arXiv subjects

Menglu Yu

Publications and source records attributed to Menglu Yu.

7 recordsLinked to original sources

arXiv2026

Optimizing Effective Training Time for Large-Scale Recommendation Systems

Lifecycle overhead silently consumes accelerator capacity across large-scale recommendation training fleets. Our largest recommendation workloads process tens of billions train- ing examples per day on thousands of GPUs. Before this work, only 50-60% of their end-to-end wall time advanced training on new data. We present a fleet-scale study of this lifecycle overhead and a set of optimizations spanning the full training stack. We use Effective Training Time (ETT%) as an operational framework to instrument lost time, localize it to independently owned infrastructure components, and expose work repeated across job restarts. This analysis guides optimizations like communication elimination and pipeline overlap during trainer initialization; dynamic-shape handling, autotuning pruning, and reusable Py- Torch 2 compilation caches; asynchronous checkpointing; stan- dalone model publishing; and reductions in recovery cost. We evaluate the optimizations on representative models and measure their impacts in our training fleet. ETT% improves on every benchmark, by 15.5% on average, and reaches 85% on our largest workload. Fleet-wide ETT% rose from about 80% to above 90% after deployment.

cs.IR↗

arXiv2026

To Solve Bilevel Optimization with Nonconvex Lower Levels, We Need Second-Order Stationarity

Although bilevel optimization (BLO) has emerged as a powerful framework for addressing many complex and nested machine learning problems in recent years, most existing studies are confined to the lower-level strongly convex (LLSC) or lower-level generally convex (LLGC) settings (i.e., the lower-level objective function is assumed to be, at least, convex). While the LLSC/LLGC assumptions render more tractable algorithmic design and theoretical analysis, they are too rigid to encompass many machine learning problems in practice. The limitations of LLSC/LLGC assumptions in BLO motivate us to investigate solving the BLO problem in the general lower-level nonconvex (LLNC) settings, which remains in its infancy. In the literature on LLNC-BLO, most of the existing works either require additional structures in the lower-level objective function for tractable theoretical analysis, or adopt the first-order stationarity reformulation as a lower-level surrogate problem, which is inherited from the LLSC/LLGC settings but could lose their effectiveness in the LLNC setting. To bridge this gap, we propose to reformulate the nonconvex lower-level problem using a second-order stationarity-based surrogate, the solution of which guarantees a local optimal solution at the lower level. Based on this reformulation, we propose the PROBE (Perturbed gradient algorithm for bilevel problem) and show that it overcomes the limitations of prior works by probing and escaping lower-level saddle points. We prove that PROBE achieves a finite-time convergence rate of $O(T^{-2/5})$, where T denotes iterations. To our knowledge, this work is the first to establish the finite-time convergence for achieving lower-level second-order stationary solutions in general LLNC-BLO. Our experiments on both a large language model-based data curation task and a meta-learning task also show that PROBE outperforms SOTA methods.

cs.LG↗

arXiv2026

Optimus: A Generic Operator-Level PyTorch Model Transformation Framework

In large-scale industrial applications, deep learning models that power recommendation and ranking have complex and diverse model architectures. These models are continuously developed and refined by large teams of machine learning engineers, rendering manual optimization infeasible. Consequently, graph-based optimization techniques have become an industry standard for boosting performance, with PyTorch FX transformations leading the charge. These transformations typically rely on a set of human-engineered module-level rewrite rules which are not scalable to diverse model architectures. To address this limitation, we introduce Optimus, a general-purpose model transformation framework built in the PyTorch 2.x (PT2) machine learning compiler. With a concise set of predefined patterns, Optimus applies an efficient greedy search algorithm for pattern matching and replacement, while preserving model semantic. It is designed and implemented as a highly customizable and extensible framework integrated into the PT2 stack. Our evaluation shows that the framework can achieve up to 63% speedup, 6% peak memory reduction, and over 400 second compile time decrease for our industry-scale recommendation models compared to baselines. Optimus is open-sourced together with PyTorch 2.x as a customizable model transformation layer.

cs.PF↗

arXiv2022

On Scheduling Ring-All-Reduce Learning Jobs in Multi-Tenant GPU Clusters with Communication Contention

