An official website of the United States government
Training large language models (LLMs) increasingly relies on geographically distributed accelerators, causing prohibitive communication costs across regions and uneven utilization of heterogeneous hardware. We propose HALoS, a hierarchical asynchronous optimization framework that tackles these issues by introducing local parameter servers (LPSs) within each region and a global parameter server (GPS) that merges updates across regions. This hierarchical design minimizes expensive inter-region communication, reduces straggler effects, and leverages fast intra-region links. We provide a rigorous convergence analysis for HALoS under non-convex objectives, including theoretical guarantees on the role of hierarchical momentum in asynchronous training. Empirically, HALoS attains up to 7.5x faster convergence than synchronous baselines in geo-distributed LLM training and improves upon existing asynchronous methods by up to 2.1x. Crucially, HALoS preserves the model quality of fully synchronous SGD-matching or exceeding accuracy on standard language modeling and downstream benchmarks-while substantially lowering total training time. These results demonstrate that hierarchical, server-side update accumulation and global model merging are powerful tools for scalable, efficient training of new-era LLMs in heterogeneous, geo-distributed environments.
more »
« less
This content will become publicly available on June 5, 2026
HALoS: Hierarchical Asynchronous Local SGD over Slow Networks for Geo-Distributed Large Language Model Training
Training large language models (LLMs) increasingly relies on geographically distributed accelerators, causing prohibitive communication costs across regions and uneven utilization of heterogeneous hardware. We propose HALoS, a hierarchical asynchronous optimization framework that tackles these issues by introducing local parameter servers (LPSs) within each region and a global parameter server (GPS) that merges updates across regions. This hierarchical design minimizes expensive inter-region communication, reduces straggler effects, and leverages fast intra-region links. We provide a rigorous convergence analysis for HALoS under non-convex objectives, including theoretical guarantees on the role of hierarchical momentum in asynchronous training. Empirically, HALoS attains up to 7.5x faster convergence than synchronous baselines in geo-distributed LLM training and improves upon existing asynchronous methods by up to 2.1x. Crucially, HALoS preserves the model quality of fully synchronous SGD-matching or exceeding accuracy on standard language modeling and downstream benchmarks-while substantially lowering total training time. These results demonstrate that hierarchical, server-side update accumulation and global model merging are powerful tools for scalable, efficient training of new-era LLMs in heterogeneous, geo-distributed environments.
more »
« less
- Award ID(s):
- 2505865
- PAR ID:
- 10631424
- Publisher / Repository:
- https://doi.org/10.48550/arXiv.2506.04531
- Date Published:
- ISSN:
- 2506.04531
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Federated learning (FL) involves training a model over massive distributed devices, while keeping the training data localized and private. This form of collaborative learning exposes new tradeoffs among model convergence speed, model accuracy, balance across clients, and communication cost, with new challenges including: (1) straggler problem—where clients lag due to data or (computing and network) resource heterogeneity, and (2) communication bottleneck—where a large number of clients communicate their local updates to a central server and bottleneck the server. Many existing FL methods focus on optimizing along only one single dimension of the tradeoff space. Existing solutions use asynchronous model updating or tiering-based, synchronous mechanisms to tackle the straggler problem. However, asynchronous methods can easily create a communication bottleneck, while tiering may introduce biases that favor faster tiers with shorter response latencies. To address these issues, we present FedAT, a novel Federated learning system with Asynchronous Tiers under Non-i.i.d. training data. FedAT synergistically combines synchronous, intra-tier training and asynchronous, cross-tier training. By bridging the synchronous and asynchronous training through tiering, FedAT minimizes the straggler effect with improved convergence speed and test accuracy. FedAT uses a straggler-aware, weighted aggregation heuristic to steer and balance the training across clients for further accuracy improvement. FedAT compresses uplink and downlink communications using an efficient, polyline-encoding-based compression algorithm, which minimizes the communication cost. Results show that FedAT improves the prediction performance by up to 21.09% and reduces the communication cost by up to 8.5×, compared to state-of-the-art FL methods.more » « less
-
Data parallel frameworks become essential for training machine learning models. The classic Bulk Synchronous Parallel (BSP) model updates the model parameters through pre-defined synchronization barriers. However, when a worker computes significantly slower than other workers, waiting for the slow worker will lead to excessive waste of computing resources. In this paper, we propose a novel proactive data-parallel (PDP) framework. PDP enables the parameter server to initiate the update of the model parameter. That is, we can perform the update at any time without pre-defined update points. PDP not only initiates the update but also determines when to update. The global decision on the frequency of updates will accelerate the training. We further propose asynchronous PDP to reduce the idle time caused by synchronizing parameter updates. We theoretically prove the convergence property of asynchronous PDP. We implement a distributed PDP framework and evaluate PDP with several popular machine learning algorithms including Multilayer Perceptron, Convolutional Neural Network, K-means, and Gaussian Mixture Model. Our evaluation shows that PDP can achieve up to 20X speedup over the BSP model and scale to large clusters.more » « less
-
In this paper, we address the challenges of asynchronous gradient descent in distributed learning environments, particularly focusing on addressing the challenges of stale gradients and the need for extensive communication resources. We develop a novel communication efficient framework that incorporates a gradient evaluation algorithm to assess and utilize delayed gradients based on their quality, ensuring efficient and effective model updates while significantly reducing communication overhead. Our proposed algorithm requires agents to only send the norm of the gradients rather than the computed gradient. The server then decides whether to accept the gradient if the ratio between the norm of the gradient and the distance between the global model parameter and the local model parameter exceeds a certain threshold. With the proper choice of the threshold, we show that the convergence rate achieves the same order as the synchronous stochastic gradient without depending on the staleness value unlike most of the existing works. Given the computational complexity of the initial algorithm, we introduce a simplified variant that prioritizes the practical applicability without compromising on the convergence rates. Our simulations demonstrate that our proposed algorithms outperform existing state-of-the-art methods, offering improved convergence rates, stability, accuracy, and resource consumption.more » « less
-
Distributed Deep Neural Network (DDNN) training on cloud spot instances is increasingly compelling as it can significantly save the user budget. To handle unexpected instance revocations, provisioning a heterogeneous cluster using the asynchronous parallel mechanism becomes the dominant method for DDNN training with spot instances. However, blindly provisioning a cluster of spot instances can easily result in unpredictable DDNN training performance, mainly because bottlenecks occur on the parameter server network bandwidth and PCIe bandwidth resources, as well as the inadequate cluster heterogeneity. To address the challenges above, we propose spotDNN, a heterogeneity-aware spot instance provisioning framework that provides predictable performance for DDNN training in the cloud. By explicitly considering the contention for bottle-neck resources, we first build an analytical performance model of DDNN training in heterogeneous clusters. It leverages the weighted average batch size and convergence coefficient to quantify the DDNN training loss in heterogeneous clusters. Through a lightweight workload profiling, we further design a cost-efficient instance provisioning strategy which incorporates the bounds calculation and sliding window techniques to effectively guarantee the training performance service level objectives (SLOs). We have implemented a prototype of spotDNN and conducted extensive experiments on Amazon EC2. Experiment results show that spotDNN can deliver predictable DDNN training performance while reducing the monetary cost by up to 68:1% compared to the existing solutions, yet with acceptable runtime overhead.more » « less
