Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ScaleLLM: A Resource-Frugal LLM Serving Framework by Optimizing End-to-End Efficiency

Yuhang Yao, Han Jin, Alay Dilipbhai Shah, Shanshan Han, Zijian Hu, Yide Ran, Dimitris Stripelis, Zhaozhuo Xu, Salman Avestimehr, Chaoyang He

TL;DR

ScaleLLM addresses the gap in end-to-end LLM serving by optimizing both the gateway/router and the inference engine to reduce total latency under real-world concurrency. The approach combines a Rust-based, gRPC-enabled gateway with advanced engine optimizations, including MOE/TP/EP parallelism, quantization, and continuous batching, along with a dynamic routing blueprint that adapts to workload levels. Experimental results show substantial improvements in throughput and latency over baselines, validating the end-to-end design across streaming and non-streaming generation scenarios. The work highlights the practical importance of end-to-end optimization in production LLM serving and provides a blueprint for dynamic, resource-frugal deployment in commercial settings.

Abstract

Large language models (LLMs) have surged in popularity and are extensively used in commercial applications, where the efficiency of model serving is crucial for the user experience. Most current research focuses on optimizing individual sub-procedures, e.g. local inference and communication, however, there is no comprehensive framework that provides a holistic system view for optimizing LLM serving in an end-to-end manner. In this work, we conduct a detailed analysis to identify major bottlenecks that impact end-to-end latency in LLM serving systems. Our analysis reveals that a comprehensive LLM serving endpoint must address a series of efficiency bottlenecks that extend beyond LLM inference. We then propose ScaleLLM, an optimized system for resource-efficient LLM serving. Our extensive experiments reveal that with 64 concurrent requests, ScaleLLM achieves a 4.3x speed up over vLLM and outperforms state-of-the-arts with 1.5x higher throughput.

ScaleLLM: A Resource-Frugal LLM Serving Framework by Optimizing End-to-End Efficiency

TL;DR

ScaleLLM addresses the gap in end-to-end LLM serving by optimizing both the gateway/router and the inference engine to reduce total latency under real-world concurrency. The approach combines a Rust-based, gRPC-enabled gateway with advanced engine optimizations, including MOE/TP/EP parallelism, quantization, and continuous batching, along with a dynamic routing blueprint that adapts to workload levels. Experimental results show substantial improvements in throughput and latency over baselines, validating the end-to-end design across streaming and non-streaming generation scenarios. The work highlights the practical importance of end-to-end optimization in production LLM serving and provides a blueprint for dynamic, resource-frugal deployment in commercial settings.

Abstract

Large language models (LLMs) have surged in popularity and are extensively used in commercial applications, where the efficiency of model serving is crucial for the user experience. Most current research focuses on optimizing individual sub-procedures, e.g. local inference and communication, however, there is no comprehensive framework that provides a holistic system view for optimizing LLM serving in an end-to-end manner. In this work, we conduct a detailed analysis to identify major bottlenecks that impact end-to-end latency in LLM serving systems. Our analysis reveals that a comprehensive LLM serving endpoint must address a series of efficiency bottlenecks that extend beyond LLM inference. We then propose ScaleLLM, an optimized system for resource-efficient LLM serving. Our extensive experiments reveal that with 64 concurrent requests, ScaleLLM achieves a 4.3x speed up over vLLM and outperforms state-of-the-arts with 1.5x higher throughput.
Paper Structure (13 sections, 10 figures, 1 table)

This paper contains 13 sections, 10 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Benchmark LLM Serving Solutions
  4. System Breakdown
  5. Performance of Baseline Solutions
  6. Optimizations
  7. Optimize Inference Engine
  8. Optimize Replica Router
  9. Experiments
  10. Evaluation Metrics
  11. Serving Performance Evaluation
  12. Blueprint Architecture of Dynamic Inference Load Balancing System
  13. Conclusion and Future work

Figures (10)

  • Figure 1: Overview of ScaleLLM Serving System. ScaleLLM provides an optimized gateway for balancing workloads of user requests to different inference replicas and an efficient serving engine for promptly response with high concurrent requests.
  • Figure 2: Comparisons with the two baseline solutions. ScaleLLM is applied without gateway optimization.
  • Figure 3: ScaleLLM Gateway Architecture
  • Figure 4: Lifecycle of Concurrent and Single Request
  • Figure 5: Endpoint Throughput Comparison.
  • ...and 5 more figures