With the rapid advancement of large language models (LLMs), reinforcement learning (RL) has emerged as a pivotal methodology for enhancing the reasoning capabilities of LLMs. Unlike traditional pre-training approaches, RL encompasses multiple stages: rollout, reward, and training, which necessitates collaboration among various worker types. However, current RL systems continue to grapple with substantial GPU underutilization, due to two primary factors: (1) The rollout stage dominates the overall RL process due to test-time scaling; (2) Imbalances in rollout lengths (within the same batch) result in GPU bubbles. While prior solutions like asynchronous execution and truncation offer partial relief, they may compromise training accuracy for efficiency. Our key insight stems from a previously overlooked observation: rollout responses exhibit remarkable similarity across adjacent training epochs. Based on the insight, we introduce RhymeRL, an LLM RL system designed to accelerate RL training with two key innovations. First, to enhance rollout generation, we present HistoSpec, a speculative decoding inference engine that utilizes the similarity of historical rollout token sequences to obtain accurate drafts. Second, to tackle rollout bubbles, we introduce HistoPipe, a two-tier scheduling strategy that leverages the similarity of historical rollout distributions to balance workload among rollout workers. Experimental results demonstrate that RhymeRL achieves up to a 2.6x performance improvement over existing methods, without compromising accuracy or modifying the RL paradigm.

Cloud-native systems increasingly rely on infrastructure services (e.g., service meshes, monitoring agents), which compete for resources with user applications, thereby degrading performance and scalability. We propose XpuPod, a new abstraction that offloads these services to Data Processing Units (DPUs) to enforce strict isolation while reducing host resource contention and operational costs. XpuPod enables a cross-processing-unit (XPU) network system. This system features two key components: (1) transparent XPU networking, which provides a unified network abstraction for processes spanning both the CPU and DPU, and (2) elastic and efficient XPU communication, a mechanism that achieves shared-memory performance without the costs of pinned resources. By leveraging the XPU network system and the compositional nature of cloud-native workloads, XpuPod can optimally offload infrastructure containers to DPUs. We implement XpuPod on Linux and a corresponding cloud-native system, XpuK8s, on Kubernetes. We evaluate our system using NVIDIA BlueField-2 DPUs and a simulator backed by a CXL memory device. The results show that XpuK8s effectively supports complex, unmodified, commodity cloud-native applications. Compared to a kernel-bypass design, XpuK8s provides up to 31.9x lower latency and consumes 64x fewer resources. Furthermore, compared to state-of-the-art systems, XpuK8s can achieve 60% lower end-to-end latency and 55% higher scalability.

As large language models (LLMs) are widely applied across various domains, improving the quality of LLM inference services is essential. In this paper, we find that optimizing configuration parameters of LLM inference engines can significantly improve LLM inference performance in terms of latency and throughput. Therefore, we propose LLMConf, an automated performance tuning system that optimizes multiple LLM inference performance metrics by searching for the optimal configuration parameters of the LLM inference engine. We first introduce a knowledge-enhanced approach to identify the set of configuration parameters (LLMConfigs) that most significantly impact LLM performance from the adjustable parameters provided by the LLM inference engine. We then perform automated data collection to build functional relationships between LLMConfigs and each performance metric. Additionally, LLMConf employs a multi-objective optimization module to obtain optimal LLMConfigs for simultaneously optimizing multiple performance metrics. The experimental results show that LLMConf significantly outperforms existing methods. Compared to the default configuration parameters of the LLM inference engine, LLMConf achieves an average improvement of 20.1% across 7 key performance metrics. Moreover, experiments demonstrate that LLMConf has strong transferability across diverse datasets, varying concurrency levels and different LLM base models.

In modern systems, memory copy remains a critical performance bottleneck across various scenarios, playing a pervasive role in system-wide execution such as syscalls, IPC, and user-mode applications. Numerous efforts have aimed at optimizing copy performance, including zero-copy with page remapping and hardware-accelerated copy. However, they typically target specific use cases, such as Linux zero-copy send() for messages of ≥10KB. This paper argues for copy as a first-class OS service, offering three key benefits: (1) with the asynchronous copy abstraction provided by the service, applications can overlap their execution with copy; (2) the service can effectively utilize hardware capabilities to enhance copy performance; (3) the service's global view of copies further enables holistic optimization. To this end, we introduce Copier, a new OS service of coordinated asynchronous copy, to serve both user-mode applications and OS services. We build Copier-Linux to demonstrate Copier's ability to improve performance for diverse use cases, including Redis, Protobuf, network stack, proxy, etc. Evaluations show that Copier achieves up to a 1.8 × speedup for real-world applications like Redis and a 1.6 × improvement over zIO, the state-of-the-art in optimizing copy efficiency. To further facilitate adoption, we develop a toolchain to ease the use of Copier. We also integrate Copier into a commercial smartphone OS (HarmonyOS 5.0), achieving promising results.