Deep neural networks (DNNs) continue to grow rapidly in size, making them infeasible to train on a single device (e.g. GPU). Pipeline parallelism is commonly used in existing DNN systems to support large-scale DNN training by partitioning a DNN into multiple stages, which concurrently perform DNN computation for different micro-batches of training samples in a pipeline fashion. However, existing pipeline-parallel approaches only consider sequential pipeline stages and thus ignore the topology of a DNN, resulting in missed model-parallel opportunities.

The strong demand for efficient and performant deployment of Deep Learning (DL) applications prompts the rapid development of a rich DL ecosystem. To keep up with this fast advancement, it is crucial for modern DL frameworks to efficiently integrate a variety of optimized tensor algebra libraries and runtimes as their backends and generate the fastest possible executable using these backends. However, current DL frameworks require significant manual effort and expertise to integrate every new backend while failing to unleash its full potential. Given the fast-evolving nature of the DL ecosystem, this manual approach often slows down continuous innovations across different layers; it prevents hardware vendors from the fast deployment of their cutting-edge libraries, DL framework developers must repeatedly adjust their hand-coded rules to accommodate new versions of libraries, and machine learning practitioners need to wait for the integration of new technologies and often encounter unsatisfactory performance.

Despite ceaseless efforts, extremely large and complex optimization space makes even the state-of-the-art compilers fail in delivering the most performant setting that can fully utilize the underlying hardware. Although this inefficiency suggests opportunity for tuning, it has been challenging for prior tuning methods to consider the complex interactions between optimizations and maximize the tuning quality while handling local optima efficiently. To tackle this problem, we suggest an intelligent auto-tuning strategy, called SRTuner, which searches for the best optimization setting by exposing important optimization interactions and directly using them to focus on promising subspaces. To reveal high-impact inter-optimization relations, SRTuner proposes a multistage structure and a distribution-based estimation method that approximates the impact of an optimization effectively. Besides, to efficiently handle local optima, our technique defines optimization decisions as a series of multi-armed bandit problems to formulate the exploration-exploitation dilemma. SRTuner is evaluated with three representative compilers from various domains on different target hardware: GCC (traditional C/ C++ compiler) on CPU, TVM (domain-specific machine learning compiler) on GPU, and OpenCL compilers (kernel compiler for heterogeneous computing) on both CPU/GPU. Results show that SRTuner accelerates target executions by $1. 24 \times$, $2. 03 \times$ and $34. 4 \times$ compared to the highest level of optimization provided by each compiler and outperforms state-of-the-art works by $1. 04 \times - 1. 14 \times$. As a byproduct of our unique tuning strategy, SRTuner can offer synergistic optimizations for each workload, which allows it to in part identify why it outperformed current compilers. With this information, we are able to find important optimizations that each compiler misused and demonstrate how this information can benefit future tuning strategies.

Yohan Jo, Shivani Poddar, Byungsoo Jeon, Qinlan Shen, Carolyn Rosé, Graham Neubig. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Many real-world data are naturally represented as tensors, or multi-dimensional arrays. Tensor decomposition is an important tool to analyze tensors for various applications such as latent concept discovery, trend analysis, clustering, and anomaly detection. However, existing tools for tensor analysis do not scale well for billion-scale tensors or offer limited functionalities. In this paper, we propose BIGtensor, a large-scale tensor mining library that tackles both of the above problems. Carefully designed for scalability, BIGtensor decomposes at least 100× larger tensors than the current state of the art. Furthermore, BIGtensor provides a variety of distributed tensor operations and tensor generation methods. We demonstrate how BIGtensor can help users discover hidden concepts and analyze trends from large-scale tensors that are hard to be processed by existing tensor tools.

How can we analyze very large real-world tensors where additional information is coupled with certain modes of tensors? Coupled matrix-tensor factorization is a useful tool to simultaneously analyze matrices and a tensor, and has been used for important applications including collaborative filtering, multi-way clustering, and link prediction. However, existing single machine or distributed algorithms for coupled matrix-tensor factorization do not scale for tensors with billions of elements in each mode. In this paper, we propose SCOUT, a large-scale coupled matrix-tensor factorization algorithm running on the distributed MAPREDUCE platform. By carefully reorganizing operations, and reusing intermediate data, SCOUT decomposes up to 100× larger tensors than existing methods, and shows linear scalability for order and machines while other methods are limited in scalability. We also apply SCOUT on real world tensors and discover interesting hidden patterns like seasonal spike, and steady attentions for healthy food on Yelp dataset containing user-business-yearmonth tensor and two coupled matrices.