stripe management on a per-group basis under Zone Append. It further adopts hybrid data management to simultaneously achieve intra-zone and inter-zone parallelism through a careful combination of both Zone Write and Zone Append primitives. We implement ZapRAID as a user-space block device, and evaluate ZapRAID using microbenchmarks, trace-driven experiments, and real-application experiments. Our evaluation results show that ZapRAID achieves high write throughput and maintains high performance in normal reads, degraded reads, crash recovery, and full-drive recovery.

cs.DC↗

A Unified Frequency Domain Decomposition Framework for Interpretable and Robust Time Series Forecasting

Current approaches for time series forecasting, whether in the time or frequency domain, predominantly use deep learning models based on linear layers or transformers. They often encode time series data in a black-box manner and rely on trial-and-error optimization solely based on forecasting performance, leading to limited interpretability and theoretical understanding. Furthermore, the dynamics in data distribution over time and frequency domains pose a critical challenge to accurate forecasting. We propose FIRE, a unified frequency domain decomposition framework that provides a mathematical abstraction for diverse types of time series, so as to achieve interpretable and robust time series forecasting. FIRE introduces several key innovations: (i) independent modeling of amplitude and phase components, (ii) adaptive learning of weights of frequency basis components, (iii) a targeted loss function, and (iv) a novel training paradigm for sparse data. Extensive experiments demonstrate that FIRE consistently outperforms state-of-the-art models on long-term forecasting benchmarks, achieving superior predictive performance and significantly enhancing interpretability of time series

cs.LG↗

MVPipe: Enabling Lightweight Updates and Fast Convergence in Hierarchical Heavy Hitter Detection

Finding hierarchical heavy hitters (HHHs) (i.e., hierarchical aggregates with exceptionally huge amounts of traffic) is critical to network management, yet it is often challenged by the requirements of fast packet processing, real-time and accurate detection, as well as resource efficiency. Existing HHH detection schemes either incur expensive packet updates for multiple aggregation levels in the IP address hierarchy, or need to process sufficient packets to converge to the required detection accuracy. We present MVPipe, an invertible sketch that achieves both lightweight updates and fast convergence in HHH detection. MVPipe builds on the skewness property of IP traffic to process packets via a pipeline of majority voting executions, such that most packets can be updated for only one or few aggregation levels in the IP address hierarchy. We show how MVPipe can be feasibly deployed in P4-based programmable switches subject to limited switch resources. We also theoretically analyze the accuracy and coverage properties of MVPipe. Evaluation with real-world Internet traces shows that MVPipe achieves high accuracy, high throughput, and fast convergence compared to six state-of-the-art HHH detection schemes. It also incurs low resource overhead in the Tofino switch deployment.

cs.NI↗

An In-Depth Comparative Analysis of Cloud Block Storage Workloads: Findings and Implications

Cloud block storage systems support diverse types of applications in modern cloud services. Characterizing their I/O activities is critical for guiding better system designs and optimizations. In this paper, we present an in-depth comparative analysis of production cloud block storage workloads through the block-level I/O traces of billions of I/O requests collected from two production systems, Alibaba Cloud and Tencent Cloud Block Storage. We study their characteristics of load intensities, spatial patterns, and temporal patterns. We also compare the cloud block storage workloads with the notable public block-level I/O workloads from the enterprise data centers at Microsoft Research Cambridge, and identify the commonalities and differences of the three sources of traces. To this end, we provide 6 findings through the high-level analysis and 16 findings through the detailed analysis on load intensity, spatial patterns, and temporal patterns. We discuss the implications of our findings on load balancing, cache efficiency, and storage cluster management in cloud block storage systems.

cs.DC↗

A Generalization of Array Codes with Local Properties and Efficient Encoding/Decoding

A maximum distance separable (MDS) array code is composed of $m\times (k+r)$ arrays such that any $k$ out of $k+r$ columns suffice to retrieve all the information symbols. Expanded-Blaum-Roth (EBR) codes and Expanded-Independent-Parity (EIP) codes are two classes of MDS array codes that can repair any one symbol in a column by locally accessing some other symbols within the column, where the number of symbols $m$ in a column is a prime number. By generalizing the constructions of EBR and EIP codes, we propose new MDS array codes, such that any one symbol can be locally recovered and the number of symbols in a column can be not only a prime number but also a power of an odd prime number. Also, we present an efficient encoding/decoding method for the proposed generalized EBR (GEBR) and generalized EIP (GEIP) codes based on the LU factorization of a Vandermonde matrix. We show that the proposed decoding method has less computational complexity than existing methods. Furthermore, we show that the proposed GEBR codes have both a larger minimum symbol distance and a larger recovery ability of erased lines for some parameters when compared to EBR codes. We show that EBR codes can recover any $r$ erased lines of a slope for any parameter $r$, which was an open problem in [2].

