Training neural networks has become a big bottleneck. For example, training ImageNet dataset on one NVIDIA K20 GPU needs 21 days. To speed up the training process, the current deep learning systems heavily rely on the hardware accelerators. However, these accelerators have limited on-chip memory compared with CPUs. We use both self-host Intel Knights Landing (KNL) clusters and multi-GPU clusters as our target platforms. From the algorithm aspect, we focus on Elastic Averaging SGD (EASGD) to design algorithms for HPC clusters. We redesign four efficient algorithms for HPC systems to improve EASGD's poor scaling on clusters. Async EASGD, Async MEASGD, and Hogwild EASGD are faster than existing counterpart methods (Async SGD, Async MSGD, and Hogwild SGD) in all comparisons. Sync EASGD achieves 5.3X speedup over original EASGD on the same platform. We achieve 91.5% weak scaling efficiency on 4253 KNL cores, which is higher than the state-of-the-art implementation.

The nature of dark energy and the complete theory of gravity are two central questions currently facing cosmology. A vital tool for addressing them is the 3-point correlation function (3PCF), which probes deviations from a spatially random distribution of galaxies. However, the 3PCF's formidable computational expense has prevented its application to astronomical surveys comprising millions to billions of galaxies. We present Galactos, a high-performance implementation of a novel, O(N2) algorithm that uses a load-balanced k-d tree and spherical harmonic expansions to compute the anisotropic 3PCF. Our implementation is optimized for the Intel Xeon Phi architecture, exploiting SIMD parallelism, instruction and thread concurrency, and significant L1 and L2 cache reuse, reaching 39% of peak performance on a single node. Galactos scales to the full Cori system, achieving 9.8 PF (peak) and 5.06 PF (sustained) across 9636 nodes, making the 3PCF easily computable for all galaxies in the observable universe.

With the appearance of the heterogeneous platform Open- Power, many-core accelerator devices have been coupled with Power host processors for the first time. Towards utilizing their full potential, it is worth investigating performance portable algorithms that allow to choose the best-fitting hardware for each domain-specific compute task. Suiting even the high level of parallelism on modern GPGPUs, our presented approach relies heavily on abstract meta-programming techniques, which are essential to focus on fine-grained tuning rather than code porting. With this in mind, the CUDA-based open-source plasma simulation code PIConGPU is currently being abstracted to support the heterogeneous OpenPower platform using our fast porting interface cupla, which wraps the abstract parallel C++11 kernel acceleration library Alpaka. We demonstrate how PIConGPU can benefit from the tunable kernel execution strategies of the Alpaka library, achieving portability and performance with single-source kernels on conventional CPUs, Power8 CPUs and NVIDIA GPUs.

The Roofline Performance Model is a visually intuitive method used to bound the sustained peak floating-point performance of any given arithmetic kernel on any given processor architecture. In the Roofline, performance is nominally measured in floating-point operations per second as a function of arithmetic intensity (operations per byte of data). In this study we determine the Roofline for the Intel Knights Landing (KNL) processor, determining the sustained peak memory bandwidth and floating-point performance for all levels of the memory hierarchy, in all the different KNL cluster modes.We then determine arithmetic intensity and performance for a suite of application kernels being targeted for the KNL based supercomputer Cori, and make comparisons to current Intel Xeon processors. Cori is the National Energy Research Scientific Computing Center’s (NERSC) next generation supercomputer. Scheduled for deployment mid-2016, it will be one of the earliest and largest KNL deployments in the world.