By Bill Bryce, VP of Products, Univa

Five HPC predictions for 2020 and beyond

Back in the 1980s, a German chemical company (BASF) found a slogan that resonates to this day - "We don't make a lot of the products you buy. We make a lot of the products you buy better". I see parallels between BASF and our business at Univa. We don't build HPC applications to be clear - rather, we build tools that help HPC experts run applications more efficiently. In this capacity, we're fortunate to work with leading HPC users in disciplines from aerospace to life sciences to AI supercomputing. This provides a unique vantage point to how HPC workloads are evolving. As we start the next decade of HPC innovation, I thought it would be interesting to make some predictions for 2020 and beyond by extrapolating existing trends.

Computing gets more diverse - A decade ago, GPU-optimized applications were few and far between. Today, fueled by better developer tools, GPU-optimized libraries, and dramatic performance gains, most leading HPC applications support GPUs. Without GPUs, AI models for computer vision, natural language processing, and language translation would be hard to imagine. Our customers at Univa now demand advanced GPU-aware scheduling features, reflecting how integral GPUs have become to modern workloads. While NVIDIA is the reigning data center GPU champ, the market is poised to get much more competitive. Intel re-entered the market in late 2019 with a slew of GPU-related announcements (OneAPI, Ponte Vecchio, Rambo Cache, and Gelato), and AMD did the same with new EPYC servers, Radeon GPUs and GPU-related software solutions. New technologies such as neuromorphic processors are already in production, promising to reduce power consumption for specific AI workloads by up to 10,000x by modeling neural networks directly in silicon. Look for a decade of increased diversity as GPUs, FPGAs, and new types of accelerators augment CPUs, delivering dramatic performance gains, and picking up from where Moore's Law left off.  

Machine architectures matter again - Despite the dominance of x86-64/x64 in HPC, it's started to feel like processor architectures were less relevant. Most HPC users run Linux distros such as Debian, Ubuntu, or CentOS, and the user experience is largely the same across all. The same apt-get, yum, or pip commands work identically on x86, Arm, and Power systems auto-installing the same widely used components and frameworks. While containers have emerged as a preferred way of making HPC applications portable, they've also had the effect of making machine architectures important again. The availability of containerized software varies widely by platform. In Docker Hub at present, there are almost 2.4 million Linux container images for x86. For Arm and Arm-64, there are ~40,000 images, and for IBM Power LE (little-endian), there are ~5,000 images. The emergence of GPU-specific container clouds such as NVIDIA's NGC shows the way of the future. As new CPU and GPU architectures proliferate, we can expect increased fragmentation in container ecosystems and registries as hardware, software, and cloud providers race to promote and deliver containers tailored to their own architectures and software ecosystems. HPC users will need management software that transparently supports multiple container formats and registries across multiple CPU and accelerator architectures.

Clouds get cloudier - In HPC, cloud usually implies infrastructure-as-a-service (IaaS). While SaaS and PaaS offerings exist, the economics of running HPC in the cloud can be challenging. HPC users often prefer to burst into IaaS cloud services to preserve flexibility, portability, and deploy virtualized or containerized software environments that match on-premise environments. As new workloads emerge however, this dynamic is changing. Analytic and data science applications are frequently accessed via interactive, collaborative Notebooks - analysts code to high-level libraries such as Keras or Fast.AI that abstract away underlying software frameworks and infrastructure. New types of domain-specific cloud offerings are emerging tailored to these workloads including Paperspace, Gradient, Onepanel and FloydHub, where users can bring their own containers and data to avoid lock-in but pay for only the capacity that they use. While on-premise and cloud HPC clusters are here to stay, we can expect cloud usage to become more fragmented as users leverage an increasingly diverse spectrum of specialized cloud services offering new types of cloud consumption models. 

HPC cloud spending emerges as a key concern - While HPC users have been slow to embrace cloud historically, this is changing fast. Recent research shows a dramatic 60 percent increase in HPC cloud spending from just under $2.5 billion in 2018 to approximately $4 billion in 2019. Managing cloud spending is especially challenging in HPC, where users have an enormous appetite for computing resources. Look for cloud spend-management and budget-aware workload management tools to be increasingly important as the use of HPC in the cloud accelerates.

HPC shifts to the edge - Today, most HPC workloads run in centralized data centers operated by governments, enterprises, and cloud providers. The internet of things (IoT) accelerated by 5G networks will be a game-changer, promising network speeds as high as 20 Gbps, millisecond latency, and up to a million devices per square kilometer. It's easy to see autonomous vehicles, drones, and vast arrays of sensors sharing vastly more telemetry than ever before. Backhauling all this data to far away clouds with tens of milliseconds of latency won't work - Clouds will need to get much closer to the ground - hence the term fog computing. We already see the effects of data gravity in commercial HPC workloads such as algorithmic trading, streaming data analytics, and seismic analysis, where latency and data considerations require HPC processing close to the edge of the network. With faster networks, larger datasets, and a vast increase in the number of connected devices and sensors, this shift to the edge will only accelerate.

##

About the Author

Bill Bryce brings 14 years of experience to his position as VP of Products in which he leads product management. Bill was instrumental in introducing agile development across Univa's distributed team, resulting in a doubling of engineering efficiency. Bill is also credited with the conception and development of key products including Univa's cloud management products. He received his B Math in Computer Science from the University of Waterloo.