arXiv subjects

Xuanzhe Liu

Publications and source records attributed to Xuanzhe Liu.

At least 55 records · Page 3Linked to original sources

arXiv2024

LoongTrain: Efficient Training of Long-Sequence LLMs with Head-Context Parallelism

Efficiently training LLMs with long sequences is important yet challenged by the massive computation and memory requirements. Sequence parallelism has been proposed to tackle these problems, but existing methods suffer from scalability or efficiency issues. We propose LoongTrain, a novel system to efficiently train LLMs with long sequences at scale. The core of LoongTrain is the 2D-Attention mechanism, which combines both head-parallel and context-parallel techniques to break the scalability constraints while maintaining efficiency. We introduce Double-Ring-Attention and analyze the performance of device placement strategies to further speed up training. We implement LoongTrain with the hybrid ZeRO and Selective Checkpoint++ techniques. Experiment results show that LoongTrain outperforms state-of-the-art baselines, i.e., DeepSpeed-Ulysses and Megatron Context Parallelism, in both end-to-end training speed and scalability, and improves Model FLOPs Utilization (MFU) by up to 2.88x.

cs.DC↗

arXiv2024

DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving

DistServe improves the performance of large language models (LLMs) serving by disaggregating the prefill and decoding computation. Existing LLM serving systems colocate the two phases and batch the computation of prefill and decoding across all users and requests. We find that this strategy not only leads to strong prefill-decoding interferences but also couples the resource allocation and parallelism plans for both phases. LLM applications often emphasize individual latency for each phase: time to first token (TTFT) for the prefill phase and time per output token (TPOT) of each request for the decoding phase. In the presence of stringent latency requirements, existing systems have to prioritize one latency over the other, or over-provision compute resources to meet both. DistServe assigns prefill and decoding computation to different GPUs, hence eliminating prefill-decoding interferences. Given the application's TTFT and TPOT requirements, DistServe co-optimizes the resource allocation and parallelism strategy tailored for each phase. DistServe also places the two phases according to the serving cluster's bandwidth to minimize the communication caused by disaggregation. As a result, DistServe significantly improves LLM serving performance in terms of the maximum rate that can be served within both TTFT and TPOT constraints on each GPU. Our evaluations show that on various popular LLMs, applications, and latency requirements, DistServe can serve 7.4x more requests or 12.6x tighter SLO, compared to state-of-the-art systems, while staying within latency constraints for > 90% of requests.

cs.DC↗

arXiv2024

RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.

cs.DC↗

arXiv2024

LLM as a System Service on Mobile Devices

Being more powerful and intrusive into user-device interactions, LLMs are eager for on-device execution to better preserve user privacy. In this work, we propose a new paradigm of mobile AI: LLM as a system service on mobile devices (LLMaaS). Unlike traditional DNNs that execute in a stateless manner, such a system service is stateful: LLMs execution often needs to maintain persistent states (mainly KV cache) across multiple invocations. To minimize the LLM context switching overhead under tight device memory budget, this work presents LLMS, which decouples the memory management of app and LLM contexts with a key idea of fine-grained, chunk-wise, globally-optimized KV cache compression and swapping. By fully leveraging KV cache's unique characteristics, it proposes three novel techniques: (1) Tolerance-Aware Compression: it compresses chunks based on their measured accuracy tolerance to compression. (2) IO-Recompute Pipelined Loading: it introduces recompute to swapping-in for acceleration. (3) Chunk Lifecycle Management: it optimizes the memory activities of chunks with an ahead-of-time swapping-out and an LCTRU (Least Compression-Tolerable and Recently-Used) queue based eviction. In evaluations conducted on well-established traces and various edge devices, \sys reduces context switching latency by up to 2 orders of magnitude when compared to competitive baseline solutions.

