Ennan Zhai

Yuxing Xiang, Xue Li, Kun Qian et al.

With the widespread adoption of Large Language Models (LLMs), serving LLM inference requests has become an increasingly important task, attracting active research advancements. Practical workloads play an essential role in this process: they are critical for motivating and benchmarking serving techniques and systems. However, the existing understanding of real-world LLM serving workloads is limited due to the lack of a comprehensive workload characterization. Prior analyses remain insufficient in scale and scope, thus failing to fully capture intricate workload characteristics. In this paper, we fill the gap with an in-depth characterization of LLM serving workloads collected from our worldwide cloud inference serving service, covering not only language models but also emerging multimodal and reasoning models, and unveiling important new findings in each case. Moreover, based on our findings, we propose ServeGen, a principled framework for generating realistic LLM serving workloads by composing them on a per-client basis. A practical use case in production validates that ServeGen avoids 50% under-provisioning compared to naive workload generation, demonstrating ServeGen's advantage in performance benchmarking. ServeGen is available at https://github.com/alibaba/ServeGen.

Yifei Xu, Yuning Chen, Xumiao Zhang et al.

Among the thriving ecosystem of cloud computing and the proliferation of Large Language Model (LLM)-based code generation tools, there is a lack of benchmarking for code generation in cloud-native applications. In response to this need, we present CloudEval-YAML, a practical benchmark for cloud configuration generation. CloudEval-YAML tackles the diversity challenge by focusing on YAML, the de facto standard of numerous cloud-native tools. We develop the CloudEval-YAML benchmark with practicality in mind: the dataset consists of hand-written problems with unit tests targeting practical scenarios. We further enhanced the dataset to meet practical needs by rephrasing questions in a concise, abbreviated, and bilingual manner. The dataset consists of 1011 problems that take more than 1200 human hours to complete. To improve practicality during evaluation, we build a scalable evaluation platform for CloudEval-YAML that achieves a 20 times speedup over a single machine. To the best of our knowledge, the CloudEval-YAML dataset is the first hand-written dataset targeting cloud-native applications. We present an in-depth evaluation of 12 LLMs, leading to a deeper understanding of the problems and LLMs, as well as effective methods to improve task performance and reduce cost.

Shaoke Xi, ChonLam Lao, Boyi Jia et al.

Large language model (LLM) training today runs on clusters spanning thousands of GPUs. While this scale enables rapid model advances, developing, debugging, and performance-tuning the training framework inevitably becomes complex and costly. This is because engineers often need to reproduce production behaviors to diagnose failures or evaluate optimizations, thereby demanding frequent and even exclusive access to production-scale clusters -- which becomes increasingly hard given that the majority of GPUs are already committed to production workloads. Simulation relies on complex performance models that are difficult to maintain, and downscaled experiments often fail to capture scale-dependent behaviors. We present PrismLLM to decouple large-scale execution from the need to access large clusters, enabling engineers to run and observe ranks of interest under faithful large-scale behavior using only a few GPUs. PrismLLM constructs a high-fidelity execution graph via a slicing-based approach that captures computation, communication, and dependencies of the target scale. Then, PrismLLM performs hybrid emulation where selected ranks execute the original program while the remaining ranks are replayed as virtual participants. Experiments on large-scale LLM training workloads show that PrismLLM accurately reproduces performance and memory behavior, achieving only 0.58\% average error in iteration time and less than 0.01\% error in peak GPU memory usage. PrismLLM can emulate clusters of up to 8192 GPUs using fewer than 1\% of the physical GPUs required by the original deployment.

Shan Yu, Junyi Shu, Yuanjiang Ni et al.

As LLM applications grow more complex, developers are increasingly adopting multi-agent architectures to decompose workflows into specialized, collaborative components, introducing structure that constrains agent behavior and exposes useful semantic predictability. Unlike traditional LLM serving, which operates under highly dynamic and uncertain conditions, this structured topology enables opportunities to reduce runtime uncertainty -- yet existing systems fail to exploit it, treating agentic workloads as generic traffic and incurring significant inefficiencies. Our analysis of production traces from an agent-serving platform and an internal coding assistant reveals key bottlenecks, including low prefix cache hit rates, severe resource contention from long-context requests, and substantial queuing delays due to suboptimal scaling. To address these challenges, we propose Pythia, a multi-agent serving system that captures workflow semantics through a simple interface at the serving layer, unlocking new optimization opportunities and substantially improving throughput and job completion time over state-of-the-art baselines.

ChonLam Lao, Minlan Yu, Aditya Akella et al.

Large-scale ML training jobs are frequently interrupted by hardware and software anomalies, failures, and management events. Existing solutions like checkpointing or runtime reconfiguration suffer from long downtimes, degraded performance, or undesired changes to training strategies. We present TrainMover, a resilient runtime that leverages standby machines to handle interruptions with minimal downtime and zero memory overhead. To achieve these goals, TrainMover introduces two key techniques: two-phase, delta-based communication group setups and communication-free sandboxed shadow iterations. Our evaluation shows that TrainMover consistently achieves second-level downtime across all evaluated models during migration, maintaining 99\% training efficiency during periodic 10-minute rebalancing. We also demonstrate the effectiveness of TrainMover in handling various interruptions.

Yu Guan, Zhiyu Yin, Haoyu Chen et al.

