An official website of the United States government
Optimizing performance for Distributed Deep Neural Network (DDNN) training has recently become increasingly compelling, as the DNN model gets complex and the training dataset grows large. While existing works on communication scheduling mostly focus on overlapping the computation and communication to improve DDNN training performance, the GPU and network resources are still under-utilized in DDNN training clusters. To tackle this issue, in this paper, we design and implement a predictable communication scheduling strategy named Prophet to schedule the gradient transfer in an adequate order, with the aim of maximizing the GPU and network resource utilization. Leveraging our observed stepwise pattern of gradient transfer start time, Prophet first uses the monitored network bandwidth and the profiled time interval among gradients to predict the appropriate number of gradients that can be grouped into blocks. Then, these gradient blocks can be transferred one by one to guarantee high utilization of GPU and network resources while ensuring the priority of gradient transfer (i.e., low-priority gradients cannot preempt high-priority gradients in the network transfer). Prophet can make the forward propagation start as early as possible so as to greedily reduce the waiting (idle) time of GPU resources during the DDNN training process. Prototype experiments with representative DNN models trained on Amazon EC2 demonstrate that Prophet can improve the DDNN training performance by up to 40% compared with the state-of-theart priority-based communication scheduling strategies, yet with negligible runtime performance overhead.
more »
« less
Communication Algorithm-Architecture Co-Design for Distributed Deep Learning
Large-scale distributed deep learning training has enabled developments of more complex deep neural network models to learn from larger datasets for sophisticated tasks. In particular, distributed stochastic gradient descent intensively invokes all-reduce operations for gradient update, which dominates communication time during iterative training epochs. In this work, we identify the inefficiency in widely used all-reduce algorithms, and the opportunity of algorithm-architecture co-design. We propose MultiTree all-reduce algorithm with topology and resource utilization awareness for efficient and scalable all-reduce operations, which is applicable to different interconnect topologies. Moreover, we co-design the network interface to schedule and coordinate the all-reduce messages for contention-free communications, working in synergy with the algorithm. The flow control is also simplified to exploit the bulk data transfer of big gradient exchange. We evaluate the co-design using different all-reduce data sizes for synthetic study, demonstrating its effectiveness on various interconnection network topologies, in addition to state-of-the-art deep neural networks for real workload experiments. The results show that MultiTree achieves 2.3× and 1.56× communication speedup, as well as up to 81% and 30% training time reduction compared to ring all-reduce and state-of-the-art approaches, respectively.
more »
« less
- Award ID(s):
- 2135995
- PAR ID:
- 10374122
- Date Published:
- Journal Name:
- 2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA)
- Page Range / eLocation ID:
- 181 to 194
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Multi-Agent Reinforcement Learning (MARL) is a key technology in artificial intelligence applications such as robotics, surveillance, energy systems, etc. Multi-Agent Deep Deterministic Policy Gradient (MADDPG) is a state-of-the-art MARL algorithm that has been widely adopted and considered a popular baseline for novel MARL algorithms. However, existing implementations of MADDPG on CPU and CPU-GPU platforms do not exploit fine-grained parallelism between cooperative agents and handle inter-agent communication sequentially, leading to sub-optimal throughput performance in MADDPG training. In this work, we develop the first high-throughput MADDPG accelerator on a CPU-FPGA heterogeneous platform. Specifically, we develop dedicated hardware modules that enable parallel training of each agent's internal Deep Neural Networks (DNNs) and support low-latency inter-agent communication using an on-chip agent interconnection network. Our experimental results show that the speed performance of agent neural network training improves by a factor of 3.6×−24.3× and 1.5×−29.5× compared with state-of-the-art CPU and CPU-GPU implementations. Our design achieves up to a 1.99× and 1.93× improvement in overall system throughput compared with CPU and CPU-GPU implementations, respectively.more » « less
-
null (Ed.)Deep learning (DL) is a popular technique for building models from large quantities of data such as pictures, videos, messages generated from edges devices at rapid pace all over the world. It is often infeasible to migrate large quantities of data from the edges to centralized data center(s) over WANs for training due to privacy, cost, and performance reasons. At the same time, training large DL models on edge devices is infeasible due to their limited resources. An attractive alternative for DL training distributed data is to use micro-clouds---small-scale clouds deployed near edge devices in multiple locations. However, micro-clouds present the challenges of both computation and network resource heterogeneity as well as dynamism. In this paper, we introduce DLion, a new and generic decentralized distributed DL system designed to address the key challenges in micro-cloud environments, in order to reduce overall training time and improve model accuracy. We present three key techniques in DLion: (1) Weighted dynamic batching to maximize data parallelism for dealing with heterogeneous and dynamic compute capacity, (2) Per-link prioritized gradient exchange to reduce communication overhead for model updates based on available network capacity, and (3) Direct knowledge transfer to improve model accuracy by merging the best performing model parameters. We build a prototype of DLion on top of TensorFlow and show that DLion achieves up to 4.2X speedup in an Amazon GPU cluster, and up to 2X speed up and 26% higher model accuracy in a CPU cluster over four state-of-the-art distributed DL systems.more » « less
-
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 state-of-the-art industrial recommender model by 37%, without any loss in performance.more » « less
-
null (Ed.)Although the distributed machine learning methods can speed up the training of large deep neural networks, the communication cost has become the non-negligible bottleneck to constrain the performance. To address this challenge, the gradient compression based communication-efficient distributed learning methods were designed to reduce the communication cost, and more recently the local error feedback was incorporated to compensate for the corresponding performance loss. However, in this paper, we will show that a new "gradient mismatch" problem is raised by the local error feedback in centralized distributed training and can lead to degraded performance compared with full-precision training. To solve this critical problem, we propose two novel techniques, 1) step ahead and 2) error averaging, with rigorous theoretical analysis. Both our theoretical and empirical results show that our new methods can handle the "gradient mismatch" problem. The experimental results show that we can even train faster with common gradient compression schemes than both the full-precision training and local error feedback regarding the training epochs and without performance loss.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1109/ISCA52012.2021.00023
