Search

Deep learning is a promising tool to determine the physical model that describes our universe. To handle the considerable computational cost of this problem, we present CosmoFlow: a highly scalable deep learning application built on top of the TensorFlow framework. CosmoFlow uses efficient implementations of 3D convolution and pooling primitives, together with improvements in threading for many element-wise operations, to improve training performance on Intel(C) Xeon Phi(TM) processors. We also utilize the Cray PE Machine Learning Plugin for efficient scaling to multiple nodes. We demonstrate fully synchronous data-parallel training on 8192 nodes of Cori with 77% parallel efficiency, achieving 3.5 Pflop/s sustained performance. To our knowledge, this is the first large-scale science application of the TensorFlow framework at supercomputer scale with fully-synchronous training. These enhancements enable us to process large 3D dark matter distribution and predict the cosmological parameters $\Omega_M$, $\sigma_8$ and n$_s$ with unprecedented accuracy., Comment: 11 pages, 6 pages, presented at SuperComputing 2018

QMCPACK is an open source quantum Monte Carlo package for ab-initio electronic structure calculations. It supports calculations of metallic and insulating solids, molecules, atoms, and some model Hamiltonians. Implemented real space quantum Monte Carlo algorithms include variational, diffusion, and reptation Monte Carlo. QMCPACK uses Slater-Jastrow type trial wave functions in conjunction with a sophisticated optimizer capable of optimizing tens of thousands of parameters. The orbital space auxiliary field quantum Monte Carlo method is also implemented, enabling cross validation between different highly accurate methods. The code is specifically optimized for calculations with large numbers of electrons on the latest high performance computing architectures, including multicore central processing unit (CPU) and graphical processing unit (GPU) systems. We detail the program's capabilities, outline its structure, and give examples of its use in current research calculations. The package is available at http://www.qmcpack.org .

We explore scaling of the standard distributed Tensorflow with GRPC primitives on up to 512 Intel Xeon Phi (KNL) nodes of Cori supercomputer with synchronous stochastic gradient descent (SGD), and identify causes of scaling inefficiency at higher node counts. To our knowledge, this is the first exploration of distributed GRPC Tensorflow scalability on a HPC supercomputer at such large scale with synchronous SGD. We studied scaling of two convolution neural networks - ResNet-50, a state-of-the-art deep network for classification with roughly 25.5 million parameters, and HEP-CNN, a shallow topology with less than 1 million parameters for common scientific usages. For ResNet-50, we achieve >80% scaling efficiency on up to 128 workers, using 32 parameter servers (PS tasks) with a steep decline down to 23% for 512 workers using 64 PS tasks. Our analysis of the efficiency drop points to low network bandwidth utilization due to combined effect of three factors. (a) Heterogeneous distributed parallelization algorithm which uses PS tasks as centralized servers for gradient averaging is suboptimal for utilizing interconnect bandwidth. (b) Load imbalance among PS tasks hinders their efficient scaling. (c) Underlying communication primitive GRPC is currently inefficient on Cori high-speed interconnect. The HEP-CNN demands less interconnect bandwidth, and shows >80% weak scaling efficiency for up to 256 nodes with only 1 PS task. Our findings are applicable to other deep learning networks. Big networks with millions of parameters stumble upon the issues discussed here. Shallower networks like HEP-CNN with relatively lower number of parameters can efficiently enjoy weak scaling even with a single parameter server., Comment: Published as a poster in NIPS 2017 Workshop: Deep Learning At Supercomputer Scale

QMCPACK has enabled cutting-edge materials research on supercomputers for over a decade. It scales nearly ideally but has low single-node efficiency due to the physics-based abstractions using array-of-structures objects, causing inefficient vectorization. We present a systematic approach to transform QMCPACK to better exploit the new hardware features of modern CPUs in portable and maintainable ways. We develop miniapps for fast prototyping and optimizations. We implement new containers in structure-of-arrays data layout to facilitate vectorizations by the compilers. Further speedup and smaller memory-footprints are obtained by computing data on the fly with the vectorized routines and expanding single-precision use. All these are seamlessly incorporated in production QMCPACK. We demonstrate upto 4.5x speedups on recent Intel processors and IBM Blue Gene/Q for representative workloads. Energy consumption is reduced significantly commensurate to the speedup factor. Memory-footprints are reduced by up-to 3.8x, opening the possibility to solve much larger problems of future., Comment: 12 pages, 10 figures, 2 tables, to be published at SC17

B-spline based orbital representations are widely used in Quantum Monte Carlo (QMC) simulations of solids, historically taking as much as 50% of the total run time. Random accesses to a large four-dimensional array make it challenging to efficiently utilize caches and wide vector units of modern CPUs. We present node-level optimizations of B-spline evaluations on multi/many-core shared memory processors. To increase SIMD efficiency and bandwidth utilization, we first apply data layout transformation from array-of-structures to structure-of-arrays (SoA). Then by blocking SoA objects, we optimize cache reuse and get sustained throughput for a range of problem sizes. We implement efficient nested threading in B-spline orbital evaluation kernels, paving the way towards enabling strong scaling of QMC simulations. These optimizations are portable on four distinct cache-coherent architectures and result in up to 5.6x performance enhancements on Intel Xeon Phi processor 7250P (KNL), 5.7x on Intel Xeon Phi coprocessor 7120P, 10x on an Intel Xeon processor E5v4 CPU and 9.5x on BlueGene/Q processor. Our nested threading implementation shows nearly ideal parallel efficiency on KNL up to 16 threads. We employ roofline performance analysis to model the impacts of our optimizations. This work combined with our current efforts of optimizing other QMC kernels, result in greater than 4.5x speedup of miniQMC on KNL., Comment: 11 pages, 10 figures, 4 tables