Powered by advances in deep learning (DL) techniques, machine learning and artificial intelligence have achieved astonishing successes. However, the rapidly growing needs for DL also led to communication- and resource-intensive distributed training jobs for large-scale DL training, which are typically deployed over GPU clusters. To sustain the ever-increasing demand for DL training, the so-called "ring-all-reduce" (RAR) technologies have recently emerged as a favorable computing architecture to efficiently process network communication and computation load in GPU clusters. The most salient feature of RAR is that it removes the need for dedicated parameter servers, thus alleviating the potential communication bottleneck. However, when multiple RAR-based DL training jobs are deployed over GPU clusters, communication bottlenecks could still occur due to contentions between DL training jobs. So far, there remains a lack of theoretical understanding on how to design contention-aware resource scheduling algorithms for RAR-based DL training jobs, which motivates us to fill this gap in this work. Our main contributions are three-fold: i) We develop a new analytical model that characterizes both communication overhead related to the worker distribution of the job and communication contention related to the co-location of different jobs; ii) Based on the proposed analytical model, we formulate the problem as a non-convex integer program to minimize the makespan of all RAR-based DL training jobs. To address the unique structure in this problem that is not amenable for optimization algorithm design, we reformulate the problem into an integer linear program that enables provable approximation algorithm design called SJF-BCO (Smallest Job First with Balanced Contention and Overhead); and iii) We conduct extensive experiments to show the superiority of SJF-BCO over existing schedulers.

cs.DC↗

arXiv2022

Toward Efficient Online Scheduling for Distributed Machine Learning Systems

Recent years have witnessed a rapid growth of distributed machine learning (ML) frameworks, which exploit the massive parallelism of computing clusters to expedite ML training. However, the proliferation of distributed ML frameworks also introduces many unique technical challenges in computing system design and optimization. In a networked computing cluster that supports a large number of training jobs, a key question is how to design efficient scheduling algorithms to allocate workers and parameter servers across different machines to minimize the overall training time. Toward this end, in this paper, we develop an online scheduling algorithm that jointly optimizes resource allocation and locality decisions. Our main contributions are three-fold: i) We develop a new analytical model that considers both resource allocation and locality; ii) Based on an equivalent reformulation and observations on the worker-parameter server locality configurations, we transform the problem into a mixed packing and covering integer program, which enables approximation algorithm design; iii) We propose a meticulously designed approximation algorithm based on randomized rounding and rigorously analyze its performance. Collectively, our results contribute to the state of the art of distributed ML system optimization and algorithm design.

cs.DC↗

arXiv2022

GADGET: Online Resource Optimization for Scheduling Ring-All-Reduce Learning Jobs

Fueled by advances in distributed deep learning (DDL), recent years have witnessed a rapidly growing demand for resource-intensive distributed/parallel computing to process DDL computing jobs. To resolve network communication bottleneck and load balancing issues in distributed computing, the so-called ``ring-all-reduce'' decentralized architecture has been increasingly adopted to remove the need for dedicated parameter servers. To date, however, there remains a lack of theoretical understanding on how to design resource optimization algorithms for efficiently scheduling ring-all-reduce DDL jobs in computing clusters. This motivates us to fill this gap by proposing a series of new resource scheduling designs for ring-all-reduce DDL jobs. Our contributions in this paper are three-fold: i) We propose a new resource scheduling analytical model for ring-all-reduce deep learning, which covers a wide range of objectives in DDL performance optimization (e.g., excessive training avoidance, energy efficiency, fairness); ii) Based on the proposed performance analytical model, we develop an efficient resource scheduling algorithm called GADGET (greedy ring-all-reduce distributed graph embedding technique), which enjoys a provable strong performance guarantee; iii) We conduct extensive trace-driven experiments to demonstrate the effectiveness of the GADGET approach and its superiority over the state of the art.

cs.DC↗

arXiv2021

A Sum-of-Ratios Multi-Dimensional-Knapsack Decomposition for DNN Resource Scheduling

In recent years, to sustain the resource-intensive computational needs for training deep neural networks (DNNs), it is widely accepted that exploiting the parallelism in large-scale computing clusters is critical for the efficient deployments of DNN training jobs. However, existing resource schedulers for traditional computing clusters are not well suited for DNN training, which results in unsatisfactory job completion time performance. The limitations of these resource scheduling schemes motivate us to propose a new computing cluster resource scheduling framework that is able to leverage the special layered structure of DNN jobs and significantly improve their job completion times. Our contributions in this paper are three-fold: i) We develop a new resource scheduling analytical model by considering DNN's layered structure, which enables us to analytically formulate the resource scheduling optimization problem for DNN training in computing clusters; ii) Based on the proposed performance analytical model, we then develop an efficient resource scheduling algorithm based on the widely adopted parameter-server architecture using a sum-of-ratios multi-dimensional-knapsack decomposition (SMD) method to offer strong performance guarantee; iii) We conduct extensive numerical experiments to demonstrate the effectiveness of the proposed schedule algorithm and its superior performance over the state of the art.

cs.DC↗