Parallelization framework has become a necessity to speed up the training of deep neural networks (DNN) recently. Such framework typically employs the Model Average approach, denoted as MA-DNN, in which parallel workers conduct respective training based on their own local data while the parameters of local models are periodically communicated and averaged to obtain a global model which serves as the new start of local models. However, since DNN is a highly non-convex model, averaging parameters cannot ensure that such global model can perform better than those local models. To tackle this problem, we introduce a new parallel training framework called Ensemble-Compression, denoted as EC-DNN. In this framework, we propose to aggregate the local models by ensemble, i.e., averaging the outputs of local models instead of the parameters. As most of prevalent loss functions are convex to the output of DNN, the performance of ensemble-based global model is guaranteed to be at least as good as the average performance of local models. However, a big challenge lies in the explosion of model size since each round of ensemble can give rise to multiple times size increment. Thus, we carry out model compression after each ensemble, specialized by a distillation based method in this paper, to reduce the size of the global model to be the same as the local ones. Our experimental results demonstrate the prominent advantage of EC-DNN over MA-DNN in terms of both accuracy and speedup.
Yonsei University（延世大学）；Seoul National University（首尔国立大学）；Samsung Electronics（三星电子）
In training of modern large natural language processing (NLP) models, it has become a common practice to split models using 3D parallelism to multiple GPUs. Such technique, however, suffers from a high overhead of inter-node communication. Compressing the communication is one way to mitigate the overhead by reducing the inter-node traffic volume; however, the existing compression techniques have critical limitations to be applied for NLP models with 3D parallelism in that 1) only the data parallelism traffic is targeted, and 2) the existing compression schemes already harm the model quality to…
Lossy compression is one of the most efficient solutions to reduce storage overhead and improve I/O performance for HPC applications. However, existing parallel I/O libraries cannot fully utilize lossy compression to accelerate parallel write due to the lack of deep understanding on compression-write performance. To this end, we propose to deeply integrate predictive lossy compression with HDF5 to significantly improve the parallel-write performance. Specifically, we propose analytical models to predict the time of compression and parallel write before the actual compression to enable compress…
Many scientific data sets contain temporal dimensions. These are the data storing information at the same spatial location but different time stamps. Some of the biggest temporal datasets are produced by parallel computing applications such as simulations of climate change and fluid dynamics. Temporal datasets can be very large and cost a huge amount of time to transfer among storage locations. Using data compression techniques, files can be transferred faster and save storage space. NUMARCK is a lossy data compression algorithm for temporal data sets that can learn emerging distributions of e…
A rich body of prior work has highlighted the existence of communication bottlenecks in synchronous data-parallel training. To alleviate these bottlenecks, a long line of recent work proposes gradient and model compression methods. In this work, we evaluate the efficacy of gradient compression methods and compare their scalability with optimized implementations of synchronous data-parallel SGD across more than 200 different setups. Surprisingly, we observe that only in 6 cases out of more than 200, gradient compression methods provide speedup over optimized synchronous data-parallel training i…
King Fahd University of Petroleum & Minerals（法赫德国王石油与矿业大学）；Tilburg University（蒂尔堡大学）；Banglalink Digital Communications Limited（Banglalink数字通信有限公司）
In this paper, we implemented both sequential and parallel version of fractal image compression algorithms using CUDA (Compute Unified Device Architecture) programming model for parallelizing the program in Graphics Processing Unit for medical images, as they are highly similar within the image itself. There are several improvement in the implementation of the algorithm as well. Fractal image compression is based on the self similarity of an image, meaning an image having similarity in majority of the regions. We take this opportunity to implement the compression algorithm and monitor the effe…
Today's exponentially increasing data volumes and the high cost of storage make compression essential for the Big Data industry. Although research has concentrated on efficient compression, fast decompression is critical for analytics queries that repeatedly read compressed data. While decompression can be parallelized somewhat by assigning each data block to a different process, break-through speed-ups require exploiting the massive parallelism of modern multi-core processors and GPUs for data decompression within a block. We propose two new techniques to increase the degree of parallelism du…
Zhejiang University of Science and Technology（浙江科技学院）；Tsinghua University（清华大学）
Inferring the causal direction between two variables from their observation data is one of the most fundamental and challenging topics in data science. A causal direction inference algorithm maps the observation data into a binary value which represents either x causes y or y causes x. The nature of these algorithms makes the results unstable with the change of data points. Therefore the accuracy of the causal direction inference can be improved significantly by using parallel ensemble frameworks. In this paper, new causal direction inference algorithms based on several ways of parallel ensemb…
University of Rochester（罗切斯特大学）；Hong Kong University of Science and Technology（香港科技大学）；Kwai Inc（快手公司）
A standard approach in large scale machine learning is distributed stochastic gradient training, which requires the computation of aggregated stochastic gradients over multiple nodes on a network. Communication is a major bottleneck in such applications, and in recent years, compressed stochastic gradient methods such as QSGD (quantized SGD) and sparse SGD have been proposed to reduce communication. It was also shown that error compensation can be combined with compression to achieve better convergence in a scheme that each node compresses its local stochastic gradient and broadcast the result…
