اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدWhale: Scaling Deep Learning Model Training to the Trillions
51
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Scaling up deep neural networks has been proven effective in improving model quality, while it also brings ever-growing training challenges. This paper presents Whale, an automatic and hardware-aware distributed training framework for giant models. Whale generalizes the expression of parallelism with four primitives, which can define various parallel strategies, as well as flexible hybrid strategies including combination and nesting patterns. It allows users to build models at an arbitrary scale by adding a few annotations and automatically transforms the local model to a distributed implementation. Moreover, Whale is hardware-aware and highly efficient even when training on GPUs of mixed types, which meets the growing demand of heterogeneous training in industrial clusters. Whale sets a milestone for training the largest multimodal pretrained model M6. The success of M6 is achieved by Whales design to decouple algorithm modeling from system implementations, i.e., algorithm developers can focus on model innovation, since it takes only three lines of code to scale the M6 model to trillions of parameters on a cluster of 480 GPUs.
قيم البحث
اقرأ أيضاً
We present scalable hybrid-parallel algorithms for training large-scale 3D convolutional neural networks. Deep learning-based emerging scientific workflows often require model training with large, high-dimensional samples, which can make training muc
h more costly and even infeasible due to excessive memory usage. We solve these challenges by extensively applying hybrid parallelism throughout the end-to-end training pipeline, including both computations and I/O. Our hybrid-parallel algorithm extends the standard data parallelism with spatial parallelism, which partitions a single sample in the spatial domain, realizing strong scaling beyond the mini-batch dimension with a larger aggregated memory capacity. We evaluate our proposed training algorithms with two challenging 3D CNNs, CosmoFlow and 3D U-Net. Our comprehensive performance studies show that good weak and strong scaling can be achieved for both networks using up 2K GPUs. More importantly, we enable training of CosmoFlow with much larger samples than previously possible, realizing an order-of-magnitude improvement in prediction accuracy.
The increased use of deep learning (DL) in academia, government and industry has, in turn, led to the popularity of on-premise and cloud-hosted deep learning platforms, whose goals are to enable organizations utilize expensive resources effectively,
and to share said resources among multiple teams in a fair and effective manner. In this paper, we examine the elastic scaling of Deep Learning (DL) jobs over large-scale training platforms and propose a novel resource allocation strategy for DL training jobs, resulting in improved job run time performance as well as increased cluster utilization. We begin by analyzing DL workloads and exploit the fact that DL jobs can be run with a range of batch sizes without affecting their final accuracy. We formulate an optimization problem that explores a dynamic batch size allocation to individual DL jobs based on their scaling efficiency, when running on multiple nodes. We design a fast dynamic programming based optimizer to solve this problem in real-time to determine jobs that can be scaled up/down, and use this optimizer in an autoscaler to dynamically change the allocated resources and batch sizes of individual DL jobs. We demonstrate empirically that our elastic scaling algorithm can complete up to $approx 2 times$ as many jobs as compared to a strong baseline algorithm that also scales the number of GPUs but does not change the batch size. We also demonstrate that the average completion time with our algorithm is up to $approx 10 times$ faster than that of the baseline.
Distributed deep learning systems (DDLS) train deep neural network models by utilizing the distributed resources of a cluster. Developers of DDLS are required to make many decisions to process their particular workloads in their chosen environment ef
ficiently. The advent of GPU-based deep learning, the ever-increasing size of datasets and deep neural network models, in combination with the bandwidth constraints that exist in cluster environments require developers of DDLS to be innovative in order to train high quality models quickly. Comparing DDLS side-by-side is difficult due to their extensive feature lists and architectural deviations. We aim to shine some light on the fundamental principles that are at work when training deep neural networks in a cluster of independent machines by analyzing the general properties associated with training deep learning models and how such workloads can be distributed in a cluster to achieve collaborative model training. Thereby we provide an overview of the different techniques that are used by contemporary DDLS and discuss their influence and implications on the training process. To conceptualize and compare DDLS, we group different techniques into categories, thus establishing a taxonomy of distributed deep learning systems.
Stochastic gradient descent (SGD) is an inherently sequential training algorithm--computing the gradient at batch $i$ depends on the model parameters learned from batch $i-1$. Prior approaches that break this dependence do not honor them (e.g., sum t
he gradients for each batch, which is not what sequential SGD would do) and thus potentially suffer from poor convergence. This paper introduces a novel method to combine gradients called Adasum (for adaptive sum) that converges faster than prior work. Adasum is easy to implement, almost as efficient as simply summing gradients, and is integrated into the open-source toolkit Horovod. This paper first provides a formal justification for Adasum and then empirically demonstrates Adasum is more accurate than prior gradient accumulation methods. It then introduces a series of case-studies to show Adasum works with multiple frameworks, (TensorFlow and PyTorch), scales multiple optimizers (Momentum-SGD, Adam, and LAMB) to larger batch-sizes while still giving good downstream accuracy. Finally, it proves that Adasum converges. To summarize, Adasum scales Momentum-SGD on the MLPerf Resnet50 benchmark to 64K examples before communication (no MLPerf v0.5 entry converged with more than 16K), the Adam optimizer to 64K examples before communication on BERT-LARGE (prior work showed Adam stopped scaling at 16K), and the LAMB optimizer to 128K before communication on BERT-LARGE (prior work used 64K), all while maintaining downstream accuracy metrics. Finally, if a user does not need to scale, we show LAMB with Adasum on BERT-LARGE converges in 30% fewer steps than the baseline.
During the last two years, the goal of many researchers has been to squeeze the last bit of performance out of HPC system for AI tasks. Often this discussion is held in the context of how fast ResNet50 can be trained. Unfortunately, ResNet50 is no lo
nger a representative workload in 2020. Thus, we focus on Recommender Systems which account for most of the AI cycles in cloud computing centers. More specifically, we focus on Facebooks DLRM benchmark. By enabling it to run on latest CPU hardware and software tailored for HPC, we are able to achieve more than two-orders of magnitude improvement in performance (110x) on a single socket compared to the reference CPU implementation, and high scaling efficiency up to 64 sockets, while fitting ultra-large datasets. This paper discusses the optimization techniques for the various operators in DLRM and which component of the systems are stressed by these different operators. The presented techniques are applicable to a broader set of DL workloads that pose the same scaling challenges/characteristics as DLRM.
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