cs.OS↗

arXiv2023

LLMCad: Fast and Scalable On-device Large Language Model Inference

Generative tasks, such as text generation and question answering, hold a crucial position in the realm of mobile applications. Due to their sensitivity to privacy concerns, there is a growing demand for their execution directly on mobile devices. Currently, the execution of these generative tasks heavily depends on Large Language Models (LLMs). Nevertheless, the limited memory capacity of these devices presents a formidable challenge to the scalability of such models. In our research, we introduce LLMCad, an innovative on-device inference engine specifically designed for efficient generative Natural Language Processing (NLP) tasks. The core idea behind LLMCad revolves around model collaboration: a compact LLM, residing in memory, takes charge of generating the most straightforward tokens, while a high-precision LLM steps in to validate these tokens and rectify any identified errors. LLMCad incorporates three novel techniques: (1) Instead of generating candidate tokens in a sequential manner, LLMCad employs the smaller LLM to construct a token tree, encompassing a wider range of plausible token pathways. Subsequently, the larger LLM can efficiently validate all of these pathways simultaneously. (2) It employs a self-adjusting fallback strategy, swiftly initiating the verification process whenever the smaller LLM generates an erroneous token. (3) To ensure a continuous flow of token generation, LLMCad speculatively generates tokens during the verification process by implementing a compute-IO pipeline. Through an extensive series of experiments, LLMCad showcases an impressive token generation speed, achieving rates up to 9.3x faster than existing inference engines.

cs.NI↗

arXiv2023

Energy-Efficient GPU Clusters Scheduling for Deep Learning

Training deep neural networks (DNNs) is a major workload in datacenters today, resulting in a tremendously fast growth of energy consumption. It is important to reduce the energy consumption while completing the DL training jobs early in data centers. In this paper, we propose PowerFlow, a GPU clusters scheduler that reduces the average Job Completion Time (JCT) under an energy budget. We first present performance models for DL training jobs to predict the throughput and energy consumption performance with different configurations. Based on the performance models, PowerFlow dynamically allocates GPUs and adjusts the GPU-level or job-level configurations of DL training jobs. PowerFlow applies network packing and buddy allocation to job placement, thus avoiding extra energy consumed by cluster fragmentations. Evaluation results show that under the same energy consumption, PowerFlow improves the average JCT by 1.57 - 3.39 x at most, compared to competitive baselines.

cs.DC↗

arXiv2023

Rise of Distributed Deep Learning Training in the Big Model Era: From a Software Engineering Perspective

Deep learning (DL) has become a key component of modern software. In the "big model" era, the rich features of DL-based software substantially rely on powerful DL models, e.g., BERT, GPT-3, and the recently emerging GPT-4, which are trained on the powerful cloud with large datasets. Hence, training effective DL models has become a vital stage in the whole software lifecycle. When training deep learning models, especially those big models, developers need to parallelize and distribute the computation and memory resources amongst multiple devices in the training process, which is known as distributed deep learning training, or distributed training for short. However, the unique challenges that developers encounter in distributed training process have not been studied in the software engineering community. Given the increasingly heavy dependence of current DL-based software on distributed training, this paper aims to fill in the knowledge gap and presents the first comprehensive study on developers' issues in distributed training. To this end, we analyze 1,131 real-world developers' issues about using these frameworks reported on Stack Overflow and GitHub. We construct a fine-grained taxonomy consisting of 30 categories regarding the fault symptoms and summarize common fix patterns for different symptoms. Based on the results, we suggest actionable implications on research avenues that can potentially facilitate the distributed training to develop DL-based software, such as focusing on the frequent and common fix patterns when designing testing or debugging tools, developing efficient testing and debugging techniques for communication configuration along with the synthesis of network configuration analysis, designing new multi-device checkpoint-and-replay techniques to help reproduction, and designing serverless APIs for cloud platforms.

cs.SE↗

arXiv2023

MuxFlow: Efficient and Safe GPU Sharing in Large-Scale Production Deep Learning Clusters

