skip to main content

Title: Training Recommender Systems at Scale: Communication-Efficient Model and Data Parallelism
In this paper, we consider hybrid parallelism—a paradigm that em- ploys both Data Parallelism (DP) and Model Parallelism (MP)—to scale distributed training of large recommendation models. We propose a compression framework called Dynamic Communication Thresholding (DCT) for communication-efficient hybrid training. DCT filters the entities to be communicated across the network through a simple hard-thresholding function, allowing only the most relevant information to pass through. For communication efficient DP, DCT compresses the parameter gradients sent to the parameter server during model synchronization. The threshold is updated only once every few thousand iterations to reduce the computational overhead of compression. For communication efficient MP, DCT incorporates a novel technique to compress the activations and gradients sent across the network during the forward and backward propagation, respectively. This is done by identifying and updating only the most relevant neurons of the neural network for each training sample in the data. We evaluate DCT on publicly available natural language processing and recommender models and datasets, as well as recommendation systems used in production at Facebook. DCT reduces communication by at least 100× and 20× during DP and MP, respectively. The algorithm has been deployed in production, and it improves end-to-end training time for a more » state-of-the-art industrial recommender model by 37%, without any loss in performance. « less
Authors:
; ; ; ; ; ; ; ;
Award ID(s):
1703678
Publication Date:
NSF-PAR ID:
10327111
Journal Name:
KDD '21: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining
Page Range or eLocation-ID:
2928 to 2936
Sponsoring Org:
National Science Foundation
More Like this
  1. Promising for digital signal processing applications, approximate computing has been extensively considered to tradeoff limited accuracy for improvements in other circuit metrics such as area, power, and performance. In this paper, approximate arithmetic circuits are proposed by using emerging nanoscale spintronic devices. Leveraging the intrinsic current-mode thresholding operation of spintronic devices, we initially present a hybrid spin-CMOS majority gate design based on a composite spintronic device structure consisting of a magnetic domain wall motion stripe and a magnetic tunnel junction. We further propose a compact and energy-efficient accuracy-configurable adder design based on the majority gate. Unlike most previous approximate circuit designs that hardwire a constant degree of approximation, this design is adaptive to the inherent resilience in various applications to different degrees of accuracy. Subsequently, we propose two new approximate compressors for utilization in fast multiplier designs. The device-circuit SPICE simulation shows 34.58% and 66% improvement in power consumption, respectively, for the accurate and approximate modes of the accuracy-configurable adder, compared to the recently reported domain wall motion-based full adder design. In addition, the proposed accuracy-configurable adder and approximate compressors can be efficiently utilized in the discrete cosine transform (DCT) as a widely-used digital image processing algorithm. The results indicatemore »that the DCT and inverse DCT (IDCT) using the approximate multiplier achieve ~2x energy saving and 3x speed-up compared to an exactly-designed circuit, while achieving comparable quality in its output result.« less
  2. Model parameter synchronization across GPUs introduces high overheads for data-parallel training at scale. Existing parameter synchronization protocols cannot effectively leverage available network resources in the face of ever increasing hardware heterogeneity. To address this issue, we propose Blink, a collective communication library that dynamically generates optimal communication primitives by packing spanning trees. We propose techniques to minimize the number of trees generated and extend Blink to leverage heterogeneous communication channels for hybrid, and faster, data transfers. Evaluations show that compared to the state-of-the-art (NCCL), Blink can achieve up to 8× faster model synchronization (AllReduce), and reduce end-to-end DNN training time for image classification tasks by up to 40%.
  3. Distributed model training suffers from communication bottlenecks due to frequent model updates transmitted across compute nodes. To alleviate these bottlenecks, practitioners use gradient compression techniques like sparsification, quantization, or low-rank updates. The techniques usually require choosing a static compression ratio, often requiring users to balance the trade-off between model accuracy and per-iteration speedup. In this work, we show that such performance degradation due to choosing a high compression ratio is not fundamental. An adaptive compression strategy can reduce communication while maintaining final test accuracy. Inspired by recent findings on critical learning regimes, in which small gradient errors can have irrecoverable impact on model performance, we propose Accordion a simple yet effective adaptive compression algorithm. While Accordion maintains a high enough compression rate on average, it avoids over-compressing gradients whenever in critical learning regimes, detected by a simple gradient-norm based criterion. Our extensive experimental study over a number of machine learning tasks in distributed environments indicates that Accordion, maintains similar model accuracy to uncompressed training, yet achieves up to 5.5x better compression and up to 4.1x end-to-end speedup over static approaches. We show that Accordion also works for adjusting the batch size, another popular strategy for alleviating communication bottlenecks.
  4. Smola, A. ; Dimakis, A. ; Stoica, I. (Ed.)
    Distributed model training suffers from communication bottlenecks due to frequent model updates transmitted across compute nodes. To alleviate these bottlenecks, practitioners use gradient compression techniques like sparsification, quantization, low rank updates etc. The techniques usually require choosing a static compression ratio, often requiring users to balance the trade-off between model accuracy and per-iteration speedup. In this work, we show that such performance degradation due to choosing a high compression ratio is not fundamental and that an adaptive compression strategy can reduce communication while maintaining final test accuracy.Inspired by recent findings on critical learning regimes, in which small gradient errors can have irrecoverable impact on model performance, we propose ACCORDION a simple yet effective adaptive compression algorithm. While ACCORDION maintains a high enough compression rate on average, it avoids detrimental impact by not compressing gradients too much whenever in critical learning regimes, detected by a simple gradient-norm based criterion. Our extensive experimental study over a number of machine learning tasks in distributed environments indicates that ACCORDION, maintains similar model accuracy to uncompressed training, yet achieves up to 5.5×better compression and up to 4.1×end-to-end speedup over static approaches. We show that ACCORDION also works for adjusting the batch size, another popular strategymore »for alleviating communication bottlenecks. Our code is available at https://github.com/uw-mad-dash/Accordion« less
  5. In this paper, we propose and analyze SPARQSGD, an event-triggered and compressed algorithm for decentralized training of large-scale machine learning models over a graph. Each node can locally compute a condition (event) which triggers a communication where quantized and sparsified local model parameters are sent. In SPARQ-SGD, each node first takes a fixed number of local gradient steps and then checks if the model parameters have significantly changed compared to its last update; it communicates further compressed model parameters only when there is a significant change, as specified by a (design) criterion. We prove that SPARQ-SGD converges as O(1/nT ) and O(1/√nT ) in the strongly-convex and non-convex settings, respectively, demonstrating that aggressive compression, including event-triggered communication, model sparsification and quantization does not affect the overall convergence rate compared to uncompressed decentralized training; thereby theoretically yielding communication efficiency for `free'. We evaluate SPARQ-SGD over real datasets to demonstrate significant savings in communication bits over the state-of-the-art.