Benchmarking Harp-DAAL: High Performance Hadoop on KNL Clusters

A survey on evaluating and optimizing performance of Intel Xeon Phi

“…They use "pencil"-shape tiles which have large stride-1 dimension for maximizing the efficacy of prefetching and are small in remaining two dimensions to easily fit the tiles in the small 1MB last-level cache of KNL which is shared between two cores. The typical size of tile and box are (64,4,4) and (64,64,64), respectively. Vertical kernels are more difficult to optimize, for example, the kernels having data-dependence across iterations cannot be vectorized.…”

“…This optimization is especially effective on Phi since its clock frequency is lower than that of CPU. Their technique reduces execution time by up to 10%.Chen et al64 interface HARP (a Hadoop-based communication library) with DAAL (a "data analytics acceleration library"). Their work replaces Java kernels of original HARP project with efficient native kernels from Intel's DAAL at the node level.…”

Exploring HPC and Big Data Convergence: A Graph Processing Study on Intel Knights Landing

“…As computation demand and volumes of data to be analyzed are constantly increasing, the lack of convergence is likely to become unsustainable: the costs of energy, computational, and human resources far exceed what most organizations can afford. The community has started to address the software convergence problem, discussing joint HPC-big data software-stacks [1], [2], [3], proposing techniques for software-integration [4], and augmenting big-data software-frameworks with efficient HPC libraries [1], [5], [6]. We envision a form of deep convergence, across both the software level and the infrastructure level.…”

“…In our vision, the deep convergence of HPC and big data entails defining common software and hardware ecosystems, catering to the needs of both worlds. This could enable: (1) efficient operation of software and hardware, which is currently available only to the HPC community, for a wideraudience including the big data community, (2) sustainable development of new solutions, addressing the human-resource scarcities affecting both communities, and (3) innovation, especially for more efficient software-and-hardware solutions, exploiting the full capability of modern hardware, and avoiding the consequences of lack of convergence already observed by practitioners [5], [6], [7].…”

“…To address the convergence problem, the community focuses already on porting big-data software to HPCinfrastructure [4], or, conversely, porting and integrating HPCsoftware into big-data infrastructure [1], [5], [6]. These alternatives pose research and engineering challenges related to portability: not only code has to be re-written for new hardware, but also intricate software parameter tuning and customization must be performed.…”