Large-scale GPU clusters are widely-used to speed up both latency-critical (online) and best-effort (offline) deep learning (DL) workloads. However, most DL clusters either dedicate each GPU to one workload or share workloads in time, leading to very low GPU resource utilization. We present MuxFlow, the first production cluster system that supports efficient and safe space-sharing for DL workloads. NVIDIA MPS provides an opportunity to share multiple workloads in space on widely-deployed NVIDIA GPUs, but it cannot guarantee the performance and safety of online workloads. MuxFlow introduces a two-level protection mechanism for memory and computation to guarantee the performance of online workloads. Based on our practical error analysis, we design a mixed error-handling mechanism to guarantee the safety of online workloads. MuxFlow further proposes dynamic streaming multiprocessor (SM) allocation and matching-based scheduling to improve the efficiency of offline workloads. MuxFlow has been deployed at CompanyX's clusters with more than 20,000 GPUs. The deployment results indicate that MuxFlow substantially improves the GPU utilization from 26$\%$ to 76$\%$, SM activity from 16$\%$ to 33$\%$, and GPU memory from 42$\%$ to 48$\%$.

cs.DC↗

arXiv2023

FaaSLight: General Application-Level Cold-Start Latency Optimization for Function-as-a-Service in Serverless Computing

Serverless computing is a popular cloud computing paradigm that frees developers from server management. Function-as-a-Service (FaaS) is the most popular implementation of serverless computing, representing applications as event-driven and stateless functions. However, existing studies report that functions of FaaS applications severely suffer from cold-start latency. In this paper, we propose an approach namely FaaSLight to accelerating the cold start for FaaS applications through application-level optimization. We first conduct a measurement study to investigate the possible root cause of the cold start problem of FaaS. The result shows that application code loading latency is a significant overhead. Therefore, loading only indispensable code from FaaS applications can be an adequate solution. Based on this insight, we identify code related to application functionalities by constructing the function-level call graph, and separate other code (i.e., optional code) from FaaS applications. The separated optional code can be loaded on demand to avoid the inaccurate identification of indispensable code causing application failure. In particular, a key principle guiding the design of FaaSLight is inherently general, i.e., platform- and language-agnostic. The evaluation results on real-world FaaS applications show that FaaSLight can significantly reduce the code loading latency (up to 78.95%, 28.78% on average), thereby reducing the cold-start latency. As a result, the total response latency of functions can be decreased by up to 42.05% (19.21% on average). Compared with the state-of-the-art, FaaSLight achieves a 21.25X improvement in reducing the average total response latency.

cs.DC↗

arXiv2023

Characterizing and Detecting WebAssembly Runtime Bugs

WebAssembly (abbreviated WASM) has emerged as a promising language of the Web and also been used for a wide spectrum of software applications such as mobile applications and desktop applications. These applications, named as WASM applications, commonly run in WASM runtimes. Bugs in WASM runtimes are frequently reported by developers and cause the crash of WASM applications. However, these bugs have not been well studied. To fill in the knowledge gap, we present a systematic study to characterize and detect bugs in WASM runtimes. We first harvest a dataset of 311 real-world bugs from hundreds of related posts on GitHub. Based on the collected high-quality bug reports, we distill 31 bug categories of WASM runtimes and summarize their common fix strategies. Furthermore, we develop a pattern-based bug detection framework to automatically detect bugs in WASM runtimes. We apply the detection framework to five popular WASM runtimes and successfully uncover 53 bugs that have never been reported previously, among which 14 have been confirmed and 6 have been fixed by runtime developers.

cs.SE↗

arXiv2022

Rise of the Planet of Serverless Computing: A Systematic Review

Serverless computing is an emerging cloud computing paradigm, being adopted to develop a wide range of software applications. It allows developers to focus on the application logic in the granularity of function, thereby freeing developers from tedious and error-prone infrastructure management. Meanwhile, its unique characteristic poses new challenges to the development and deployment of serverless-based applications. To tackle these challenges, enormous research efforts have been devoted. This paper provides a comprehensive literature review to characterize the current research state of serverless computing. Specifically, this paper covers 164 papers on 17 research directions of serverless computing, including performance optimization, programming framework, application migration, multi-cloud development, testing and debugging, etc. It also derives research trends, focus, and commonly-used platforms for serverless computing, as well as promising research opportunities.

