arXiv subjects

Minxian Xu

Publications and source records attributed to Minxian Xu.

At least 19 recordsLinked to original sources

arXiv2026

SSD-LLaMA: SSD-Native Inference for Trillion-Parameter MoE at 1+ Token/s on a Consumer PC

Frontier open-weight language models increasingly use Mixture-of-Experts (MoE) architectures to expand model capacity while activating only a small subset of experts per token. Local inference must nevertheless keep the complete expert pool available, which remains far beyond consumer-grade RAM and VRAM capacity even after quantization. SSDs provide practical capacity at this scale, but turning that capacity into executable model memory requires efficient expert delivery, coordinated management of SSD, RAM, and VRAM, and CPU--GPU hybrid execution under bounded bandwidth. We present \textit{SSD-LLaMA}, an SSD-native local MoE inference system that addresses these challenges with an SSD I/O pipeline optimized for expert delivery, a native three-tier storage hierarchy that delivers and retains experts dynamically, and balanced CPU--GPU hybrid execution. \textit{SSD-LLaMA} executes every selected expert without pruning or substitution. Across three frontier MoE model families, \textit{SSD-LLaMA} improves prefill token rate by 1.52$\times$--4.19$\times$ and decode token rate by 2.10$\times$--15.58$\times$ over the evaluated baselines. We also achieve higher than 1 token/s for running trillion-parameter model with a single RTX 5090 and no more than 32GB RAM.

cs.DC↗

arXiv2026

RolloutPipe: Overlapping Pipelined Rollout and Training in Disaggregated On-Policy LLM Reinforcement Learning

Large language model (LLM) post-training for reasoning increasingly relies on reinforcement learning with verifiable rewards (RLVR), where models learn from ground-truth feedback on mathematical, logical, and scientific tasks. To enable flexible resource allocation and support heterogeneous training setups, modern RLVR systems adopt disaggregated architectures that decouple rollout generation and policy training across independent GPU pools. However, existing synchronous on-policy GRPO (Group Relative Policy Optimization) RLVR systems finish an entire rollout before starting training, leaving the trainer GPU pool idle while rollout is still ongoing. Asynchronous RL pipelines overlap the two stages, but at the cost of training on stale data. To address these challenges, we propose RolloutPipe, a post-training framework for disaggregated RLVR systems, which turns the fixed-weight rollout into a complete-group pipeline where trainable groups move to the trainer while later groups are still being generated. RolloutPipe achieves this through two techniques including complete-group pipelining (CGP) and frontier-group dispatch (FGD). CGP dispatches each trainable complete group to the trainer FIFO as soon as group materialization finishes, and FGD is an admission policy on the Rollout node that first admits requests for the frontier groups needed to form the next training batch, so that trainer-ready groups arrive earlier and more steadily. The design starts training before the rollout completes while maintaining on-policy correctness. Evaluated on Qwen3-1.7B across four reasoning and science benchmarks and twelve rollout settings, RolloutPipe shortens the rollout-to-train-end time by 30.7%-42.3%, and lowers the trainer waiting ratio by 37%-76% compared to Slime, a state-of-the-art rollout and training system.

cs.DC↗

arXiv2026

BrownoutMoE: Structure-Aware Expert Grouping for Efficient and Accurate LLM Web-based Services

Mixture-of-Experts (MoE) large language models (LLMs) are increasingly deployed in Web-facing services, where inference must be both accurate and responsive under bursty demand. Although MoE models improve parameter efficiency through sparse expert activation, efficient MoE inference remains challenging in practice. A major reason is the highly imbalanced expert access pattern during inference: a few hot experts process most routed tokens, while many cold experts are rarely activated, leaving GPU parallelism underutilized. Existing systems mainly optimize runtime execution, such as scheduling, communication overlap, and kernel fusion, but usually preserve the original expert organization and therefore do not address the structural inefficiency caused by fragmented expert usage. In this paper, we present \textbf{BrownoutMoE}, a structure-aware optimization framework for efficient and accurate MoE inference services. Inspired by the brownout paradigm in service computing, BrownoutMoE reorganizes experts into groups to improve utilization and system efficiency while maintaining service quality. Specifically, we formulate layer-wise expert grouping as a learning problem and employ reinforcement learning to discover grouping strategies that minimize accuracy degradation. We further introduce a grouping-consistent distillation process to produce deployable models that are compatible with standard inference pipelines. Experimental results demonstrate that BrownoutMoE reduces accuracy degradation by up to 71.4% and improves throughput by up to 2.24x over baselines.

