arXiv subjects

Xingda Wei

Publications and source records attributed to Xingda Wei.

18 recordsLinked to original sources

arXiv2026

SMetric: Rethink LLM Scheduling for Serving Agents with Balanced Session-centric Scheduling

LLM scheduling is critical to serving, yet it remains unclear how well existing designs fit agentic serving--with LLM requests issued by agents instead of humans. This shifts the workload in two ways: (1) agents act only on complete responses, making the cluster's tokens per second (TPS) the primary goal and relaxing--not eliminating--per-token latency requirements; and (2) requests share much of their KV\$-reuse exceeds 80% of request tokens in a production trace from BAILIAN, versus 54-62% in chat. This paper first contributes a systematic study of request scheduling for agents on two real-world traces. We find that to increase KV\$ reuse, existing schedulers overly prioritize routing requests to instances caching their KV\$, overloading a few while leaving the rest idle, capping TPS. We thus present two key insights: (1) load balance need not sacrifice all KV\$ reuse, thanks to the global-tier KV\$ store and (2) by utilizing the workload's intra-session locality, balancing a small fraction of requests--the first request in each agent session--suffices to balance the cluster without sacrificing most KV\$ reuse on local instances. SMETRIC realizes these insights with balanced session-centric scheduling: it routes each session's first request purely for load balance and its follow-up requests in a cache-aware manner, preserving load balance and local reuse while keeping demand on the global tier low. Using the session turn information as the scheduling metric is deliberate: it is derived efficiently and accurately from the user inputs alone, so the scheduler stays clean and stateless. SMETRIC improves cluster TPS by 10-16% under prefill-decode colocation with a global store and prefill TPS by 2-34% under disaggregation over state-of-the-art schedulers, also with a better per-token latency.

cs.DC↗

arXiv2026

CoAgent: Concurrency Control for Multi-Agent Systems

Multi-agent LLM systems -- coding agents, devops agents, document agents -- now routinely run several agents in parallel against the same git tree, Kubernetes cluster, or document. As soon as two of them mutate shared state, they enter the regime classical concurrency control has studied for decades, but classical mechanisms fit LLM agents poorly. A single agent transaction spans minutes of inference, read sets are broad and opaque rather than statically inferable, and the live state agents act on admits neither fork nor buffer, so writes take effect the moment they execute. Locks block long inference intervals; OCC abort-and-retry discards minutes of work on every conflict. This paper builds concurrency control on a capability classical transactions lack: the LLM inside each agent can judge whether a conflicting write invalidates its plan, and can repair exactly the operations that depended on it. Control therefore turns advisory: the runtime informs, the agent repairs. Our protocol, MTPO (Monotonic Trajectory Pre-Order), fixes a serialization order at launch, serves each read the order-filtered value, and applies writes speculatively in place; a one-way notification asks an affected reader to re-judge and patch its plan, while the framework mechanically undoes and reorders misplaced writes through the saga-style inverse each tool registers in advance. At quiescence the run is serializable in the pre-decided order. We realize MTPO as CoAgent, toolcall middleware whose privileged ToolSmith grows footprint-declared, undoable tools online. On ten contended workloads, CoAgent stays within 5\% of serial correctness at a $1.4\times$ speedup and near-serial token cost, where 2PL and OCC surrender nearly all concurrency gains; on a bash-only target system, it grows a 25-tool library online and lifts the task pass rate from 45/71 to 63/71 at $0.80\times$ the time and $0.86\times$ the cost.

cs.DC↗

arXiv2026

Simple is Better: Multiplication May Be All You Need for LLM Request Scheduling

High-quality LLM request scheduling requires meeting two key objectives: ensuring the routed instance has KVCache to accelerate request execution, and ensuring that the workload is balanced across instances. Achieving both objectives is challenging because pursuing one may compromise the other. Current approaches use various combinators (e.g., linear combinations) to compute a scheduling score that combines indicators for the two objectives. These approaches are complex: they either require significant workload-specific hyperparameter tuning or model-hardware-aware simulator development, yet could still lead to suboptimal performance. In this paper, we show that using a simple multiplication of two carefully chosen indicators: one KVCache-aware (new prefill tokens if routed to an instance) and one load-balancing-aware (current batch size of the instance), as the scheduling score (LMETRIC) can achieve both objectives simultaneously without any hyperparameter tuning. The key idea is that the simply multiplied score considers both objectives in a manner similar to a linear combination, but the original hyperparameters cancel out during comparison, so no tuning is needed to find the best parameters. The two indicators are chosen based on our analysis of LLM characteristics. Our extensive experiments show that this simple approach can reduce TTFT by 92% and 39%, and TPOT by 24% and 51%, compared to vLLM-v1 and an in-production scheduler on real-world workloads covering chatbots and coding agents. We also derive the mathematical conditions under which multiplication may fail, and find that such conditions are extremely rare in practice and can be detected (and mitigated) beforehand. LMETRIC has been deployed in production and canary release confirms its effectiveness