cs.SE↗

arXiv2022

Software Engineering for Serverless Computing

Serverless computing is an emerging cloud computing paradigm that has been applied to various domains, including machine learning, scientific computing, video processing, etc. To develop serverless computing-based software applications (a.k.a., serverless applications), developers follow the new cloud-based software architecture, where they develop event-driven applications without the need for complex and error-prone server management. The great demand for developing serverless applications poses unique challenges to software developers. However, Software Engineering (SE) has not yet wholeheartedly tackled these challenges. In this paper, we outline a vision for how SE can facilitate the development of serverless applications and call for actions by the SE research community to reify this vision. Specifically, we discuss possible directions in which researchers and cloud providers can facilitate serverless computing from the SE perspective, including configuration management, data security, application migration, performance, testing and debugging, etc.

cs.SE↗

arXiv2022

Federated Neural Architecture Search

To preserve user privacy while enabling mobile intelligence, techniques have been proposed to train deep neural networks on decentralized data. However, training over decentralized data makes the design of neural architecture quite difficult as it already was. Such difficulty is further amplified when designing and deploying different neural architectures for heterogeneous mobile platforms. In this work, we propose an automatic neural architecture search into the decentralized training, as a new DNN training paradigm called Federated Neural Architecture Search, namely federated NAS. To deal with the primary challenge of limited on-client computational and communication resources, we present FedNAS, a highly optimized framework for efficient federated NAS. FedNAS fully exploits the key opportunity of insufficient model candidate re-training during the architecture search process, and incorporates three key optimizations: parallel candidates training on partial clients, early dropping candidates with inferior performance, and dynamic round numbers. Tested on large-scale datasets and typical CNN architectures, FedNAS achieves comparable model accuracy as state-of-the-art NAS algorithm that trains models with centralized data, and also reduces the client cost by up to two orders of magnitude compared to a straightforward design of federated NAS.

cs.LG↗

arXiv2022

Benchmarking of DL Libraries and Models on Mobile Devices

Deploying deep learning (DL) on mobile devices has been a notable trend in recent years. To support fast inference of on-device DL, DL libraries play a critical role as algorithms and hardware do. Unfortunately, no prior work ever dives deep into the ecosystem of modern DL libs and provides quantitative results on their performance. In this paper, we first build a comprehensive benchmark that includes 6 representative DL libs and 15 diversified DL models. We then perform extensive experiments on 10 mobile devices, which help reveal a complete landscape of the current mobile DL libs ecosystem. For example, we find that the best-performing DL lib is severely fragmented across different models and hardware, and the gap between those DL libs can be rather huge. In fact, the impacts of DL libs can overwhelm the optimizations from algorithms or hardware, e.g., model quantization and GPU/DSP-based heterogeneous computing. Finally, atop the observations, we summarize practical implications to different roles in the DL lib ecosystem.

cs.LG↗

arXiv2022

Mandheling: Mixed-Precision On-Device DNN Training with DSP Offloading

This paper proposes Mandheling, the first system that enables highly resource-efficient on-device training by orchestrating the mixed-precision training with on-chip Digital Signal Processing (DSP) offloading. Mandheling fully explores the advantages of DSP in integer-based numerical calculation by four novel techniques: (1) a CPU-DSP co-scheduling scheme to mitigate the overhead from DSP-unfriendly operators; (2) a self-adaptive rescaling algorithm to reduce the overhead of dynamic rescaling in backward propagation; (3) a batch-splitting algorithm to improve the DSP cache efficiency; (4) a DSP-compute subgraph reusing mechanism to eliminate the preparation overhead on DSP. We have fully implemented Mandheling and demonstrate its effectiveness through extensive experiments. The results show that, compared to the state-of-the-art DNN engines from TFLite and MNN, Mandheling reduces the per-batch training time by 5.5$\times$ and the energy consumption by 8.9$\times$ on average. In end-to-end training tasks, Mandheling reduces up to 10.7$\times$ convergence time and 13.1$\times$ energy consumption, with only 1.9%-2.7% accuracy loss compared to the FP32 precision setting.

