October 2016 Summaries

Fixing bugs in code often leads to increased complexity and the risk of introducing additional bugs due to the need to handle various edge cases, typically addressed with if statements. These conditions complicate the code, making it harder to understand and maintain, especially during refactoring, which can lead to regressions. An example from Rescale illustrates a bug in the local UI state handling for cloud-based desktop instances, where optimistic state updates could lead to confusing transitions for users. Instead of relying on additional if statements, a more robust solution was implemented by maintaining a set of statuses for each desktop and displaying the appropriate status based on the set's contents, eliminating any dependence on the order of status updates. This approach not only resolved the immediate issue but also addressed similar bugs, demonstrating the value of considering the broader class of problems when fixing bugs to avoid future complexity.

Rescale Deep Learning Cloud introduces an interactive notebook feature that facilitates iterative deep learning workflows by enabling users to switch between interactive data preprocessing and batch neural network training. The guide provides a detailed example using the CIFAR10 image classification dataset, demonstrating how to start a Rescale Linux Desktop with a NVIDIA K80 GPU, attach TensorFlow software, and utilize Jupyter Notebook for network training. Users are shown how to adapt TensorFlow's CIFAR10 training example, run different convolutional layer configurations, and observe training progress using TensorBoard via SSH tunnel. The process allows launching batch training jobs directly from the notebook, with a focus on comparing performance between two-layer and three-layer convolutional networks. The ability to iterate development in a consistent environment with the batch training cluster minimizes configuration discrepancies, enhancing reliability. For more intensive workloads, Rescale offers configurations supporting up to eight K80 GPUs, and users are encouraged to explore this workflow by signing up for the service.

A recent study by Accenture highlights the growing importance of cloud strategies among IT leaders, with 95% of respondents already having a five-year plan in place. Despite security concerns, over 80% believe that public cloud security is more robust than in-house solutions, while 89% see cloud implementation as a competitive advantage for innovation and agility. The expanding cloud market, with its low barrier to entry, presents a challenge due to the overwhelming choices of SaaS, IaaS, and PaaS technologies, necessitating a strategic approach in selecting a high-performance computing partner. Key considerations for IT leaders include scalability, flexibility, compatibility with hybrid environments, and robust security measures. The ability to manage both cloud and on-premises resources and align cost with demand is crucial, with emphasis placed on choosing a partner that can adapt to an enterprise's evolving needs. As cloud adoption accelerates, IT managers are tasked with navigating the "how" of cloud integration, while the "why" of cloud adoption, such as its advantages, is largely understood.

A recent revisit to Google Compute Engine's (GCE) instances has revealed significant improvements in network performance, particularly in latency and bandwidth for different instance types. Using the OSU Micro Benchmarks version 5.3.2, tests were conducted on Debian GNU/Linux 8 with MPICH3 across three instance types: n1-standard-32, n1-highmem-32, and n1-highcpu-32, ensuring they were on separate physical hosts for accurate network measurements. The average latency for these instances was found to be around 40 to 45 microseconds, a notable enhancement from previous results of around 180 microseconds. In terms of bidirectional bandwidth, the n1-highmem-32 instance exhibited the highest performance, reaching up to 1076.29 MB/s, consistent with Google’s official network throughput documentation. These findings suggest that GCE has optimized its offerings to better support high-performance computing applications.

The relationship between IT departments and engineering teams in organizations is often marked by friction due to differing objectives and resource constraints, with IT being viewed as a bottleneck by engineers seeking more advanced tools and resources. However, the advent of cloud computing has begun to alleviate this tension by democratizing IT resources and shifting budget control to engineering managers, allowing them to select the hardware and software solutions that best fit their needs. This shift transforms the traditional capital expenditure model into an operating expense model, requiring engineers to consider project costs while designing workloads. As IT managers focus on ensuring the security and integrity of cloud solutions, both departments can collaborate more effectively, as demonstrated by a successful case where an engineering manager found a cloud-based solution for high-performance computing needs. With public cloud services growing rapidly, the focus has shifted from whether organizations will adopt the cloud to how they will utilize it, enabling innovation to thrive without being hindered by traditional IT processes, provided there is sufficient operational budget.