arXiv subjects

Gangmuk Lim

Publications and source records attributed to Gangmuk Lim.

4 recordsLinked to original sources

arXiv2026

Lodestar: An Online-Learning LLM Inference Router

Efficiently serving large language model (LLM) inference tasks is crucial both for user-perceived latency such as time-to-first-token (TTFT) and for GPU utilization. However, LLM request routing, that is, assigning each inference request to a GPU instance, is particularly challenging: execution is highly input-dependent; batching and KV-cache reuse create strong cross-request coupling; and latency responds nonlinearly to context length, model/engine settings, and heterogeneous accelerators. As a result, simple traditional load balancing algorithms, and even heuristics tailored for LLM inference, fail to achieve good performance. We present Lodestar, a novel learning-based request routing system for distributed GPU clusters. Lodestar continuously collects a snapshot of the cluster at per-request level, including real-time instance state, request characteristics, and observed performance, and trains an online reward predictor that it uses to route inference requests to the instance that will maximize given reward (e.g., minimizing TTFT). Lodestar is cloud-native and works seamlessly with existing serving stacks (vLLM). With continuous online adaptation to changing workloads and infrastructure conditions, Lodestar achieves 1.41x lower average TTFT and 1.47x lower P99 TTFT on average (up to 2.15x/1.86x on homogeneous and 4.38x/4.42x on heterogeneous clusters) compared to a state-of-the-art prefix cache and load-aware heuristic, and learns these efficient routing strategies within about 5 minutes, based on experiments in a public cloud GPU cluster.

cs.DC↗

arXiv2025

AIBrix: Towards Scalable, Cost-Effective Large Language Model Inference Infrastructure

We introduce AIBrix, a cloud-native, open-source framework designed to optimize and simplify large-scale LLM deployment in cloud environments. Unlike traditional cloud-native stacks, AIBrix follows a co-design philosophy, ensuring every layer of the infrastructure is purpose-built for seamless integration with inference engines like vLLM. AIBrix introduces several key innovations to reduce inference costs and enhance performance including high-density LoRA management for dynamic adapter scheduling, LLM-specific autoscalers, and prefix-aware, load-aware routing. To further improve efficiency, AIBrix incorporates a distributed KV cache, boosting token reuse across nodes, leading to a 50% increase in throughput and a 70% reduction in inference latency. AIBrix also supports unified AI runtime which streamlines model management while maintaining vendor-agnostic engine compatibility. For large-scale multi-node inference, AIBrix employs hybrid orchestration -- leveraging Kubernetes for coarse-grained scheduling and Ray for fine-grained execution -- to balance efficiency and flexibility. Additionally, an SLO-driven GPU optimizer dynamically adjusts resource allocations, optimizing heterogeneous serving to maximize cost efficiency while maintaining service guarantees. Finally, AIBrix enhances system reliability with AI accelerator diagnostic tools, enabling automated failure detection and mock-up testing to improve fault resilience. AIBrix is available at https://github.com/vllm-project/aibrix.

cs.DC↗

arXiv2023

Kivi: Verification for Cluster Management

Modern cloud infrastructure is powered by cluster management systems such as Kubernetes and Docker Swarm. While these systems seek to minimize users' operational burden, the complex, dynamic, and non-deterministic nature of these systems makes them hard to reason about, potentially leading to failures ranging from performance degradation to outages. We present Kivi, the first system for verifying controllers and their configurations in cluster management systems. Kivi focuses on the popular system Kubernetes, and models its controllers and events into processes whereby their interleavings are exhaustively checked via model checking. Central to handling autoscaling and large-scale deployments is our design that seeks to find violations in a smaller and reduced topology. We also develop several model optimizations in Kivi to scale to large clusters. We show that Kivi is effective and accurate in finding issues in realistic and complex scenarios and showcase two new issues in Kubernetes controller source code.

cs.DC↗

arXiv2019

Approximate Quantiles for Datacenter Telemetry Monitoring

Datacenter systems require efficient troubleshooting and effective resource scheduling so as to minimize downtimes and to efficiently utilize limited resources. In doing so, datacenter operators employ streaming analytics for collecting and processing datacenter telemetry over a temporal window. The quantile operator is key to these systems as it can summarize the typical and abnormal behavior of the monitored system. Computing quantiles in real-time is resource-intensive as it requires processing hundreds of millions of events in seconds while providing high accuracy. We overcome these challenges in real-time quantile computation through workload-driven approximation, motivated by three insights in our study: (i) values are dominated by a set of recurring small values, (ii) distribution of small values is consistent across different time scales, and (iii) tail values are dominated by a small set of large values. That is, we propose AOMG, an efficient and accurate quantile approximation algorithm that capitalizes on these insights. AOMG minimizes memory footprint of the quantile operator via compression and frequency-based summarization of small values. While these summaries are stored and processed at sub-window granularity for memory efficiency, they can extend to compute quantiles on user-defined temporal windows. Low value error for tail quantiles is achieved by retaining a few tail values per subwindow. AOMG estimates quantiles with high throughput and less than 5% relative value error across a wide range of use cases while state-of-the-art algorithms either have a high relative value error (9.3-137.0%) or deliver lower throughput (15-92%).

cs.DB↗