cs.DC↗

arXiv2026

CoCoScale: Leveraging Layer-wise Scaling to Unlock the Potential of Online LLM Serving

Online large language model (LLM) serving has become the backbone of modern AI applications, powering diverse downstream services through shared hardware clusters. However, modern serving systems frequently encounter highly dynamic workloads characterized by severe workload skewness, where a small fraction of model instances receives the vast majority of traffic. Existing instance-level scaling mechanisms are limited by coarse-grained resource adjustment: scaling up requires the cold-start of full-model replicas, incurring substantial latency, while scaling down leaves the system vulnerable to performance degradation during sudden traffic surges. The key insight of this work is that LLM serving offers a unique opportunity for fine-grained scaling. In this paper, we propose CoCoScale, a layer-wise dynamic scaling mechanism that selectively expands the parallelism of hot layers onto idle resources reclaimed from underutilized devices, enabling elastic data parallelism without altering model architectures or adding hardware overhead. Evaluations demonstrate that CoCoScale significantly reduces cold start latency by 97.9%-99.3% compared to traditional scale up. Under production traces, CoCoScale reduces average latency by 20.7\%--28.1\% and achieves full Service Level Objective (SLO) attainment, demonstrating superior dynamic adaptability and resource efficiency.

cs.DC↗

arXiv2026

SwiftCache: Efficient LLM Serving for Multi-turn Conversations with Heterogeneous KV Cache Sharing

Multi-turn conversation is a fundamental scenario in LLM applications, widely used in chatbots and AI agents. As the conversation evolves, historical tokens accumulate continuously. Existing systems cache their key-value (KV) pairs to avoid redundant computation. However, limited GPU memory (HBM) capacity often forces these KV caches to be offloaded to CPU memory or SSD, making KV cache reloads increasingly costly in terms of latency as the context grows. Meanwhile, the constrained HBM capacity also limits the maximum inference length, thereby restricting the number of turns that can be supported in a conversation. To address these two challenges, we propose SwiftCache, a collaborative inference system that enables heterogeneous models to share underutilized GPU memory and NVLink bandwidth within a server. Specifically, models with low KV cache demand donate idle GPU memory to store the prefix cache of high-demand models, allowing cross-model KV cache sharing over NVLink and avoiding slow PCIe transfers. SwiftCache further reduces memory pressure by keeping only the KV cache of the currently active layer in local GPU memory, thereby enabling longer-context inference. Our experiments on real-world workloads show that SwiftCache reduces P99 time-to-first-token (TTFT) by up to 69% and extends maximum context length by up to 3.98x compared to vLLM and SGLang, with minimal interference to co-located models.

cs.DC↗

arXiv2026

Coordinated Scheduling for MoE LLM Serving

Serving Mixture-of-Experts (MoE) large language models (LLMs) is challenging because dynamic request workloads interact with sparse expert routing, creating both data-parallel (DP) engine imbalance and expert-level hotspots. Existing LLM serving systems typically make these decisions in isolation: frontend schedulers route requests using coarse request counters, while backend expert balancers rely mainly on aggregate expert activation counts. This separation prevents the serving system from reacting to fine-grained engine pressure, backend MoE pressure, and source-dependent expert traffic. To address this gap, we propose Gimbal, a coordinated cross-level scheduling system for efficient MoE-based LLM serving. First, Gimbal presents a fine-grained DP-engine scheduler that uses online backend pressure signals, including key-value (KV) cache usage, remaining prefill work, queue pressure, and MoE expert pressure, to dispatch requests away from overloaded engines. Inside each engine, Gimbal further applies a lightweight prefill-aware queue ordering policy with aging to reduce head-of-line blocking without output-length prediction. Second, Gimbal extends expert load balancing with online source-DP-to-expert routing statistics and uses a heuristic guided by a mixed-integer nonlinear program (MINLP) to place experts while jointly considering expert load, source-aware communication, and migration stability. Our evaluation shows that Gimbal reduces average Time To First Token (TTFT) by 42.9% and average Time Per Output Token (TPOT) by 33.3% compared with the state-of-the-art serving system vLLM, while improving high-load request throughput by 3.0%.

