February 2016 Summaries

In React development, an elegant approach to creating flexible component APIs is using functions as children, which allows for a more dynamic and maintainable structure. The example of a Dropdown component illustrates this method by exposing the toggle functionality directly to its children through a function, eliminating the need for manual event listener setup on DOM elements. This approach leverages React's event system and enhances the component's flexibility, exemplified by the dropdown's ability to toggle the 'opened' CSS class using a function passed as a child. The component's implementation involves using the 'classnames' package for managing class names and the 'react-click-outside' package to handle clicks outside the component, demonstrating a practical application of this pattern in real-world scenarios.

In the process of developing web applications with Django, developers often start by placing all models and business logic into a single app due to the simplicity of initial projects. However, as a business grows, this approach can lead to difficulties in debugging and expanding features. The blog discusses the benefits of reorganizing code into multiple Django apps, each responsible for specific domains, to improve maintainability and scalability. A case study involving a blogging application illustrates the transition from a monolithic structure to a more modular one, where billing-related features are moved to a separate app. This reorganization, while initially requiring more effort, results in clearer separation of concerns, easier debugging, and a more manageable codebase. Additionally, it simplifies onboarding for new developers and allows for easier deprecation of specific business logic. Overall, the piece highlights the importance of periodically reassessing and restructuring the application architecture to align with business growth.

A webinar scheduled for March 2, 2016, aims to showcase the benefits of using Rescale's on-demand, high-performance computing (HPC) cloud platform for SHIPFLOW users involved in computational fluid dynamics (CFD) for ship hydrodynamic design. Hosted by Leif Broberg from FLOWTECH and Sarah Dietz and Hiraku Nakamura from Rescale, alongside Magnus Östberg from FLOWTECH, the session will address challenges such as limited local computing resources and tight deadlines that often restrict CFD project scopes. By leveraging cloud resources, both large and small computations, including hull shape optimization and parameter studies, can be enhanced. The webinar will cover how to run SHIPFLOW directly on the Rescale platform and conduct remote optimizations and parameter studies, providing attendees with practical insights and solutions.

In the evolving landscape of Azure Cloud Services, Microsoft has transitioned from "Classic" resources to the Azure Resource Manager (ARM), which simplifies the deployment of Windows MPI applications by allowing users to submit declarative JSON templates defining necessary cloud resources. ARM, akin to AWS's CloudFormation, offers advantages like parallelized resource deployment, although it presents challenges with conditional resource deployment due to JSON's limitations. For high-performance computing (HPC), Azure offers the HPC Pack and Azure Batch service, with the latter recently supporting multi-instance MPI tasks on VM pools, albeit requiring developer familiarity. A new ARM template has been devised to facilitate the deployment of a basic Windows MPI cluster without the HPC Pack's overhead, supporting SSH for remote access and simplifying MPI job execution on Azure. This approach aims to provide a more accessible entry point for users beginning with MS-MPI, featuring easy deployment via GitHub and enabling both command line and GUI access options.

Rescale has expanded its capabilities to support various neural network software packages, including Keras, a Python library that facilitates the creation and training of neural networks. In this tutorial, a convolutional neural network (CNN) is trained using the CIFAR10 dataset, which contains 60,000 images across 10 categories, to classify images. The process involves loading and pre-processing the dataset, defining the network architecture with Keras, and training the model using stochastic gradient descent. After training, the model is tested for accuracy and saved for future use. The tutorial also demonstrates how to upload the training script and dataset to Rescale, conduct training on a GPU node, and classify new images by loading and transforming real images to match the training format. Finally, the tutorial explains how to use the trained model for image classification tasks and outlines the process of setting up a Rescale job for labeling images. This comprehensive guide showcases the power of using cloud-based resources for efficient neural network training and image classification.