cs.DC↗

arXiv2026

Efficient Vector Search in the Wild: One Model for Multi-K Queries

Learned top-K search is a promising approach for serving vector queries with both high accuracy and performance. However, current models trained for a specific K value fail to generalize to real-world multi-K queries: they suffer from accuracy degradation (for larger Ks) and performance loss (for smaller Ks). Training the model to generalize on different Ks requires orders of magnitude more preprocessing time and is not suitable for serving vector queries in the wild. We present OMEGA, a K-generalizable learned top-K search method that simultaneously achieves high accuracy, high performance, and low preprocessing cost for multi-K vector queries. The key idea is that a base model properly trained on K=1 with our trajectory-based features can be used to accurately predict larger Ks with a dynamic refinement procedure and smaller Ks with minimal performance loss. To make our refinements efficient, we further leverage the statistical properties of top-K searches to reduce excessive model invocations. Extensive evaluations on multiple public and production datasets show that, under the same preprocessing budgets, OMEGA achieves 6-33% lower average latency compared to state-of-the-art learned search methods, while all systems achieve the same recall target. With only 16-30% of the preprocessing time, OMEGA attains 1.01-1.28x of the optimal average latency of these baselines.

cs.DB↗

arXiv2025

Fast LLM Post-training via Decoupled and Fastest-of-N Speculation

Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. This work, SpecActor, achieves fast rollout with speculative decoding that deploys a fast draft path to accelerate the unparallelizable generation, while the correctness is guaranteed by fast parallel verification of the outputs with the original model. SpecActor addresses two foundational challenges that hinder speculation efficiency: (1) a Decoupled speculation method that overcomes the computation inefficiency issue when executing speculative decoding with relative large per-worker batch size -- a common configuration in training but unfriendly to speculation, and (2) a Fastest-of-N speculation method that selects and combines different draft methods according to the rollout progress to approximate the optimal draft method even when the best one is unknown a priori. Extensive evaluations on production traces show that SpecActor accelerates mean rollout speed by 2.0--2.4x, with up to 2.7x speedup, over common post-training baselines. The results are consistent across both dense and MoE models and across different RL algorithms. Notably, SpecActor is 1.1--2.6x faster compared to vanilla speculative rollout in different traces. The accelerated rollout achieves 1.4--2.3x faster end-to-end training time.

cs.DC↗

arXiv2025

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

Serving large language models (LLMs) is important for cloud providers, and caching intermediate results (KV\$) after processing each request substantially improves serving throughput and latency. However, there is limited understanding of how LLM serving benefits from KV\$ caching, where system design decisions like cache eviction policies are highly workload-dependent. In this paper, we present the first systematic characterization of the KV\$ workload patterns from one of the leading LLM service providers. We draw observations that were not covered by previous studies focusing on synthetic workloads, including: KV\$ reuses are skewed across requests, where reuses between single-turn requests are equally important as multi-turn requests; the reuse time and probability are diverse considering all requests, but for a specific request category, the pattern tends to be predictable; and the overall cache size required for an ideal cache hit ratio is moderate. Based on the characterization, we further propose a workload-aware cache eviction policy that improves the serving performance under real-world traces, especially with limited cache capacity.