cs.DC↗

arXiv2026

Cloud-native and Distributed Systems for Efficient and Scalable Large Language Models -- A Research Agenda

The rapid rise of Large Language Models (LLMs) has revolutionized various artificial intelligence (AI) applications, from natural language processing to code generation. However, the computational demands of these models, particularly in training and inference, present significant challenges. Traditional systems are often unable to meet these requirements, necessitating the integration of cloud-native and distributed architectures. This paper explores the role of cloud platforms and distributed systems in supporting the scalability, efficiency, and optimization of LLMs. We discuss the complexities of LLM deployment, including data management, resource optimization, and the need for microservices, autoscaling, and hybrid cloud-edge solutions. Additionally, we examine emerging research trends, such as serverless inference, quantum computing, and federated learning, and their potential to drive the next phase of LLM innovation. The paper concludes with a roadmap for future developments, emphasizing the need for continued research, standardization, and cross-sector collaboration to sustain the growth of LLMs in both research and enterprise applications.

cs.DC↗

arXiv2026

Auto-scaling Approaches for Microservice Applications: A Survey and Taxonomy

Microservice applications are created as loosely coupled application components and they leverage cloud elasticity to reduce costs and increase development speed. However, microservice applications exhibit complex interactions among dynamically evolving services and highly variable workloads, posing significant challenges to auto-scaling mechanisms. Key issues include service dependency management, performance profiling, anomaly detection, workload characterization, and fine-grained resource allocation. To address these challenges, recent auto-scaling approaches leverage historical and runtime data to adapt resource provisioning and optimize system efficiency. Since 2018, marked by the graduation of Kubernetes as the first Cloud Native Computing Foundation (CNCF) project, microservice applications have been widely deployed on standardized orchestration platforms, fundamentally shifting auto-scaling from coarse-grained to service-level, dependency-aware strategies. Accordingly, this paper surveys state-of-the-art auto-scaling approaches for microservice applications since 2018 and presents a taxonomy along five dimensions: infrastructure, architecture, scaling methods, optimization objectives, and behavior modeling. These perspectives collectively target key objectives, including resource efficiency, cost efficiency, and Service Level Agreement (SLA) assurance, aiming to balance system optimization with SLA compliance. We further present a comprehensive comparison and in-depth analysis of representative approaches, examining their core features, strengths, limitations, and applicable scenarios, as well as their performance across diverse environments and workload conditions.

cs.DC↗

arXiv2026

DOPD: A Dynamic PD-Disaggregation Architecture for Maximizing Goodput in LLM Inference Serving

To meet strict Service-Level Objectives (SLOs),contemporary Large Language Models (LLMs) decouple the prefill and decoding stages and place them on separate GPUs to mitigate the distinct bottlenecks inherent to each phase. However, the heterogeneity of LLM workloads causes producerconsumer imbalance between the two instance types in such disaggregated architecture. To address this problem, we propose DOPD (Dynamic Optimal Prefill/Decoding), a dynamic LLM inference system that adjusts instance allocations to achieve an optimal prefill-to-decoding (P/D) ratio based on real-time load monitoring. Combined with an appropriate request-scheduling policy, DOPD effectively resolves imbalances between prefill and decoding instances and mitigates resource allocation mismatches due to mixed-length requests under high concurrency. Experimental evaluations show that, compared with vLLM and DistServe (representative aggregation-based and disaggregationbased approaches), DOPD improves overall system goodput by up to 1.5X, decreases P90 time-to-first-token (TTFT) by up to 67.5%, and decreases P90 time-per-output-token (TPOT) by up to 22.8%. Furthermore, our dynamic P/D adjustment technique performs proactive reconfiguration based on historical load, achieving over 99% SLOs attainment while using less additional resources.