cs.IT↗

Two New Piggybacking Designs with Lower Repair Bandwidth

Piggybacking codes are a special class of MDS array codes that can achieve small repair bandwidth with small sub-packetization by first creating some instances of an $(n,k)$ MDS code, such as a Reed-Solomon (RS) code, and then designing the piggyback function. In this paper, we propose a new piggybacking coding design which designs the piggyback function over some instances of both $(n,k)$ MDS code and $(n,k')$ MDS code, when $k\geq k'$. We show that our new piggybacking design can significantly reduce the repair bandwidth for single-node failures. When $k=k'$, we design piggybacking code that is MDS code and we show that the designed code has lower repair bandwidth for single-node failures than all existing piggybacking codes when the number of parity node $r=n-k\geq8$ and the sub-packetization $α<r$. Moreover, we propose another piggybacking codes by designing $n$ piggyback functions of some instances of $(n,k)$ MDS code and adding the $n$ piggyback functions into the $n$ newly created empty entries with no data symbols. We show that our code can significantly reduce repair bandwidth for single-node failures at a cost of slightly more storage overhead. In addition, we show that our code can recover any $r+1$ node failures for some parameters. We also show that our code has lower repair bandwidth than locally repairable codes (LRCs) under the same fault-tolerance and redundancy for some parameters.

cs.IT↗

Efficient LSM-Tree Key-Value Data Management on Hybrid SSD/HDD Zoned Storage

Zoned storage devices, such as zoned namespace (ZNS) solid-state drives (SSDs) and host-managed shingled magnetic recording (HM-SMR) hard-disk drives (HDDs), expose interfaces for host-level applications to support fine-grained, high-performance storage management. Combining ZNS SSDs and HM-SMR HDDs into a unified hybrid storage system is a natural direction to scale zoned storage at low cost, yet how to effectively incorporate zoned storage awareness into hybrid storage is a non-trivial issue. We make a case for key-value (KV) stores based on log-structured merge trees (LSM-trees) as host-level applications, and present HHZS, a middleware system that bridges an LSM-tree KV store with hybrid zoned storage devices based on hints. HHZS leverages hints issued by the flushing, compaction, and caching operations of the LSM-tree KV store to manage KV objects in placement, migration, and caching in hybrid ZNS SSD and HM-SMR HDD zoned storage. Experiments show that our HHZS prototype, when running on real ZNS SSD and HM-SMR HDD devices, achieves the highest throughput compared with all baselines under various settings.

cs.PF↗

Separating Data via Block Invalidation Time Inference for Write Amplification Reduction in Log-Structured Storage

Log-structured storage has been widely deployed in various domains of storage systems, yet its garbage collection incurs write amplification (WA) due to the rewrites of live data. We show that there exists an optimal data placement scheme that minimizes WA using the future knowledge of block invalidation time (BIT) of each written block, yet it is infeasible to realize in practice. We propose a novel data placement algorithm for reducing WA, SepBIT, that aims to infer the BITs of written blocks from storage workloads and separately place the blocks into groups with similar estimated BITs. We show via both mathematical and production trace analyses that SepBIT effectively infers the BITs by leveraging the write skewness property in practical storage workloads. Trace analysis and prototype experiments show that SepBIT reduces WA and improves I/O throughput, respectively, compared with state-of-the-art data placement schemes. SepBIT is currently deployed to support the log-structured block storage management at Alibaba Cloud.

cs.DC↗

Repair Pipelining for Erasure-Coded Storage: Algorithms and Evaluation

We propose repair pipelining, a technique that speeds up the repair performance in general erasure-coded storage. By carefully scheduling the repair of failed data in small-size units across storage nodes in a pipelined manner, repair pipelining reduces the single-block repair time to approximately the same as the normal read time for a single block in homogeneous environments. We further design different extensions of repair pipelining algorithms for heterogeneous environments and multi-block repair operations. We implement a repair pipelining prototype, called ECPipe, and integrate it as a middleware system into two versions of Hadoop Distributed File System (HDFS) (namely HDFS-RAID and HDFS-3) as well as Quantcast File System (QFS). Experiments on a local testbed and Amazon EC2 show that repair pipelining significantly improves the performance of degraded reads and full-node recovery over existing repair techniques.

cs.DC↗