cs.DC↗

arXiv2025

DecLock: A Case of Decoupled Locking for Disaggregated Memory

This paper reveals that locking can significantly degrade the performance of applications on disaggregated memory (DM), sometimes by several orders of magnitude, due to contention on the NICs of memory nodes (MN-NICs). To address this issue, we present DecLock, a locking mechanism for DM that employs decentralized coordination for ownership transfer across compute nodes (CNs) while retaining centralized state maintenance on memory nodes (MNs). DecLock features cooperative queue-notify locking that queues lock waiters on MNs atomically, enabling clients to transfer lock ownership via message-based notifications between CNs. This approach conserves MN-NIC resources for DM applications and ensures fairness. Evaluations show DecLock achieves throughput improvements of up to 43.37$\times$ and 1.81$\times$ over state-of-the-art RDMA-based spinlocks and MCS locks, respectively. Furthermore, DecLock helps two DM applications, including an object store and a real-world database index (Sherman), avoid performance degradation under high contention, improving throughput by up to 35.60$\times$ and 2.31$\times$ and reducing 99th-percentile latency by up to 98.8% and 82.1%.

cs.DC↗

arXiv2025

DiFache: Efficient and Scalable Caching on Disaggregated Memory using Decentralized Coherence

The disaggregated memory (DM) architecture offers high resource elasticity at the cost of data access performance. While caching frequently accessed data in compute nodes (CNs) reduces access overhead, it requires costly centralized maintenance of cache coherence across CNs. This paper presents DiFache, an efficient, scalable, and coherent CN-side caching framework for DM applications. Observing that DM applications already serialize conflicting remote data access internally rather than relying on the cache layer, DiFache introduces decentralized coherence that aligns its consistency model with memory nodes instead of CPU caches, thereby eliminating the need for centralized management. DiFache features a decentralized invalidation mechanism to independently invalidate caches on remote CNs and a fine-grained adaptive scheme to cache objects with varying read-write ratios. Evaluations using 54 real-world traces from Twitter show that DiFache outperforms existing approaches by up to 10.83$\times$ (5.53$\times$ on average). By integrating DiFache, the peak throughput of two real-world DM applications increases by 7.94$\times$ and 2.19$\times$, respectively.

cs.DC↗

arXiv2025

Enabling Efficient Transaction Processing on CXL-Based Memory Sharing

Transaction processing systems are the crux for modern data-center applications, yet current multi-node systems are slow due to network overheads. This paper advocates for Compute Express Link (CXL) as a network alternative, which enables low-latency and cache-coherent shared memory accesses. However, directly adopting standard CXL primitives leads to performance degradation due to the high cost of maintaining cross-node cache coherence. To address the CXL challenges, this paper introduces CtXnL, a software-hardware co-designed system that implements a novel hybrid coherence primitive tailored to the loosely coherent nature of transactional data. The core innovation of CtXnL is empowering transaction system developers with the ability to selectively achieve data coherence. Our evaluations on OLTP workloads demonstrate that CtXnL enhances performance, outperforming current network-based systems and achieves with up to 2.08x greater throughput than vanilla CXL memory sharing architectures across universal transaction processing policies.

cs.AR↗

arXiv2024

KunServe: Parameter-centric Memory Management for Efficient Memory Overloading Handling in LLM Serving

Serving LLMs with a cluster of GPUs is common nowadays, where the serving system must meet strict latency SLOs required by applications. However, the stateful nature of LLM serving requires maintaining huge states (i.e., KVCache) in limited GPU memory. Under spikes in real-world workloads, GPU memory can be easily throttled, leading to orders of magnitude higher response latency due to queuing introduced by waiting for KVCache to be reclaimed. Prior KVCache-centric approaches handle load throttling by dropping, migrating, or swapping KVCache. These methods fail to release sufficient memory quickly with requests still queued. This paper proposes the first parameter-centric approach to handling throttling by selectively dropping replicated parameters to instantly free memory for requests, based on an unnoticed observation that model parameters are commonly replicated across GPUs for serving LLMs. With additional memory, all requests can be served with a larger batch without queuing. To make the parameter-centric approach correct and efficient, we cooperatively execute requests on GPUs with a complete copy of parameters using pipeline parallelism, and derive an appropriate drop plan without unnecessary cooperation. We also design techniques to minimize the performance overhead due to pipeline parallelism with the execution patterns of requests under drop. Evaluations show that {\sys} reduces the tail TTFT of requests under throttling by up to 72.2 times compared to the state-of-the-art systems including Llumnix, vLLM and InferCept.