cs.DC↗

arXiv2026

Multi-Layer Scheduling for MoE-Based LLM Reasoning

Large Language Models (LLMs) have achieved remarkable success across a wide range of tasks, but serving them efficiently at scale remains a critical challenge due to their substantial computational and latency demands. While most existing inference frameworks rely on simple scheduling strategies such as First-Come-First-Serve (FCFS) at the engine level and Round-Robin (RR) at the scheduler or coordinator level, they often fail to fully utilize system resources and may suffer from issues such as head-of-line blocking and load imbalance. Recent advances in Mixture-of-Experts (MoE) models have also introduced new challenges in scheduling arising from expert parallelism and routing complexity. This research proposes a multi-layer scheduling framework tailored for MoE-based LLM serving. It targets scheduling at three levels: request-level, enginelevel, and expert-level. At the request level, we explore algorithms such as Shortest-Job-First (SJF) and priority-aware aging to improve throughput and reduce latency. At the engine level, we design load-aware dispatching strategies that account for the current prefix token load, KV cache utilization, and user stickiness to achieve better resource matching. At the expert level, we focus on alleviating expert hotspots and strategically placing inter-layer expert dependencies to balance load and improve routing efficiency. Extensive experimental results from more than 100 experiments conducted under diverse workload distributions show that our approach consistently outperforms the state-of-theart inference framework vLLM, achieving up to 17.8% reduction in Time To First Token (TTFT) latency and 13.3% reduction in Time-Per-Output-Token (TPOT) latency.

cs.DC↗

arXiv2026

ORACL: Optimized Reasoning for Autoscaling via Chain of Thought with LLMs for Microservices

Applications are moving away from monolithic designs to microservice and serverless architectures, where fleets of lightweight and independently deployable components run on public clouds. Autoscaling serves as the primary control mechanism for balancing resource utilization and quality of service, yet existing policies are either opaque learned models that require substantial per-deployment training or brittle hand-tuned rules that fail to generalize. We investigate whether large language models can act as universal few-shot resource allocators that adapt across rapidly evolving microservice deployments. We propose ORACL, Optimized Reasoning for Autoscaling via Chain of Thought with LLMs for Microservices, a framework that leverages prior knowledge and chain-of-thought reasoning to diagnose performance regressions and recommend resource allocations. ORACL transforms runtime telemetry, including pods, replicas, CPU and memory usage, latency, service-level objectives, and fault signals, into semantic natural-language state descriptions and invokes an LLM to produce an interpretable intermediate reasoning trace. This reasoning identifies likely root causes, prunes the action space, and issues safe allocation decisions under policy constraints. Experiments on representative open-source microservice workloads show that ORACL improves root-cause identification accuracy by 15 percent, accelerates training by up to 24x, and improves quality of service by 6 percent in short-term scenarios, without deployment-specific retraining.

cs.DC↗

arXiv2025

BanaServe: Unified KV Cache and Dynamic Module Migration for Balancing Disaggregated LLM Serving in AI Infrastructure

Large language models (LLMs) are increasingly deployed in AI infrastructure, driving the need for high throughput, resource efficient serving systems. Disaggregated LLM serving, which separates prompt prefill from auto-regressive decode, has emerged as a promising architecture by isolating their heterogeneous compute and memory demands. However, current disaggregated systems face three key limitations: (i) static resource allocation cannot adapt to highly dynamic workloads, causing over-provisioning that wastes resources or under-provisioning that violates service level objectives (SLOs); (ii) inherent load imbalance between prefill and decode stages, where prefill is compute-bound and decode is memory-bound, causes under-utilization in one tier while the other becomes a bottleneck; and (iii) prefix cache aware routing skews load distribution, as high cache hit rate prefill nodes attract disproportionately more requests, further degrading balance and efficiency. To address these issues, we present BanaServe, a dynamic orchestration framework that continuously rebalances computational and memory resources across prefill and decode instances while eliminating hotspots induced by cache. BanaServe introduces layer level weight migration, attention level Key Value Cache (KV Cache) migration, and Global KV Cache Store sharing with layer wise overlapped transmission, enabling both coarse grained (layer level) and fine grained (attention level) load redistribution with minimal latency overhead. These mechanisms allow routers to perform purely load aware scheduling, unconstrained by cache placement. Compared to vLLM, BanaServe achieves 1.2x-3.9x higher throughput with 3.9%-78.4% lower total processing time, and outperforms DistServe by 1.1x-2.8x in throughput with 1.4%-70.1% latency reduction.

