You may have seen Cray’s recent announcement regarding our next generation supercomputer (codenamed “Shasta”) that we anticipate delivering to the Department of Energy’s Argonne National Laboratory (ANL) in the future. We don’t normally talk about a new system architecture this far in advance, but since it’s come out in the announcement, I thought I’d provide a brief overview of the Shasta system.
Shasta, a next generation supercomputer, will be the successor to both our Cray® XC™ line of performance computers (previously code-named “Cascade”) and our Cray® CS™ line of standards-based cluster systems. As such, Shasta is the most flexible cluster architecture system we’ve ever designed and the full embodiment of our adaptive supercomputing vision. We’ve been out talking with customers from HPC markets like meteorology and high-energy physics, as well as commercial markets like financial services and energy. Based on their feedback, we’ve designed Shasta to support requirements for both traditional HPC customers and commercial customers. Yes, this will be the infrastructure that takes us all the way to exascale computing, but it’s really designed to provide cost-effective solutions in standard datacenters, starting at quite modest system sizes.
Shasta’s cluster architecture will accommodate multiple processor families. We expect future generation Intel® Xeon™ and Xeon Phi™ processors to be an important part of that mix, but the system is designed to host other processor types as well. This, of course, is something our current Cascade design does too. But to accommodate a broader range of processors, Shasta will also offer multiple network types. We’ll naturally support the second generation Intel® Omni-Path network, and we’ve been working closely with Intel to provide a unique implementation of this network in Shasta. We’ll also support standards-based open networks to provide additional flexibility.
The Shasta network will have significantly higher performance and even lower diameter (ergo, low latency and bandwidth-efficient) than the groundbreaking Aries network in the Cray XC40 system. Using this network, we’ll be able to build truly balanced systems, where the network is tightly integrated and well matched to the processing nodes. High network bandwidth relative to local compute and memory bandwidth is key for sustained performance on real world codes, and a hallmark of Cray systems.
On the packaging side, some of our customers value high-density, scale-optimized cabinets, while others need standard cabinets that fit into their datacenter ecosystem. Shasta will support both scale-optimized cabinets and standard cabinets for rack-mounted equipment. We’ll also provide more flexibility in cooling approaches than with our XC line, which we’ll discuss in more detail as we approach general availability. Shasta will optimize TCO with warm water cooling, and simultaneously accommodate the trend toward higher-power processors. Cray systems over the past decade have been designed for easy and cost-effective upgrades, and Shasta will be no exception. We’re shooting to make this our most flexible, efficient and long-lived system infrastructure yet.
The software, too, will be highly flexible. Again, in response to varying customer needs, we’ll natively support both Cray’s scale-optimized software stack for high-end HPC solutions as well as standards-based software stacks. We’re working closely with Intel to define common APIs and layers to enhance modularity, interoperability and manageability of our runtime and management software stacks. Our management software will incorporate standard OpenStack interfaces and components, making it easy to manage and interoperate with existing environments. Support for Linux® containers, such as Docker, will provide enhanced portability for third party software.
With the Urika-GD™ and Urika-XA™ platforms, Cray systems are now being used for a variety of high-performance data analytics workloads, from in-memory Spark™-based applications to complex graph analytics. It turns out that high performance data analytics and traditional supercomputing benefit from the same system attributes, notably a strong interconnect and storage system. Not surprising since they’re both really about data movement. Shasta will support our full data analytics software stack as a runtime option, leveraging the Shasta interconnect to enable analytics workloads to be run at large scale and facilitating mixed simulation/data analytics workflows to be run on a single system, keeping the data in memory throughout the flow. This capability is a request we’ve been hearing from both our HPC and commercial customers.
I’m excited about Shasta and its potential for Cray’s customers. I’ve seen many systems introduced at Cray over the years, and I think Shasta will be our best yet. We’ll continue to evolve the XC line through the 2018 Shasta introduction. As with previous generations, we expect codes to transition seamlessly to the new system. I’m looking forward to our partnership with Argonne, and to the years of breakthrough science and engineering the Shasta series will enable.