cs.DC↗

arXiv2024

BLITZSCALE: Fast and Live Large Model Autoscaling with O(1) Host Caching

Model autoscaling is the key mechanism to achieve serverless model-as-a-service, but it faces a fundamental trade-off between scaling speed and storage/memory usage to cache parameters, and cannot meet frequent scaling requirements across multiple hosts. The key problem is that data plane performance is slow, and scaled instances remain stopped while parameters are loading. In this paper, we first show that the data plane can be made fast with no or O(1) caching by loading parameters through the compute network between GPUs because: (1) its speed is comparable to host cache and is underutilized, and (2) scaling multiple instances requires no or O(1) caching with network-optimized multicast. Second, autoscaling can be made live by breaking the scaling abstraction for inference from a coarse-grained instance-level to a fine-grained layer-level. This allows us to offload the layer computation from the overloaded serving instances to the scaled ones without waiting for the parameters to be fully loaded. Under real-world workloads, our system BLITZSCALE achieves up to 94 % lower tail latency reductions compared to state-of-the-art autoscaling system (ServerlessLLM), and it reduces the GPU time used for serving by 49 % when compared with serving systems that do not support autoscaling like DistServe and vLLM with the same service-level-agreement.

cs.DC↗

arXiv2024

PhoenixOS: Concurrent OS-level GPU Checkpoint and Restore with Validated Speculation

PHOENIXOS (PHOS) is the first OS service that can concurrently checkpoint and restore (C/R) GPU processes--a fundamental capability for critical tasks such as fault tolerance, process migration, and fast startup. While concurrent C/R is well-established on CPUs, it poses unique challenges on GPUs due to their lack of essential features for efficiently tracing concurrent memory reads and writes, such as specific hardware capabilities (e.g., dirty bits) and OS-mediated data paths (e.g., copy-on-write). To ensure correct concurrent C/R, PHOS proactively detects GPU memory reads and writes through a two-step process: first, it speculates about GPU memory accesses based on the arguments used when launching GPU kernels; then, it validates these accesses efficiently at runtime using binary instrumentation. With this validated speculation, PHOS retrofits CPU-based concurrent C/R for GPUs through software-based approaches, including soft copy-on-write, soft recopy, and soft on-demand restore. PHOS further proposes several GPU-aware techniques for efficient GPU C/R, including coordinated checkpoint data transfer and execution context pool. For downstream tasks that use C/R for tolerating failures, migrating processes between machines, and accelerating cold starts in serverless computing, PHOS achieves orders of magnitude higher performance than state-of-the-art OS-level GPU C/R systems like NVIDIA cuda-checkpoint.

cs.DC↗

arXiv2024

Characterizing the Dilemma of Performance and Index Size in Billion-Scale Vector Search and Breaking It with Second-Tier Memory

Vector searches on large-scale datasets are critical to modern online services like web search and RAG, which necessity storing the datasets and their index on the secondary storage like SSD. In this paper, we are the first to characterize the trade-off of performance and index size in existing SSD-based graph and cluster indexes: to improve throughput by 5.7$\times$ and 1.7$\times$, these indexes have to pay a 5.8$\times$ storage amplification and 7.7$\times$ with respect to the dataset size, respectively. The root cause is that the coarse-grained access of SSD mismatches the fine-grained random read required by vector indexes with small amplification. This paper argues that second-tier memory, such as remote DRAM/NVM connected via RDMA or CXL, is a powerful storage for addressing the problem from a system's perspective, thanks to its fine-grained access granularity. However, putting existing indexes -- primarily designed for SSD -- directly on second-tier memory cannot fully utilize its power. Meanwhile, second-tier memory still behaves more like storage, so using it as DRAM is also inefficient. To this end, we build a graph and cluster index that centers around the performance features of second-tier memory. With careful execution engine and index layout designs, we show that vector indexes can achieve optimal performance with orders of magnitude smaller index amplification, on a variety of second-tier memory devices. Based on our improved graph and vector indexes on second-tier memory, we further conduct a systematic study between them to facilitate developers choosing the right index for their workloads. Interestingly, the findings on the second-tier memory contradict the ones on SSDs.

cs.DC↗

arXiv2024

Characterizing Network Requirements for GPU API Remoting in AI Applications

GPU remoting is a promising technique for supporting AI applications. Networking plays a key role in enabling remoting. However, for efficient remoting, the network requirements in terms of latency and bandwidth are unknown. In this paper, we take a GPU-centric approach to derive the minimum latency and bandwidth requirements for GPU remoting, while ensuring no (or little) performance degradation for AI applications. Our study including theoretical model demonstrates that, with careful remoting design, unmodified AI applications can run on the remoting setup using commodity networking hardware without any overhead or even with better performance, with low network demands.