cs.DC↗

arXiv2025

TD3-Sched: Learning to Orchestrate Container-based Cloud-Edge Resources via Distributed Reinforcement Learning

Resource scheduling in cloud-edge systems is challenging as edge nodes run latency-sensitive workloads under tight resource constraints, while existing centralized schedulers can suffer from performance bottlenecks and user experience degradation. To address the issues of distributed decisions in cloud-edge environments, we present TD3-Sched, a distributed reinforcement learning (DRL) scheduler based on Twin Delayed Deep Deterministic Policy Gradient (TD3) for continuous control of CPU and memory allocation, which can achieve optimized decisions for resource provisioning under dynamic workloads. On a realistic cloud-edge testbed with SockShop application and Alibaba traces, TD3-Sched achieves reductions of 17.9% to 38.6% in latency under same loads compared with other reinforcement-learning and rule-based baselines, and 16% to 31.6% under high loads. TD3-Sched also shows superior Service Level Objective (SLO) compliance with only 0.47% violations. These results indicate faster convergence, lower latency, and more stable performance while preserving service quality in container-based cloud-edge environment compared with the baselines.

cs.DC↗

arXiv2025

C-Koordinator: Interference-aware Management for Large-scale and Co-located Microservice Clusters

Microservices transform traditional monolithic applications into lightweight, loosely coupled application components and have been widely adopted in many enterprises. Cloud platform infrastructure providers enhance the resource utilization efficiency of microservices systems by co-locating different microservices. However, this approach also introduces resource competition and interference among microservices. Designing interference-aware strategies for large-scale, co-located microservice clusters is crucial for enhancing resource utilization and mitigating competition-induced interference. These challenges are further exacerbated by unreliable metrics, application diversity, and node heterogeneity. In this paper, we first analyze the characteristics of large-scale and co-located microservices clusters at Alibaba and further discuss why cycle per instruction (CPI) is adopted as a metric for interference measurement in large-scale production clusters, as well as how to achieve accurate prediction of CPI through multi-dimensional metrics. Based on CPI interference prediction and analysis, we also present the design of the C-Koordinator platform, an open-source solution utilized in Alibaba cluster, which incorporates co-location and interference mitigation strategies. The interference prediction models consistently achieve over 90.3% accuracy, enabling precise prediction and rapid mitigation of interference in operational environments. As a result, application latency is reduced and stabilized across all percentiles (P50, P90, P99) response time (RT), achieving improvements ranging from 16.7% to 36.1% under various system loads compared with state-of-the-art system. These results demonstrate the system's ability to maintain smooth application performance in co-located environments.

cs.DC↗

arXiv2025

Unlock the Potential of Fine-grained LLM Serving via Dynamic Module Scaling

The rise of large language models (LLMs) has created new opportunities across various fields but has also introduced significant challenges in resource management. Current LLM serving systems face a fundamental tension: balancing serving demands with limited resources while adapting to unpredictable traffic patterns. Static deployments lead to suboptimal resource utilization and performance degradation under dynamic workloads. Furthermore, the high cost of adjusting instances hinders dynamic scaling, limiting the true potential of efficient LLM serving. To address this, we propose CoCoServe, an elastic system that facilitates dynamic and fine-grained scaling. Its key innovation lies in the module-level operations for the replication and migration of LLM modules, such as decoder layers and projections. Through a comprehensive analysis of the trade-offs associated with these operations, we develop an auto-scaling mechanism that dynamically regulates module-level resource allocation and performance optimization, enabling a more cost-effective deployment of LLMs. Our evaluation demonstrates that the scaling operations employed by CoCoServe exhibit excellent scalability and can reduce costs by 46% while maintaining availability. Compared to state-of-the-art LLM serving systems (e.g., Hugging Face Transformers and vLLM), our approach reduces latency by 14%-75% and achieves 1.16x-4x throughput on average across different model sizes and workloads.