Troubleshooting performance problems of large model training (LMT) is immensely challenging, due to unprecedented scales of modern GPU clusters, the complexity of software-hardware interactions, and the data intensity of the training process. Existing troubleshooting approaches designed for traditional distributed systems or datacenter networks fall short and can hardly apply to real-world training systems. In this paper, we present PerfTracker, the first online troubleshooting system utilizing fine-grained profiling, to diagnose performance issues of large-scale model training in production. PerfTracker can diagnose performance issues rooted in both hardware (e.g., GPUs and their interconnects) and software (e.g., Python functions and GPU operations). It scales to LMT on modern GPU clusters. PerfTracker effectively summarizes runtime behavior patterns of fine-grained LMT functions via online profiling, and leverages differential observability to localize the root cause with minimal production impact. PerfTracker has been deployed as a production service for large-scale GPU clusters of O(10, 000) GPUs (product homepage https://help.aliyun.com/zh/pai/user-guide/perftracker-online-performance-analysis-diagnostic-tool). It has been used to diagnose a variety of difficult performance issues.

Shaoke Xi, Jiaqi Gao, Mengqi Liu et al.

With the growing performance requirements on networked applications, there is a new trend of offloading stateful network applications to SmartNICs to improve performance and reduce the total cost of ownership. However, offloading stateful network applications is non-trivial due to state operation complexity, state resource consumption, and the complicated relationship between traffic and state. Naively partitioning the program by state or traffic can result in a suboptimal partition plan with higher CPU usage or even packet drops. In this paper, we propose Cora, a compiler and runtime that offloads stateful network applications to SmartNIC-accelerated hosts. Cora compiler introduces an accurate performance model for each SmartNIC and employs an efficient compiling algorithm to search the offloading plan. Cora runtime can monitor traffic dynamics and adapt to minimize CPU usage. Cora is built atop Netronome Agilio and BlueField 2 SmartNICs. Our evaluation shows that for the same throughput target, Cora can propose partition plans saving up to 94.0% CPU cores, 1.9 times more than baseline solutions. Under the same resource constraint, Cora can accelerate network functions by 44.9%-82.3%. Cora runtime can adapt to traffic changes and keep CPU usage low.

Jianbo Dong, Bin Luo, Jun Zhang et al.

The emergence of Large Language Models (LLMs) has necessitated the adoption of distributed training techniques, involving the deployment of thousands of GPUs to train a single model. Unfortunately, the efficiency of large-scale distributed training systems is often suboptimal due to the increased likelihood of hardware errors in high-end GPU products and the heightened risk of network traffic collisions. Moreover, any local hardware failure can disrupt training tasks, and the inability to swiftly identify faulty components leads to a significant waste of GPU resources. And, prolonged communication due to traffic collisions can substantially increase GPU waiting times. To address these challenges, we propose a communication-driven solution, namely the C4. The key insights of C4 are twofold. First, the load in distributed training exhibits homogeneous characteristics and is divided into iterations through periodic synchronization, therefore hardware anomalies would incur certain syndrome in collective communication. By leveraging this feature, C4 can rapidly identify the faulty components, swiftly isolate the anomaly, and restart the task, thereby avoiding resource wastage caused by delays in anomaly detection. Second, the predictable communication model of collective communication, involving a limited number of long-lived flows, allows C4 to efficiently execute traffic planning, substantially reducing bandwidth competition among these flows. The C4 has been extensively deployed across real-world production systems in a hyperscale cloud provider, yielding a significant improvement in system efficiency, from 30% to 45%. This enhancement is attributed to a 30% reduction in error-induced overhead and a 15% reduction in communication costs.

Mark Santolucito, Jialu Zhang, Ennan Zhai et al.

Continuous Integration (CI) testing is a popular software development technique that allows developers to easily check that their code can build successfully and pass tests across various system environments. In order to use a CI platform, a developer must include a set of configuration files to a code repository for specifying build conditions. Incorrect configuration settings lead to CI build failures, which can take hours to run, wasting valuable developer time and delaying product release dates. Debugging CI configurations is challenging because users must manage configurations for the build across many system environments, to which they may not have local access. Thus, the only way to check a CI configuration is to push a commit and wait for the build result. To address this problem, we present the first approach, VeriCI, for statically checking for errors in a given CI configuration before the developer pushes a commit to build on the CI server. Our key insight is that the repositories in a CI environment contain lists of build histories which offer the time-aware repository build status. Driven by this insight, we introduce the Misclassification Guided Abstraction Refinement (MiGAR) loop that automates part of the learning process across the heterogeneous build environments in CI. We then use decision tree learning to generate constraints on the CI configuration that must hold for a build to succeed by training on a large history of continuous integration repository build results. We evaluate VeriCI on real-world data from GitHub and find that we have 83% accuracy of predicting a build failure.

Ludovic Barman, Italo Dacosta, Mahdi Zamani et al.

Organizational networks are vulnerable to traffic-analysis attacks that enable adversaries to infer sensitive information from the network traffic - even if encryption is used. Typical anonymous communication networks are tailored to the Internet and are poorly suited for organizational networks. We present PriFi, an anonymous communication protocol for LANs, which protects users against eavesdroppers and provides high-performance traffic-analysis resistance. PriFi builds on Dining Cryptographers networks but reduces the high communication latency of prior work via a new client/relay/server architecture, in which a client's packets remain on their usual network path without additional hops, and in which a set of remote servers assist the anonymization process without adding latency. PriFi also solves the challenge of equivocation attacks, which are not addressed by related works, by encrypting the traffic based on the communication history. Our evaluation shows that PriFi introduces a small latency overhead (~100ms for 100 clients) and is compatible with delay-sensitive applications such as VoIP.