cs.OS↗

arXiv2023

Transactional Indexes on (RDMA or CXL-based) Disaggregated Memory with Repairable Transaction

The failure atomic and isolated execution of clients operations is a default requirement for a system that serve multiple loosely coupled clients at a server. However, disaggregated memory breaks this requirement in remote indexes because a client operation is disaggregated to multiple remote reads/writes. Current indexes focus on performance improvements and largely ignore tolerating client failures. We argue that a practical DM index should be transactional: each index operation should be failure atomic and isolated in addition to being concurrency isolated. We present repairable transaction (rTX), a lightweight primitive to execute DM index operations. Each rTX can detect other failed rTXes on-the-fly with the help of concurrency control. Upon detection, it will repair their non-atomic updates online with the help of logging, thus hiding their failures from healthy clients. By further removing unnecessary logging and delegating concurrency control to existing carefully-tuned index algorithms, we show that transactional indexes can be built at a low performance overhead on disaggregated memory. We have refactored two state-of-the-art DM indexes, RaceHashing and Sherman (B+Tree), with rTX. Evaluations show that rTX is 1.2 to 2X faster than other alternatives, e.g., distributed transaction. Meanwhile, its overhead is up to 42% compared to non-fault-tolerant indexes.

cs.DB↗

arXiv2022

Characterizing Off-path SmartNIC for Accelerating Distributed Systems

SmartNICs have recently emerged as an appealing device for accelerating distributed systems. However, there has not been a comprehensive characterization of SmartNICs, and existing designs typically only leverage a single communication path for workload offloading. This paper presents the first holistic study of a representative off-path SmartNIC, specifically the Bluefield-2, from a communication-path perspective. Our experimental study systematically explores the key performance characteristics of communication among the client, on-board SoC, and host, and offers insightful findings and advice for designers. Moreover, we propose the concurrent use of multiple communication paths of a SmartNIC and present a pioneering guideline to expose new optimization opportunities for various distributed systems. To demonstrate the effectiveness of our approach, we conducted case studies on a SmartNIC-based distributed file system (LineFS) and an RDMA-based disaggregated key-value store (DrTM-KV). Our experimental results show improvements of up to 30% and 25% for LineFS and DrTM-KV, respectively.

cs.DC↗

arXiv2022

No Provisioned Concurrency: Fast RDMA-codesigned Remote Fork for Serverless Computing

Serverless platforms essentially face a tradeoff between container startup time and provisioned concurrency (i.e., cached instances), which is further exaggerated by the frequent need for remote container initialization. This paper presents MITOSIS, an operating system primitive that provides fast remote fork, which exploits a deep codesign of the OS kernel with RDMA. By leveraging the fast remote read capability of RDMA and partial state transfer across serverless containers, MITOSIS bridges the performance gap between local and remote container initialization. MITOSIS is the first to fork over 10,000 new containers from one instance across multiple machines within a second, while allowing the new containers to efficiently transfer the pre-materialized states of the forked one. We have implemented MITOSIS on Linux and integrated it with FN, a popular serverless platform. Under load spikes in real-world serverless workloads, MITOSIS reduces the function tail latency by 89% with orders of magnitude lower memory usage. For serverless workflow that requires state transfer, MITOSIS improves its execution time by 86%.

cs.OS↗

arXiv2021

KRCORE: a microsecond-scale RDMA control plane for elastic computing

We present KRCORE, an RDMA library with a microsecond-scale control plane on commodity RDMA hardware for elastic computing. KRCORE can establish a full-fledged RDMA connection within 10{\mu}s (hundreds or thousands of times faster than verbs), while only maintaining a (small) fixed-sized connection metadata at each node, regardless of the cluster scale. The key ideas include virtualizing pre-initialized kernel-space RDMA connections instead of creating one from scratch, and retrofitting advanced RDMA dynamic connected transport with static transport for both low connection overhead and high networking speed. Under load spikes, KRCORE can shorten the worker bootstrap time of an existing disaggregated key-value store (namely RACE Hashing) by 83%. In serverless computing (namely Fn), KRCORE can also reduce the latency for transferring data through RDMA by 99%.

cs.NI↗