cs.DC↗

arXiv2025

Cloud Native System for LLM Inference Serving

Large Language Models (LLMs) are revolutionizing numerous industries, but their substantial computational demands create challenges for efficient deployment, particularly in cloud environments. Traditional approaches to inference serving often struggle with resource inefficiencies, leading to high operational costs, latency issues, and limited scalability. This article explores how Cloud Native technologies, such as containerization, microservices, and dynamic scheduling, can fundamentally improve LLM inference serving. By leveraging these technologies, we demonstrate how a Cloud Native system enables more efficient resource allocation, reduces latency, and enhances throughput in high-demand scenarios. Through real-world evaluations using Kubernetes-based autoscaling, we show that Cloud Native architectures can dynamically adapt to workload fluctuations, mitigating performance bottlenecks while optimizing LLM inference serving performance. This discussion provides a broader perspective on how Cloud Native frameworks could reshape the future of scalable LLM inference serving, offering key insights for researchers, practitioners, and industry leaders in cloud computing and artificial intelligence.

cs.DC↗

arXiv2025

BucketServe: Bucket-Based Dynamic Batching for Smart and Efficient LLM Inference Serving

Large language models (LLMs) have become increasingly popular in various areas, traditional business gradually shifting from rule-based systems to LLM-based solutions. However, the inference of LLMs is resource-intensive or latency-sensitive, posing significant challenges for serving systems. Existing LLM serving systems often use static or continuous batching strategies, which can lead to inefficient GPU memory utilization and increased latency, especially under heterogeneous workloads. These methods may also struggle to adapt to dynamic workload fluctuations, resulting in suboptimal throughput and potential service level objective (SLO) violations. In this paper, we introduce BucketServe, a bucket-based dynamic batching framework designed to optimize LLM inference performance. By grouping requests into size-homogeneous buckets based on sequence length, BucketServe minimizes padding overhead and optimizes GPU memory usage through real-time batch size adjustments preventing out-of-memory (OOM) errors. It introduces adaptive bucket splitting/merging and priority-aware scheduling to mitigate resource fragmentation and ensure SLO compliance. Experiment shows that BucketServe significantly outperforms UELLM in throughput, achieving up to 3.58x improvement. It can also handle 1.93x more request load under the SLO attainment of 80% compared with DistServe and demonstrates 1.975x higher system load capacity compared to the UELLM.

cs.DC↗

arXiv2025

BrownoutServe: SLO-Aware Inference Serving under Bursty Workloads for MoE-based LLMs

In recent years, the Mixture-of-Experts (MoE) architecture has been widely applied to large language models (LLMs), providing a promising solution that activates only a subset of the model's parameters during computation, thereby reducing overall memory requirements and allowing for faster inference compared to dense models. Despite these advantages, existing systems still face issues of low efficiency due to static model placement and lack of dynamic workloads adaptation. This leads to suboptimal resource utilization and increased latency, especially during bursty requests periods. To address these challenges, this paper introduces BrownoutServe, a novel serving framework designed to optimize inference efficiency and maintain service reliability for MoE-based LLMs under dynamic computational demands and traffic conditions. BrownoutServe introduces "united experts" that integrate knowledge from multiple experts, reducing the times of expert access and inference latency. Additionally, it proposes a dynamic brownout mechanism to adaptively adjust the processing of certain tokens, optimizing inference performance while guaranteeing service level objectives (SLOs) are met. Our evaluations show the effectiveness of BrownoutServe under various workloads: it achieves up to 2.07x throughput improvement compared to vLLM and reduces SLO violations by 90.28%, showcasing its robustness under bursty traffic while maintaining acceptable inference accuracy.

cs.DC↗