cs.NI↗

arXiv2021

Exploring the Generalizability of Spatio-Temporal Traffic Prediction: Meta-Modeling and an Analytic Framework

The Spatio-Temporal Traffic Prediction (STTP) problem is a classical problem with plenty of prior research efforts that benefit from traditional statistical learning and recent deep learning approaches. While STTP can refer to many real-world problems, most existing studies focus on quite specific applications, such as the prediction of taxi demand, ridesharing order, traffic speed, and so on. This hinders the STTP research as the approaches designed for different applications are hardly comparable, and thus how an application-driven approach can be generalized to other scenarios is unclear. To fill in this gap, this paper makes three efforts: (i) we propose an analytic framework, called STAnalytic, to qualitatively investigate STTP approaches regarding their design considerations on various spatial and temporal factors, aiming to make different application-driven approaches comparable; (ii) we design a spatio-temporal meta-model, called STMeta, which can flexibly integrate generalizable temporal and spatial knowledge identified by STAnalytic, (iii) we build an STTP benchmark platform including ten real-life datasets with five scenarios to quantitatively measure the generalizability of STTP approaches. In particular, we implement STMeta with different deep learning techniques, and STMeta demonstrates better generalizability than state-of-the-art approaches by achieving lower prediction error on average across all the datasets.

cs.LG↗

arXiv2021

From Cloud to Edge: A First Look at Public Edge Platforms

Public edge platforms have drawn increasing attention from both academia and industry. In this study, we perform a first-of-its-kind measurement study on a leading public edge platform that has been densely deployed in China. Based on this measurement, we quantitatively answer two critical yet unexplored questions. First, from end users' perspective, what is the performance of commodity edge platforms compared to cloud, in terms of the end-to-end network delay, throughput, and the application QoE. Second, from the edge service provider's perspective, how are the edge workloads different from cloud, in terms of their VM subscription, monetary cost, and resource usage. Our study quantitatively reveals the status quo of today's public edge platforms, and provides crucial insights towards developing and operating future edge services.

cs.NI↗

arXiv2021

Characterizing Impacts of Heterogeneity in Federated Learning upon Large-Scale Smartphone Data

Federated learning (FL) is an emerging, privacy-preserving machine learning paradigm, drawing tremendous attention in both academia and industry. A unique characteristic of FL is heterogeneity, which resides in the various hardware specifications and dynamic states across the participating devices. Theoretically, heterogeneity can exert a huge influence on the FL training process, e.g., causing a device unavailable for training or unable to upload its model updates. Unfortunately, these impacts have never been systematically studied and quantified in existing FL literature. In this paper, we carry out the first empirical study to characterize the impacts of heterogeneity in FL. We collect large-scale data from 136k smartphones that can faithfully reflect heterogeneity in real-world settings. We also build a heterogeneity-aware FL platform that complies with the standard FL protocol but with heterogeneity in consideration. Based on the data and the platform, we conduct extensive experiments to compare the performance of state-of-the-art FL algorithms under heterogeneity-aware and heterogeneity-unaware settings. Results show that heterogeneity causes non-trivial performance degradation in FL, including up to 9.2% accuracy drop, 2.32x lengthened training time, and undermined fairness. Furthermore, we analyze potential impact factors and find that device failure and participant bias are two potential factors for performance degradation. Our study provides insightful implications for FL practitioners. On the one hand, our findings suggest that FL algorithm designers consider necessary heterogeneity during the evaluation. On the other hand, our findings urge system providers to design specific mechanisms to mitigate the impacts of heterogeneity.

cs.LG↗