For parallel breadth first search (BFS) algorithm on large-scale distributed memory systems, communication often costs significantly more than arithmetic and limits the scalability of the algorithm. In this paper we sufficiently reduce the communication cost in distributed BFS by compressing and sieving the messages. First, we leverage a bitmap compression algorithm to reduce the size of messages before communication. Second, we propose a novel distributed directory algorithm, cross directory, to sieve the redundant data in messages. Experiments on a 6,144-core SMP cluster show our algorithm o…
TACO: Efficient Communication Compression of Intermediate Tensors for Scalable Tensor-Parallel LLM Training
Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences（中国科学院大学杭州高等研究院）；Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）；University of Science and Technology of China（中国科学技术大学）
ETH Zürich（苏黎世联邦理工学院）；Stanford University（斯坦福大学）；Zhejiang University（浙江大学）；Carnegie Mellon University（卡内基梅隆大学）；Meta AI
Communication compression is a crucial technique for modern distributed learning systems to alleviate their communication bottlenecks over slower networks. Despite recent intensive studies of gradient compression for data parallel-style training, compressing the activations for models trained with pipeline parallelism is still an open problem. In this paper, we propose AC-SGD, a novel activation compression algorithm for communication-efficient pipeline parallelism training over slow networks. Different from previous efforts in activation compression, instead of compressing activation values d…
Large scale simulations of complex systems ranging from climate and astrophysics to crowd dynamics, produce routinely petabytes of data and are projected to reach the zettabytes level in the coming decade. These simulations enable unprecedented insights but at the same their effectiveness is hindered by the enormous data sizes associated with the computational elements and respective output quantities of interest that impose severe constraints on storage and I/O time. In this work, we address these challenges through a novel software framework for scientific data compression. The software (Cub…
Graphics Processing Units allow for running massively parallel applications offloading the CPU from computationally intensive resources, however GPUs have a limited amount of memory. In this paper a trie compression algorithm for massively parallel pattern matching is presented demonstrating 85% less space requirements than the original highly efficient parallel failure-less aho-corasick, whilst demonstrating over 22 Gbps throughput. The algorithm presented takes advantage of compressed row storage matrices as well as shared and texture memory on the GPU.
As HPC systems continue to grow to exascale, the amount of data that needs to be saved or transmitted is exploding. To this end, many previous works have studied using error-bounded lossy compressors to reduce the data size and improve the I/O performance. However, little work has been done for effectively offloading lossy compression onto FPGA-based SmartNICs to reduce the compression overhead. In this paper, we propose a hardware-algorithm codesign for an efficient and adaptive lossy compressor for scientific data on FPGAs (called CEAZ), which is the first lossy compressor that can achieve h…
Communication overhead is the key challenge for distributed training. Gradient compression is a widely used approach to reduce communication traffic. When combining with parallel communication mechanism method like pipeline, gradient compression technique can greatly alleviate the impact of communication overhead. However, there exists two problems of gradient compression technique to be solved. Firstly, gradient compression brings in extra computation cost, which will delay the next training iteration. Secondly, gradient compression usually leads to the decrease of convergence accuracy.
Nvidia Corporation（英伟达公司）；University of Waikato（怀卡托大学）
SHAP (SHapley Additive exPlanation) values provide a game theoretic interpretation of the predictions of machine learning models based on Shapley values. While exact calculation of SHAP values is computationally intractable in general, a recursive polynomial-time algorithm called TreeShap is available for decision tree models. However, despite its polynomial time complexity, TreeShap can become a significant bottleneck in practical machine learning pipelines when applied to large decision tree ensembles. Unfortunately, the complicated TreeShap algorithm is difficult to map to hardware accelera…
Recently, Czumaj et.al. (arXiv 2017) presented a parallel (almost) $2$-approximation algorithm for the maximum matching problem in only $O({(\log\log{n})^2})$ rounds of the massive parallel computation (MPC) framework, when the memory per machine is $O(n)$. The main approach in their work is a way of compressing $O(\log{n})$ rounds of a distributed algorithm for maximum matching into only $O({(\log\log{n})^2})$ MPC rounds. In this note, we present a similar algorithm for the closely related problem of approximating the minimum vertex cover in the MPC framework. We show that one can achieve an …
The Massive Parallel Computation (MPC) model is a theoretical framework for popular parallel and distributed platforms such as MapReduce, Hadoop, or Spark. We consider the task of computing a large matching or small vertex cover in this model when the space per machine is $n^δ$ for $δ\in (0,1)$, where $n$ is the number of vertices in the input graph. A direct simulation of classic PRAM and distributed algorithms from the 1980s results in algorithms that require at least a logarithmic number of MPC rounds. We give the first algorithm that breaks this logarithmic barrier and runs in $\tilde O(\s…
Decompressing a file made by the gzip program at an arbitrary location is in principle impossible, due to the nature of the DEFLATE compression algorithm. Consequently, no existing program can take advantage of parallelism to rapidly decompress large gzip-compressed files. This is an unsatisfactory bottleneck, especially for the analysis of large sequencing data experiments. Here we propose a parallel algorithm and an implementation, pugz, that performs fast and exact decompression of any text file. We show that pugz is an order of magnitude faster than gunzip, and 5x faster than a highly-opti…
For over a decade now we have been witnessing the success of {\em massive parallel computation} (MPC) frameworks, such as MapReduce, Hadoop, Dryad, or Spark. One of the reasons for their success is the fact that these frameworks are able to accurately capture the nature of large-scale computation. In particular, compared to the classic distributed algorithms or PRAM models, these frameworks allow for much more local computation. The fundamental question that arises in this context is though: can we leverage this additional power to obtain even faster parallel algorithms? A prominent example he…
